Feasibility of remote sensing to inform site condition monitoring assessments on Scottish terrestrial SSSIs

Transcription

1 COMMISSIONED REPORT Commissioned Report No. 083 Feasibiity of remote sensing to inform site condition monitoring assessments on Scottish terrestria SSSIs (ROAME No. F02LG15) For further information on this report pease contact: David Wood Scottish Natura Heritage 1 Kimory Estate LOCHGILPHEAD Argy PA31 8RR Teephone: E-mai: david.wood@snh.gov.uk This report shoud be quoted as: Yaop, A.R., Thomas, G., Thacker, J., Brewer, T. and Sannier, C. (2004). Feasibiity of remote sensing to inform site condition monitoring assessments on Scottish terrestria SSSIs. Scottish Natura Heritage Commissioned Report No. 083 (ROAME No. F02LG15). This report, or any part of it, shoud not be reproduced without the permission of Scottish Natura Heritage. This permission wi not be withhed unreasonaby. The views expressed by the author(s) of this report shoud not be taken as the views and poicies of Scottish Natura Heritage. Scottish Natura Heritage 2004.

2

3 COMMISSIONED REPORT Summary Feasibiity of remote sensing to inform site condition monitoring assessments on Scottish terrestria SSSIs Commissioned Report No. 083 (ROAME No. F02LG15) Contractor: Cranfied University Background A review of remote sensing technoogies and cassification protocos with regard to their abiity to meet specific conservation monitoring criteria. Main findings A review of remote sensing and image cassification technoogies with regard to their suitabiity for contributing to future Site Condition Monitoring reporting requirements has been undertaken. Part of this process has invoved assessing technoogies against specific area monitoring targets. To assist in this an anaysis of the impications of the target criteria in terms of sensor resoution and cassification has been undertaken. With respect to sensor technoogy, the minimum spatia resoution required to meet the specific targets for assessing area extent changes is estimated at <2m for a arge proportion of the woodand sizes ikey to be in Scotand s SSSI inventory. Very few sensors on existing remote sensing patforms are abe to meet this requirement and this excudes many of the traditiona earth resources sateite systems such as Landsat, IRS, ERS and SPOT. Over the short-term, the ony source of digita image data ikey to have coverage for a SSSI sites and with sufficient spatia, spectra and tempora resoution to meet SNH monitoring requirements, are airborne sensors, most notaby those producing aeria photography from which ortho-corrected imagery can be derived. Sateite borne utra-high resoution systems coud feasiby provide an aternative source for those areas where data exist, and there is expected to be significanty increased avaiabiity of such data in the medium term. Over the onger-term airborne hyper-spectra and muti-band SAR data might become avaiabe at a cost that makes their use possibe. With respect to cassification methodoogies and the abiity to correcty identify features of interest, the reported accuracies of most automated cassification and segmentation agorithms is beow that required to meet SNH criteria when using traditiona remote sensing data sources. They are aso beow what can be expected from manua interpretation of high resoution air photography by trained interpreters, athough it shoud be considered that automated methods are more cost effective when appied to arger areas and when using sateite images. A number of newer techniques may be capabe of reativey high accuracies of cassification when correcty appied, and offer hope that in the near future they may find appication in a CSM programme. However, it is a concusion of this report that, for the criteria specified, there are no immediatey depoyabe techniques avaiabe. Whatever cassification techniques are in future adopted it is important to separate the issue of accurate cassification from those of consistent boundary determination. This is especiay important with compex woodand sites and appears to be one of the more important issues to be resoved. Recent research suggests a number of approaches might be empoyed to provide more automated methods of boundary pacement and recommendations about further investigation of these are made. For further information on this project contact: David Wood, Scottish Natura Heritage, 1 Kimory Estate, Lochgiphead, Argy PA31 8RR. Te: For further information on the SNH Research & Technica Support Programme contact: The Advisory Services Co-ordination Group, Scottish Natura Heritage, 2 Anderson Pace, Edinburgh EH6 5NP. Te: or ascg@snh.gov.uk

4

5 Gossary AISA ALTM API CASI DEFRA DN DLR EMR INSAR IMU ISODATA JNCC LASER LIDAR PCA SCM SNH SAR SASI TWINSPAN Airborne Imaging Spectroradiometer for Appications Airborne Laser Terrain Mapper Aeria Photograph Interpretation Compact Airborne Spectrographic Imager Department for Environment, Food and Rura Affairs Digita Number German Aerospace Centre Eectromagnetic Radiation SAR Interferometry Inertia Management Unit Iterative Sef-Organizing Data Anaysis Technique Joint Nature Conservation Committee Light Ampification by Stimuated Emission of Radiation Light Detection and Ranging Principa Component Anaysis Site Condition Monitoring Scottish Natura Heritage Synthetic Aperture Radar Shortwave Infrared Airborne Spectrographic Sensor Two-way Indicator Species Anaysis

6 EXECUTIVE SUMMARY A review of remote sensing and image cassification technoogies with regard to their suitabiity for contributing to future Site Condition Monitoring reporting requirements has been undertaken. Part of this process has invoved assessing technoogies against specific area monitoring targets. To assist in this an anaysis of the impications of the target criteria in terms of sensor resoution and cassification has been undertaken. With respect to sensor technoogy, the minimum spatia resoution required to meet the specific targets for assessing area extent changes is estimated at <2m for a arge proportion of the woodand sizes ikey to be in Scotand s SSSI inventory. Very few sensors on existing remote sensing patforms are abe to meet this requirement and this excudes many of the traditiona earth resources sateite systems such as Landsat, IRS, ERS and SPOT. Over the short-term, the ony source of digita image data ikey to have coverage for a SSSI sites and with sufficient spatia, spectra and tempora resoution to meet SNH monitoring requirements, are airborne sensors, most notaby those producing aeria photography from which ortho-corrected imagery can be derived. Sateite borne utra-high resoution systems coud feasiby provide an aternative source for those areas where data exist, and there is expected to be significanty increased avaiabiity of such data in the medium term. Over the onger-term airborne hyper-spectra and muti-band SAR data might become avaiabe at a cost that makes their use possibe. With respect to cassification methodoogies and the abiity to correcty identify features of interest, the reported accuracies of most automated cassification and segmentation agorithms is beow that required to meet SNH criteria when using traditiona remote sensing data sources. They are aso beow what can be expected from manua interpretation of high resoution air photography by trained interpreters, athough it shoud be considered that automated methods are more cost effective when appied to arger areas and when using sateite images. A number of newer techniques may be capabe of reativey high accuracies of cassification when correcty appied, and offer hope that in the near future they may find appication in a CSM programme. However, it is a concusion of this report that, for the criteria specified, there are no immediatey depoyabe techniques avaiabe. Whatever cassification techniques are in future adopted it is important to separate the issue of accurate cassification from those of consistent boundary determination. This is especiay important with compex woodand sites and appears to be one of the more important issues to be resoved. Recent research suggests a number of approaches might be empoyed to provide more automated methods of boundary pacement and recommendations about further investigation of these are made.

7 Contents Summary Gossary Executive summary 1 INTRODUCTION Aims and objectives Methods and sources of information 2 2 SELECTION OF REMOTE SENSING TECHNOLOGIES APPROPRIATE TO CONSERVATION MONITORING OBJECTIVES Sensor resoutions Identification of sensors based on resoution characteristics Identification of sensors based on conservation objectives Target sensor characteristics 7 3 REVIEW OF APPROPRIATE SENSOR TECHNOLOGIES Airborne digita ortho-photography Technica description Aeria photography data formats Review of recent appications Airborne SAR (Synthetic Aperture Radar) Technica description Review of recent appications Airborne hyper-spectra sensors Technica description Review of recent appications LIDAR (Light Detection and Ranging) Technica description Review of recent appications Sateite-based sensors Technica description Review of recent appications Operationa issues 20

8 4 IMAGE DATA PROCESSING AND INFORMATION EXTRACTION Overview Pixe cassifiers Unsupervised cassification Supervised cassification Sub-pixe cassification Object cassifiers Data fusion Use of cassifiers for vegetation mapping Unsupervised custering Supervised, maximum ikeihood cassification Spectra separabiity Artificia neura networks Poygon-based cassifiers 31 5 CONCLUSIONS AND RECOMMENDATIONS Identification of trees and woodand boundaries Additiona issues Avaiabiity of remote sensing data to meet SNH area target monitoring requirements Avaiabiity of cassification routines to meet SNH area target monitoring requirements Summary of comments, suggested strategies and techniques Summary comments Summary strategies and techniques 35 6 REFERENCES 37 7 Appendix 1 Minimum object sizes in which a 5% change be detected using pixe cassification techniques Introduction Methods Resuts 45 8 Appendix 2 Layman s summaries of key remote sensing technoogies Aeria digita ortho-photographs Passive optica sensors Active sensors Pixe cassifiers Object cassifiers 54

9 1 INTRODUCTION The rationae for conducting this review derives from a Scottish Natura Heritage (SNH) requirement, as part of the Site Condition Monitoring Programme (SCM), to report on the condition of a notified features of woodand SSSIs. Monitoring procedures foow the Common Standards Monitoring approach agreed by the conservation agencies of Engand, Northern Ireand, Scotand, and Waes with a features assessed in reation to an agreed ist of attributes at east once every six years. Notification of woodand features of SSSIs is normay based on habitat or species type and the attributes to be monitored incude area, natura processes and structura deveopment, regeneration potentia, composition (trees and shrubs) and quaity indicators (Engish Nature, 2002). Quantification of the area covered by the specified habitat types is recognised as a critica component of SCM. At present area assessments are carried out manuay, using maps derived from ground surveys or from hardcopy aeria photographs. This process is time consuming and, as importanty, acks a sufficienty objective basis to ensure accurate and repeatabe monitoring of the area attribute of woodand features. Remote sensing technoogies may provide a basis for deveoping an improved monitoring system. Recent deveopments in sensor technoogy provide a range of digita data sources with spatia and radiometric quaities potentiay suited to monitoring the area of woodand sites, whie data processing technoogies offer the possibiity of improved methods of interpretation. In summarising the current status of remote sensing technoogies a principa aim is to aow SNH to determine if any of these technoogies, either singy or in combination, coud be depoyed during either of the next two rounds of SCM (Apri 2005 March 2011; Apri 2011 March 2017). In conducting this review, nove methods have been used to identify those sensor and processing technoogies that have the highest probabiity of meeting the specific targets set by SNH and therefore form the basis for the iterature review. In addition to reviewing currenty accepted practice, this review seeks to identify approaches that ceary have potentia but have yet to be vaidated sufficienty. 1.1 Aims and objectives The overa aim of this review is to summarise the current status of remote sensing technoogies in respect of determining area assessments of natura and semi-natura habitats and with a particuar regard to seminatura woodands. Given this context the specific objectives of this review are to: Identify specific remote sensing and image cassification technoogies that are appropriate to meeting the SNH requirements for condition monitoring; Identify specific use of these technoogies in the context of monitoring semi-natura woodand environments; Provide guidance as to the direction for depoyment of current technoogies judged capabe of meeting SNH objectives; Identify remote sensing approaches most ikey to meet SNH objectives but that require further evauation or testing prior to depoyment. 1

10 1.2 Methods and sources of information Given the restricted time-frame for conducting this study seected peer-reviewed pubished and non-peer reviewed materia from seected organisations over the ast five years form the main basis of the review. Prior to this date the avaiabe technoogies used are unikey to be abe to form the basis of a readiy useabe digita remote sensing methodoogy for meeting SNH objectives. More recent deveopments incuding the increasing avaiabiity of digita ortho-imagery, utra-high resoution sateite imagery and more sophisticated pixe and object based cassifiers may provide the basis of an operationa method. The restricted time frame does not appy to techniques or exampes deemed to be semina, and in this case iterature is cited earier than Aso the current review focuses on specific sensors and does not provide a comprehensive review of a avaiabe sensors. Two recent reviews have been undertaken on behaf of the Department for Environment, Food and Rura Affairs (DEFRA, 2003) and the Forestry Commission (Mathus et a. 2000). Taken together these provide a comprehensive review and description of the main sateite and aeria sensors and their associated spatia, spectra, radiometric and tempora resoution characteristics. Such information is not repeated here. 2

11 2 SELECTION OF REMOTE SENSING TECHNOLOGIES APPROPRIATE TO CONSERVATION MONITORING OBJECTIVES Remote sensing monitoring strategies require two main components, sources of digita image data and a method for cassifying or segmenting these data into the required information. The identification of an appropriate source of data is normay based upon consideration of the fundamenta resoution characteristics of specific sensors and for competeness these are defined beow. However, the specific objective of the monitoring is important and a nove method has been devised to provide an objective basis for identifying those technoogies to be reviewed and for assessing the resuts of their recent impementation. 2.1 Sensor resoutions The key considerations regarding sensor choice centre around radiometric, spatia, spectra and, especiay important for monitoring programmes, tempora resoution. It shoud aso be emphasised that for any nationa monitoring requirement cost considerations wi aso be a significant factor. Radiometric resoution One of the principa deveopments in sensor technoogy is improved radiometric sensitivity. Modern sensors record images directy in digita form and represent the magnitude of the sensed signa in terms of the Digita Number (DN). This represents an arbitrary positive integer, and the potentia range of vaues that coud be dynamicay recorded is described as the radiometric resoution. Many sensors have operated with 8 bits of precision equivaent to a range in digita number vaue of 0 255, but increasingy sensors are being deveoped that record images with an increased eve of precision using 10 (DN range of ), 12 (DN range of ) and even 14 bits (0 4095) to represent the magnitude of the energy sensed for each image pixe. The direct consequence of these improvements in radiometric resoution is an increased abiity to measure reativey sma changes in energy from target features. The radiometric resoution shoud therefore have a direct infuence on the abiity to discriminate or identify specific features within an image. Spectra Remote sensing devices record the magnitude of eectro-magnetic radiation (EMR) that is either refected, emitted or back-scattered from the surface. One fundamenta property of EMR is waveength, and this is normay used as a basis for separating the three fundamenta forms of remote sensing. Typicay these individua forms of remote sensing are described as optica (refected radiation at soar waveengths), therma (emitted radiation at therma waveengths) and microwave (back-scattered radiation at microwave waveengths). Remote sensors are aso described as being passive and active, depending upon whether they use an existing natura source of EMR or provide their own source of iumination. It is generay acknowedged that therma remote sensing can provide ony imited discrimination of vegetated surfaces and therefore therma sensing devices are not incuded in this review. The spectra resoution of a sensor indicates the number of wavebands that can be individuay recorded across the principa sub-divisions of the eectromagnetic spectrum (ie optica, therma and microwave). A simpe categorisation can be identified (Tabe 2.1.1). 3

12 Tabe Sensor cassifications and exampes Cassification Wavebands Exampe patform/sensor Singe band 1 SPOT Pan Orb Image OrbView Pan ERS SAR Muti-spectra <10 Landsat MSS Landsat TM EOS ASTER SPOT HRV Hyper-spectra EO-1 Hyperion Hymap (Aeria) It is generay accepted that an increased number of spectra wavebands eads to an improved abiity to discriminate features present within an image and therefore considerabe emphasis is often aso given to the spectra resoution as a defining characteristic of sensor technoogy. Spatia The spatia resoution of a sensor is defined as a pixe or IFOV (instantaneous fied of view) and usuay expressed as a inear measurement of the width of the area covered on the ground. A genera categorisation of sensors based upon pixe size is presented in Tabe 2.1.2, together with some exampe patforms and associated sensors. One of the main technica deveopments in the fied of remote sensing has been the increased number of sateite patforms that can provide utra-high resoution images with pixe sizes between 1m and 10m. Super-high resoution images are aso becoming avaiabe from airborne patforms. Tempora A remote sensing patforms have a revisit time, ie the time between repeat coverage of the same area. The typica range for specific sateite sensors is between a few days and 10 s of days. Notionay, these are far shorter than monitoring periods and therefore the tempora resoution can appear unimportant. However, for both passive and active optica sensors the presence of coud severey inhibits the coection of imagery. For airborne systems tempora resoution is not fixed and can be determined based on ogistica issues. Tabe Sensor spatia resoution and exampes Resoution Pixe size (m) Exampe patform/sensor Low 1000 METEOSAT VISSR NOAA AVHRR Intermediate Envisat ATSR2 SPOT VEGETATION Terra/Acqua MODIS High Envisat MODIS EOS ASTER Landsat TM Landsat MSS Utra-high 1 10 Earth Watch QuickBird Ikonos Ikonos Orb Image OrbView SPOT-5 Super-high 1 Digita aeria photography Airborne hyper-spectra SAR sensors 4

13 2.2 Identification of sensors based on resoution characteristics From previous work it is possibe to identify those resoution characteristics most reevant to the current requirement and to identify imiting vaues as a basis for determining those sensor technoogies that shoud have the potentia to meet a specific conservation monitoring objective. In respect to the radiometric resoution an approximate ower imit is suggested by the study of Kerekes and Landgrebe (1989). Athough based on simuated data their resuts indicated itte accuracy improvement beyond 8 bits radiometric resoution. For sensors with high to super-high resoution, this resuts from the dominance of other sources of variabiity in DN over the quantization eve since the sma pixe sizes resove a greater degree of the sma-scae variabiity associated with specific feature types, eg eaves, branches, shadows for tree dominated environments. On this basis improved radiometric resoution may be more significant for sensors with intermediate spatia resoution. With respect to spectra resoution, taking visua interpretation of norma coour photography as a baseine method for assessing woodand areas suggests sensor technoogies require a minimum configuration of 3 wavebands. Experience with coour aeria photography indicates that it is beneficia if one of the wavebands is in the near-infrared. However, it is generay true that the discrimination of vegetation casses is improved by acquiring images across a wider range of waveengths and this woud impy a significant advantage in respect of hyper-spectra sensors. However often wavebands acquired at simiar waveengths are highy correated so that much of the information they contain is redundant and provides itte benefit in discriminating the required casses. Once again it is usefu to try and identify approximate upper and ower imits within the context of this review. There are no specific reported studies that identify the reation between the number and position (within the eectromagnetic spectrum) of the sensor wavebands and the abiity to discriminate specificay the woodand casses most reevant to this study. However, the study of Thenkabai et a. (2000) athough focused on monitoring agricutura crops is interesting in this context. In this study 490 wavebands of optica data were anaysed to determine an approximate upper imit in terms of improved prediction of crop characteristics. The resuts indicated that 12 narrow bands provided an optima estimation of crop biophysica characteristics. It woud appear reasonabe to extend this genera principe to the discrimination of woodand categories. Consequenty, athough many hyper-spectra sensors are capabe of recording separate images for upwards of 100 wavebands, typicay this is ony undertaken for very specific types of research and, more operationay, data are usuay recorded in separate wavebands. The identification of sensor technoogies appropriate to a specific appication is often undertaken by consideration of the fundamenta spatia resoution characteristics of a specific sensor. Wikie and Finn (1996) suggest that the minimum spatia resoution of a sensor for a specific appication can be approximated as haf the dimension of the minimum feature to be observed. The SCM criteria for monitoring open areas as sma as 10m x 10m indicates sensor technoogies of 5m or better, ie utra-high and super-high resoution systems are required. However, recognition (whether by visua interpretation or by use of a digita agorithm), as opposed to simpe detection, of a feature requires that the minimum feature be represented by a number of whoe pixes. Ignoring uncertainty in accuratey cassifying pixes, an approximate indicator can be derived based upon the requirement to derive an estimate of the mean signature for a feature. A simpe approximate rue is that any feature shoud be represented by a minimum of pixes. Given the 10m x 10m threshod, this suggests a requirement for pixe sizes of 2m. 5

14 With respect to tempora resoution the most reevant factor is the utiity of each pass of the sensor in gathering usefu data. Coud cover within the UK is a frequent occurrence and restricts the utiity of optica sateite systems. It is extremey hard to obtain enough scenes without coud cover to create a mosaic of an area as arge as Scotand, and where it is possibe it is usua for individua scenes to be gathered months apart. For vegetation monitoring this presents considerabe issues as phenoogica changes may ater the spectra appearance of pants over that time making digita cassification difficut. This can be mitigated to some extent by performing cassification individuay for each scene and then mosaicing the outputs in a GIS. 2.3 Identification of sensors based on conservation objectives In deveoping any system of monitoring based upon remote sensing technoogies it is important to consider the scae of the features that are the basis of the monitoring, in this case Scottish Terrestria Sites of Specia Scientific Interest. At 31/03/2002, there were 1,447 SSSIs in Scotand, covering a tota area of 1,007,260ha, or 12.8% of and area. Woodand SSSIs within Scotand range in size from 2.73ha to more than 5000ha, athough the majority of sites fa within a range of ha. Note however that this is the tota area of each designated SSSI and not necessariy the area of woodand parces. Each SSSI site may comprise a number of parces and these may be far smaer. In addition, SNH have identified, as an exampe of a specific area reated condition monitoring requirement, the foowing targets for the Gen Raoch to Baravaa Woods SSSI. This SSSI is a coasta site of approximatey 400ha in Knapdae, Argy and contains severa notified features, one of which is semi-natura broadeaved woodand. The area reated targets are: Target 1: To maintain the tota area covered by semi-natura broadeaved woodand within the SSSI at not ess than 95% of the area present at SSSI notification in 1985 (200ha). To determine whether or not the target has been met, it wi be necessary to define an overa enveope, or boundary, for the woodand feature within the SSSI. This enveope shoud ignore sma open areas (<5ha) in the woodand canopy. The area covered by the enveope shoud be measured with enough precision to detect a 5% change in the overa woodand area (for this site, 10ha), based on two rounds of monitoring carried out six years apart. Target 2: To maintain the canopy cover in woodand areas at between 30% and 90%. To determine whether or not this target has been met, it wi be necessary to detect sma open areas within the overa woodand enveope. The minimum area to be detected wi be 0.01ha (ie 10m x 10m) and the maximum 5ha. The tota area covered by these open areas wi be subtracted from the overa woodand enveope area to give the area of canopy cover for the woodand feature. This is then expressed as a percentage of the woodand enveope area. It has been anticipated that the requirement to reiaby detect a 5% change in area extent between monitoring rounds wi require identification of changes in wooded bocks far smaer than the overa area of the SSSI. There are few existing reviews of the abiity of any specific sensor to meet these requirements and therefore an anaysis of the major issues controing the abiity of a sensor to detect edge features has 6

15 been undertaken. The detais of the anaysis are presented in Appendix 1. The resuts indicate that to guarantee a capabiity to vaidate the specified SCM area reated target criteria across the range of seminatura woodand environments characteristic of Scottish Terrestria Sites woud require image data with 25cm spatia resoution and an interpretation method capabe of achieving 99% cassification accuracy. Reducing the requirement to monitoring 75% of woodand sites using the area target criteria requires a minimum of 50cm pixes and cassification accuracies of 90%. 2.4 Target sensor characteristics An assessment soey of resoution characteristics woud indicate that sensors with better than 2m pixes, a minimum of three bands and a minimum radiometric resoution of 8 bits are required. An anaysis that incudes the scae of the features to be assessed and specific eves of cassification accuracy indicates that ony sensors capabe of super-high spatia resoution combined with cassification techniques capabe of accuracies exceeding 90% are reevant to the current review. For the purposes of this review, and combining these resuts it has been decided to consider those sensors capabe of utra-high resoution or better and with more than a singe waveband, since these sensors shoud aow monitoring of a minimum of 50% of the Scottish Terrestria Sites. The foowing two sections review the technica characteristics and cassification approaches appied that meet the above criteria and methods of interpretation. The sensor technoogies reviewed incude airborne photographic, hyper-spectra, synthetic aperture radar (SAR) and LIDAR systems and sateite borne utra-high resoution systems. The cassification methodoogies incude pixe, sub-pixe and object based cassifiers and a variety of cassification agorithms (eg unsupervised, supervised, fuzzy, neura network). In doing so there is necessariy some overap in the review. 7

16 3 REVIEW OF APPROPRIATE SENSOR TECHNOLOGIES 3.1 Airborne digita ortho-photography Technica description Raw aeria photography sti forms the highest resoution commerciay avaiabe remotey sensed data. Ground resoution is determined by a combination of the resoving power of the camera system and the scae of the photography. For exampe, 1:10,000 scae photography acquired using a modern metric aeria camera can provide digita images with pixe sizes of 10 15cm. Most historica aeria photography has been acquired on back and white fim. However, the majority of coverage today is captured using true coour. Both types of photography have their advantages: Back and white photography has the highest spatia resoution Fying conditions are ess critica with back and white photography as fiters can be used to reduce the effects of atmospheric haze Coour photography is more immediatey interpretabe The human eye is abe to perceive more shades of coour than tones of grey. This can be an advantage as we as a disadvantage. Too many subte changes of shade can bur differences between vegetation types. Coour infrared photography is not widey fown in the UK athough it is vauabe for interpretation of vegetation communities. Much of the research that has investigated the use of aeria photography for forest monitoring has used coour infrared photography. Digita aeria cameras are now avaiabe and this has the potentia to increase the avaiabiity of coour infrared imagery. Aeria photography is commony fown as stereo photography. Between adjacent photographs aong a fight ine there wi be an overap of 55 65% and across fight ines an overap of 25 30%. Photography fown in this way aows adjacent pairs of photographs aong a fight ine to be viewed in 3-D using a stereoscope. This greaty enhances the interpretabiity of aeria photographs, particuary in forestry appications where height is an important factor in feature recognition. Softcopy photogrammetric software, for exampe, SOCET Set, aow 3-D visuaisation of stereo pairs within the digita environment. One of the most significant recent deveopments has been the provision of digita ortho-photography. An ortho-photograph is an aeria photograph that has been modified such that its scae is uniform throughout the photographic image, that is, the dispacements that arise from capturing a 3-D word onto 2-D fim are removed. A digita ortho-photograph is a product derived from stereo aeria photography. Advances in computing hardware and software have revoutionised ortho-photograph production, such that it is now widey avaiabe in the UK. The ortho-rectification process invoves photogrammetricay scanning the origina fim based aeria photographs. Commony this produces a coarser resoution image compared to the origina aeria photograph. For exampe, the Miennium Map product from GetMapping Pc has a resoution of 25cm (12cm for seected urban areas). Dispacement correction is undertaken using either ground contro points derived from high precision GPS measurements and a digita eevation mode, derived either within the photogrammetry software or from an existing dataset (ortho-correction) or rectification to existing digita map data and an existing digita eveation mode (geo-correction). 8

17 Individua ortho-images are mosaiced together, and during this process coour baancing is undertaken to minimise variations of tone/shade of coour that resut from exposure variations on the origina aeria photographs. However, the mosaic processing, especiay when appied to arge scae products wi not be perfect and both spatia and spectra artefacts may be apparent at image boundaries. This may be a particuar issue if automated cassification techniques are to be appied to the imagery. The fina product is a 2-D image. This can be viewed in a GIS in the same way as any other data ayer or can be draped over an eevation mode and viewed using 2.5-D perspective viewing. The geometric accuracy of ortho-photography is determined by the type of correction empoyed. The nationa map base is undergoing a Positiona Accuracy Improvement (PAI) programme that wi resove many of the known errors within the dataset. Ortho-photography produced using maps for dispacement correction prior to PAI wi contain inherent inaccuracies when compared to high accuracy GPS measurements taken in the fied. In particuar, areas of high reief wi be ess accurate if geo-correction to a map base has been used. Ony a fu ortho-correction using contro points and a digita eevation mode wi give high precision in these areas. However, this is a more costy method of ortho-photo production. Nationa coverage of digita ortho-photography wi be avaiabe as the imagery ayer of Ordnance Surveys MasterMap product. This is expected to have a three year update cyce athough more remote areas may go five years between updates. Frequency of update wi be dependent upon fying conditions. GetMapping Pc have captured the majority of the UK, athough ortho-photography is not avaiabe for a of Scotand at present. However, currenty data has been captured for ~45% of Scotand of which 35 40% has been orthocorrected. The pixe size wi be 50cm for the Scottish data Aeria photography data formats Individua frames of aeria photography can be purchased as fim or photographic print products. These images are normay contact prints at the origina resoution of the imagery. They can be viewed monoscopicay (2-D viewing) or stereoscopicay (3-D viewing) if stereo pairs are avaiabe. A mirror or pocket stereoscope is often used to enabe stereo viewing. Handing photographs in this format can be aborious because mapping requires visua interpretation and referencing onto an existing map base. Digita formats are either compressed or uncompressed. Two parameters are inherent within a digita fies pixe resoution and coour depth. Pixe resoution refers to the scanning resoution used and is measured either in microns or dots per inch (dpi). Coour depth refers to the number of potentia coours that can be represented in the fie and is somewhat anaogous to a combined spectra/radiometric resoution parameter. Eight bit data can ony represent coour vaues from 0 255, therefore restricting the number of coours that can be shown on the digita image. Twenty-four bit data can represent 16,777,216 coours and is therefore the preferred coour depth for coour aeria photography. Uncompressed digita format Uncompressed image fies are digita representations of origina images using pixes of a known resoution and coour depth that have not undergone any ateration to reduce fie storage requirements. High resoution images, that is, those with a sma pixe size, create arge fie sizes. For exampe, a singe metric camera true coour photograph, scanned at 22 micron (equivaent to 1200 dpi) and 24 bit coour depth, has a fie size of approximatey 350Mb. A 1km 2 tie of true coour ortho-photography suppied with the same 9

18 specification is approximatey 50Mb in size. The TIFF (Tagged Image Format Fie) format is the most commony used uncompressed image format and most image processing and GIS programs are abe to import this fie format. The advantage of the uncompressed format is that the image is unatered from the origina. No artefacts arising from a compression process wi be evident. Compressed formats Compression is a software procedure that encodes the data structure so that the storage requirement is ess. Two methods can be empoyed: Loss-ess compression which seeks to preserve a the data yet sti reduce fie sizes; Statistica compression which simpifies the data using statistica agorithms and hence reduces fie size. Using the same specifications as above, the commony appied compression methods give fie sizes of between 2 4Mb for a 1km 2 tie. The two most commony used oss-ess compression formats are MrSID and ECW and the most common statistica compression format is JPEG. The JPEG format is widey used in graphics appications and therefore fies can be imported into the majority of image processing and GIS programs. MrSID and ECW conversion is now possibe with most GIS packages. However, athough these methods significanty reduce fie size, image quaity is degraded when viewed at high magnifications. Compression patterns are visibe as coour variation between adjacent pixes which woud otherwise have simiar pixe vaues in an uncompressed fie. Visua interpretation of compressed imagery can therefore be hampered with certain cover types. The effects are most evident when a statistica compression technique is used. Recent experience of using TIFF format compared to compressed JPEG suggests the extra detai avaiabe in the former can be crucia in identifying vegetation types and shoud be used wherever possibe Review of recent appications A ong-standing roe for traditiona forms of aeria photograph interpretation has been change detection. Many studies have demonstrated its worth and accuracy (Dunn et a. 1990; Green et a. 1993; Tayor et a. 2000). The traditiona approach has been to manuay deineate boundaries either directy onto the photography or a map covering the same spatia area. However, there are two criticisms of this approach. Firsty, it is subjective as it is reiant upon the interpreter pacing the boundaries by visua interpretation. Secondy, identification and mapping of features that have sma spatia extent is time consuming and may be impractica for arge areas (Hoopainen and Wang 1998; Kadmon and Harari-Kremer 1999). Aeria photograph interpretation has been widey used for monitoring, inventory and mapping of forestry and its advantages and disadvantages we estabished (Howard 1991). The avaiabiity of digita orthophotography wi ease the process by which monitoring data can be recorded. The geometric fideity of ortho-images removes in part the positiona errors of boundary pacement, but does not remove the subjectivity of deciding where the ine shoud be mapped based upon the and cover evident in the photography. 10

19 However, once a boundary has been mapped digitay it can be overaid onto ortho-photography of a different date to visuay determine the extent of change. Various dispay toos in image processing and GIS software can be used to aid this process. A ficker too aternates images from two or more dates sequentiay aowing the interpreter to quicky see where changes have occurred (Cameron et a. 2000). Swipe dispays one image over another and the interpreter can then draw one image back over the other to identify changes (Bowman et a. 2001). Fade dissoves one image date into another. A of these toos rey upon the interpreter to visuay assess where changes are occurring but because the human operator is very good at picking up changes when presented with two images it can be done efficienty and accuratey. Severa factors can affect the interpreter s abiity to detect change using these methods, for exampe, the time of day or year between the sets of photography (Cameron et a. 2000) and processing of the image mosaics (Hoopainen and Wang 1998). Discrimination of forestry casses is a we-estabished visua aeria photograph interpretation technique that utiises spectra and textura properties from the aeria photography. The discrimination of woodand from open habitats and areas of open water is generay straightforward. Visua interpretation is abe to deineate boundaries quicky and accuratey. For exampe, in the Monitoring Landscape Change in the Nationa Parks project (Tayor et a., 2000), 95% mapping accuracy was obtained for woodand casses using traditiona photo interpretation techniques. Differentiation of broadeaf from coniferous woodand can be ess straightforward and is reiant to a arge degree on the spectra and textura properties of trees. The ease with which tree types can be separated spectray is in part determined by the date of photograph acquisition. Eary season and ate season photography can show cear differences in coour/tona vaues between broadeaf and coniferous trees, which enabe accurate and efficient visua interpretation. However, confusion can arise as a resut of simiar textura properties between tree species. This is ess of a probem in semi-natura forest, as the trees exhibit a natura crown structure that they may not in a cosey panted managed forest. Shape and shadow can aso aid interpretation where individua trees can be identified. Severa studies have investigated the use of automated cassification techniques on digita orthophotographs. Data cassified incude tree type, basa area, stem voume and mean tree height. Many digita cassification methods are based upon refectance vaues. Ortho-photographs present a number of probems in this regard. For a continuous stand of trees covering a singe frame of aeria photography, refectance wi vary due to the position of individua trees with respect to the centre or principa point of the aeria photograph. The reief dispacement effect increases radiay from the principa point, causing ta structures, such as trees, to ean away from the centre of the photograph. Combined with the position of the sun, the net effect is that some areas of the photograph wi be much ighter because more of each tree is seen to face the sun and other areas wi be darker as more shadow is evident. This is known as the bi-directiona refectance factor. Factors that increase the bi-directiona refectance effect incude: Scae of photography (the arger the scae, the more obvious the effect); Soar eevation (ow soar anges increase bi-directiona refectance effects); A features position on the photograph (reief dispacement); Sope and aspect. 11

20 Potentiay the cassifier wi identify more than one category as a resut. In addition, exposure fa off or vignetting across a singe frame wi contribute to this phenomenon. These effects are most pronounced on arge scae photography, for exampe, scaes arger than 1:10,000. Anti-vignetting fiters on the camera can be used to reduce this effect. In high reief areas, where many Scottish SSSIs are ocated, uneven iumination of and cover may cause probems. Sopes facing directy into the sun wi have higher refectance vaues compared to those on predominanty north facing sopes. This is an effect that is we known on sateite imagery and due to the higher resoution wi be more pronounced on ortho-photographs. Agorithms have been deveoped to counteract these effects, for exampe, topographic effects can be corrected to some extent by using physica radiance modes (Kadmon and Harari-Kremer 1999). Other methods utiise ratioing (normaisation) or regression to suppress bi-directiona refectance effects across an orthophoto image (Hoopainen and Wang 1998). These methods have argey been appied to test sites of sma area extent, for exampe, Hoopainen and Wang (1998) used an area of 300ha. There is therefore some uncertainty as to whether they woud be appicabe over a much arger area extent. A bi-directiona refectance distribution function (BRDF) has to be used in these correction methods. Separate BDRF vaues shoud be cacuated for each and use/cover type. Undertaking this for every pixe in an image is unreaistic. Therefore, common practice uses a standard quantification factor that may over or under correct refectance vaues in different parts of the image (Mikkoa and Peikka 2002). Derivation of the standard quantification factor invoves degrading the spatia resoution of sampe areas from the imagery. For exampe, Mikkoa and Peikka (2002) reduced the image resoutions from 60 90cm in the origina imagery to 2m to even out the shadow effects on individua trees. Hoopainen and Wang (1998) used variabe resoution sampe areas for different and use types. These agorithms are to a arge extent research based. They have not been appied in an operationa context. It shoud aso be noted that these agorithms shoud ony be appied to non-compressed imagery. Compression artefacts as described previousy wi have a detrimenta effect upon automated cassification accuracies. 3.2 Airborne SAR (Synthetic Aperture Radar) Technica description Synthetic Aperture Radar refers to those sensors that iuminate the earth s surface by generating a source of microwave energy. Known as active systems, SARs are capabe of both day and night time operation and because they operate at mm to cm waveengths are reativey itte affected by coud and rain. The aperture defines the band width over which the antenna acquires data and the onger the radar antenna the more waveengths can be samped. Long antennas are difficut to operate in space and SAR techniques empoy the motion of the sensor patform itsef to simuate a onger antenna ength. By convention SAR systems are characterised by a etter designating the waveength (in cm) or frequency (in cyces per second or Hertz) range of the EMR signa generated. The majority of SAR systems operate at X-band ( cm), C-band ( cm), L-band ( cm) or P-band ( cm). 12

21 Increasingy systems depoyed wi be capabe of muti-band rather than singe band data coection, thereby improving the capabiity for discrimination of surface features. An additiona advantage of active SAR systems is the capabiity to coect images with different poarisations of the EMR signa. Poarisation is a property of an EMR wave and with some recent active SAR systems both the signa generated and received can be separated into vertica and horizonta poarisations. Not a systems currenty record poarisation information. Recenty within the UK at east two airborne SAR systems have been depoyed athough these systems differ substantiay in their operating characteristics (Tabe 3.2.1). The German Aerospace Research Estabishment (DLR) experimenta SAR system (E-SAR) is a muti-frequency dua poarisation sensor. The STAR-3i sensor operated by Intermap Technoogies (Canada) is an X-band SAR that is operated in interferometric (INSAR) mode. An airborne interferometric system uses two SARs on the same patform to image the same area from sighty different positions. The resutant data can be processed to derive terrain height information. Additionay image data can aso be generated and under standard operating conditions the STAR-3i system can produce an image with 2.5m pixes and a 10km wide swath. Tabe Typica operating characteristics of seected SAR systems Sensor Wavebands Poarisations Swath width (km) Pixe size (m) E-SAR P, L, C, X VV and HH Star 3i X Review of recent appications From the perspective of monitoring woodand sites this review necessariy focuses on airborne SAR, given that space-borne sensors cannot at present provide images of the required spatia resoution. The capabiity to acquire data under coudy and rainfa conditions provides a significant operationa advantage compared to other passive systems. From a technica point of view, recent airborne sensors can provide images with pixe sizes of better than 5m, with 2m pixes a common operationa specification. Increasingy systems are becoming avaiabe that are capabe of simutaneousy acquiring images at more than a singe waveband and with co- and cross-poarizations, thereby increasing the dimensionaity of the avaiabe data. In addition the deveopment of interferometric processing techniques provides a basis for estimating the height of features as we as providing images. There remain, however, certain technica chaenges specific to the processing of SAR images. SAR pixe data are sensitive to a variety of surface characteristics other than the simpe presence of vegetation and other surface covers. These incude the surface dieectric constant, the oca (ground) surface roughness, and topography (Goya et a., 1999). For vegetated surfaces SAR images are sensitive to moisture content and structure. Where the emphasis is on feature recognition in regions with moderate to extreme terrain variation, the most significant confounding infuence can be the sensitivity to topography. SAR images aso contain a specific source of random noise caed specke that can ead to a 3 4 fod reduction in the actua compared to the nomina spatia resoution. This reduction in the effective spatia resoution wi be especiay significant where ony a singe microwave waveband image is avaiabe for feature mapping. 13

22 There exist a range of studies specific to forested and woodand environments. Those invoving the interpretation of images comprising a singe waveband do not suggest that the mapping accuracies wi be sufficient for the purposes of monitoring woodand sites in Scotand. C-band data in particuar do not adequatey separate grass and forested areas. However, Jager and Benz (2000) report resuts of the cassification of DLR E-SAR data acquired at L-band and VV poarization into five surface casses: smooth surface, ow and medium vegetation, forest, and urban. Accuracies of mapping exceeded 80% for the forest cass. Increasingy use is being made of airborne, muti-band SAR. Dobson et a. (1999) indicate that for forest inventory, muti-frequency SAR data provide generay better discrimination than muti-poarized data. Herad et a. (1999) report accuracies approaching 90% for forest when combining X (HH poarization), C (HH poarization), L (HH poarization), L (HV poarization) and L (VV poarization) wavebands. Recent sensor configurations provide interferometric data products derived from coherence and phase data obtained simutaneousy from two sensors on the same aeria patform. The use of an X- or C-band interferometry SAR (InSAR) appears to provide usefu information on individua trees in the upper canopy (Hoekman and Quiñones, 2000). At the onger waveengths of L and P band data, there is increased penetration of the canopy and therefore the images can provide information reated to ower forest ayers. An approach to mapping the position and crown size of arge (~10 20m) trees within a tropica forest environment was deveoped by Varekamp and Hoekman (2001). Using a C-band interferometric SAR system, data were acquired with approximatey 1m resoution. The image was segmented to identify potentia tree crown segments and the shape of these modeed parametricay using an eiptic Fourier representation of a cosed curve. The approach is nove in departing from a traditiona method of feature recognition and focussing on recognition of individua tree crowns athough no accuracy statistics are provided. In summary airborne SAR systems have been used in a range of forested environments and the avaiabe studies indicate a capabiity to separate forest from other surface categories. Current sensor technoogy is deveoping rapidy and providing muti-band, muti-poarization and interferometry image data. 3.3 Airborne hyper-spectra sensors Technica description Hyper-spectra sensors are designed to capture images with a arger number of wavebands than standard muti-spectra devices. Athough such sensors are capabe of samping in excess of 100 separate wavebands, this can normay ony be achieved for a sampe of pixes across the image swath. In addition, absorption by atmospheric water and dust can restrict the number of bands that contain data suitabe for interpretation. A range of hyper-spectra sensors are currenty in operation and Tabe detais the typica operating characteristics of those systems that have recenty been used in the UK. Athough the number of wavebands potentiay acquired can range between 100 and 200, these systems are programmabe and fewer bands are coected operationay. The Compact Airborne Spectrographic Imager (CASI) deveoped by ITRES Research Ltd. and the Airborne Imaging Spectroradiometer for Appications (AISA) of Spectrum Mapping operate simiary within the visibe and near-infrared part of the spectrum and both are capabe of acquiring more than 200 bands of data. 14

23 The Shortwave Infrared Airborne Spectrographic Sensor (SASI), aso deveoped by ITRES Research Ltd., provides some overap with the CASI sensor, but extends the coection of the data into the short-wave infrared. Tabe Operating characteristics of seected hyper-spectra sensors Instrument Across track Across track Spectra bands Spectra range Dynamic range fied of view ( ) pixes (nm) (bits) CASI CASI SASI HYMAP AISA Eage AISA The Hyperspectra Mapping (HYMAP) sensor deveoped by Integrated Spectronics Pty Ltd covers a wider range of the eectromagnetic spectrum than the CASI and AISA sensors but with fewer wavebands. The pixe resoution and swath width of hyper-spectra sensors is mainy dependent upon the fying height of the aircraft and, for exampe, under typica operating conditions images with pixe sizes of between m and swath widths of between 256m 2km and 740m to amost 6km for the CASI-2 and CASI-3 systems respectivey are obtained (Tabe 3.3.2). Tabe Comparison of CASI-2 and CASI-3 sensors CASI 2 CASI 3 Pixe size Swath width Fying height Swath width Fying height (m) (m) (m AGL) (m) (m AGL) Review of recent appications A probematic issue in the past use of hyper-spectra data has been in respect of the geometric and radiometric properties of the acquired images and identification of an appropriate seection of bands from the hyper-spectra data. Athough techniques are being deveoped to faciitate band seection (Chang and Du, 2004; Chang et a., 1999), the earier discussion does not indicate this to be a major factor of reevance to the current study. From this perspective hyper-spectra sensors provide a technoogy that offers significant promise and some recent studies support this contention. Paterson et a. (2001) focused on demonstrating the abiity of a CASI 2 hyper-spectra sensor to spectray separate a range of tree species common to woodands in centra Queensand, Austraia. Using 14 wavebands and 1m spatia resoution, the study confirmed that six woodand species generated spectray distinct profies of refectance. However, the study aso concuded that the presence of shadow within the woodand images coud provide a imitation in deveoping a 15

24 comprehensive inventory. Zhang and Frankin (2002) obtained 83% cassification accuracy for nine casses in Aberta using 2m CASI imagery. They extracted image spectra data, textura derivatives and a kriging (geo-statistica) surface from mutispectra CASI imagery (2m spatia resoution). 3.4 LIDAR (Light Detection and Ranging) Technica description LIDAR refers to a range of instruments that transmit a series of aser (ight ampification by stimuated emission of radiation) puses towards a target. The signas returned from the ground surface are sensed and changes in the properties of the ight are used to determine one or more properties of the target. The simpest LIDAR instrument uses the time deay of the response to determine accuratey the distance to the feature being imaged. When the instrument is carried on an aeria patform (aircraft or heicopter) and the position of the patform and orientation monitored using GPS and an Inertia Measuring Unit (IMU), then terrain height can be determined. The technique is sometimes caed airborne aser terrain mapping or ALTM. A range of LIDAR systems are avaiabe and severa are currenty depoyed within the UK. Tabe provides an exampe of the typica operating conditions of the Optech 2033 LIDAR, a LIDAR system currenty in use by Infoterra Ltd. (UK). Data are acquired with 1 2m spatia resoution, height accuracies of around 20cm across a swath of around 1km. Tabe Typica operating conditions of Optech 2033 ALTM LIDAR Atitude (m) Swath width (m) Point density (m) Accuracy (cm) Horizonta Vertica ±15 < ±25 < Review of recent appications From the perspective of monitoring woodand sites it is evident that LIDAR is a technoogy that aso offers some promise. From a technica point of view, recent sensors have capabiities of iuminating cross-sectiona areas beow 1m at a wide range of aser point densities that can approach 200 points/m 2, and can provide a range of aser attributes depending upon the capabiity to record mutipe returns. This effectivey provides sufficient spatia detai to resove individua trees. Additionay, whie both passive optica and active microwave systems are ess sensitive to structura properties of dense forests, recent research indicates that LIDAR instruments can estimate important structura attributes within dense forest stands incuding temperate coniferous and deciduous forests (Lefsy et a., 1999a; 1999b) as we as dense tropica forests (Drake et a., 2002). Specific research studies focus mainy on the roe of LIDAR in the measurement of forest and woodand canopy structura properties, in particuar tree heights (Popescu et a., 2002; Zimbe et a., 2003). As noted by Drake et a. (2002) sma-footprint (ie <10m) LIDAR systems have been used in estimating tree heights, percent canopy cover, timber voume and aso forest above ground biomass. For exampe, Naesset (2002) estimated mean tree height, dominant height, mean diameter, stem number, basa area, and timber voume 16

25 of both spruce and pine trees in Norway from various canopy height and canopy density measurements derived from a sma-footprint (average footprint diameter of 21cm) LIDAR sensor. However, it woud appear that these finer resoution systems are best suited to reativey open canopies whereas arge-footprint (ie >10m) LIDAR is better suited to a range of cosed canopy conditions (eg Means et a., 1999). Few studies provide an indication of the possibiity of separating individua tree species using soey measurements from a LIDAR. However, Brandberg et a. 2003) coected data from a aser profier of severa tree species in a deciduous forest in eastern USA. Oaks (Quercus spp.), red mape (Acer rubrum) and yeow popar (Liriodendron tuiperifera) were studied during eaf-off conditions in winter and an attempt was made to cassify these from the LIDAR data using a inear discriminant anaysis. The best singe measure was the maximum aser height and the maximum achieved accuracy of species cassification using a range of aser attributes was 60%. Recent research has combined the use of LIDAR data with the coection of hyperspectra (eg Hi et a., 2002) and sateite imagery (eg Hudak et a., 2002). Backburn (2002) derived a canopy-surface height image using an ALTM 1020 LIDAR with eevation accuracies of 15cm and an approximate 2m spatia resoution together with CASI data for 12 wavebands. The LIDAR eevation data were coected with a view to improving the quantification of pigment concentrations for broad-eaved deciduous and coniferous evergreen forests from the CASI data. Athough there was itte improvement for broad-eaved stands, for coniferous stands, the use of LIDAR data to remove canopy gap areas from the CASI imagery increased the accuracy of CASI derived estimates. Riano et a. (2003) suggest that increasingy LIDAR data wi be used in conjunction with data from other optica sensors as a basis for improved feature type discrimination, either of individua tree species or different surface canopy types (for exampe shrubs from young forest and grassand). In summary LIDAR systems have been depoyed across a range of forested environments mainy with a view to estimating the structura properties of woodand and forest regions. Current sensor technoogy provides a basis for identification of individua trees and the measurement precision adequate for the purpose of separating trees from shrub and grass. 3.5 Sateite-based sensors Technica description The spatia resoution from traditiona sateite remote sensing patforms is generay not sufficient to provide the eve of detai as required by SCM with perhaps the exception of the SPOT 5 sateite. The sensors provide a spatia resoution of 2.5m in panchromatic mode and 10m in mutispectra mode. One major advantage of this system compared with others and aeria photography, is the arge 60km swath that it provides (see Tabe 3.4.1). That coud sti reduce data and processing costs where arge areas need to be monitored. However, SPOT 5 might sti not be abe to achieve the eve of detai required for a sufficient proportion of the required SSSIs. Aongside traditiona sateite remote sensing patforms a new generation of very high resoution earth observation patforms are now being deveoped. This began in September 1999 with the successfu aunch 17

26 of the Space Imaging Ikonos 2 sateite and was foowed by Digita Gobe QuickBird 1 in 2001 and recenty Orbimage OrbView 3 in Ikonos 2 and OrbView 3 have very simiar characteristics with 1m spatia resoution in panchromatic mode and 4m in mutispectra mode. QuickBird 1 has improved spatia resoution with 61cm in panchromatic and 2.44m in mutispectra mode. The wavebands of a three sateites range from bue to near infrared, which is ikey to be an advantage compared with traditiona coour photography and counterbaance the sighter ower spatia resoution. A three sateites possess revisit times as ow as three days. Unfortunatey very high resoution sateite imagery companies have not had a very good reputation among civiian users, particuary in Europe, arising mainy for the high initia cost of the imagery and faiure to suppy data that had been ordered. Athough the costs associated with these data have dropped significanty, probems with data deivery appear to be due to the priority for miitary use at a time of ow system numbers. Whether this wi improve over the short-term with the three sateites in orbit is not yet cear. In addition, there is one other very high resoution sateite, EROS A, in orbit buit by Israe, with data distributed by ImageSat Internationa. It ony has a panchromatic camera, which means that it coud ony suppy back and white images at a standard resoution of 1.8m that can be increased up to 1m. Athough coour imagery is not avaiabe, EROS A coud be a suitabe aternative in the absence of other data sources, with costs from 5/sq km for new acquisition. The EROS B sateite is due to be aunched in 2004 and wi provide imagery with a spatia resoution between Ikonos 2 and QuickBird 1. Finay, the French space agency is aso due to aunch the Peïades consteation which wi provide imagery simiar to that of QuickBird 1 from Tabe Utra-high sateite operating characteristics SPOT 5 Ikonos 2 QuickBird 1 OrbView 3 EROS A EROS B Peïades Orbit 822km 681km 450km 470km 480km 600km 694km Sun Sun Sun Sun Sun Sun Sun synchronous synchronous synchronous synchronous synchronous synchronous synchronous Revisit time <5 days 3 days days <3 days days n/a 4 days Spatia resoution at nadir (m) Mutispectra Yes Yes Yes Yes No Yes Yes Pixe depth 8 bits n/a n/a Swath at nadir (km) Spectra range (µm) Avaiabiity Now Now Now 2004 Now ? Review of recent appications As discussed above the avaiabiity of sateite based sensor with resoution sufficient for SCM is fairy recent and therefore fewer studies have been pubished in the scientific iterature than for the other technoogies reviewed. From a genera perspective, a report commissioned by Digita Gobe (QuickBird suppier) concuded that QuickBird imagery coud be used for standard USGS mapping and it is generay accepted that Ikonos and QuickBird imagery are suitabe for mapping at 1:10,000 scae. 18

27 From the range of imaging technoogies reviewed in this report, utra-high resoution sateite imagery is the cosest to airborne ortho-photography in terms of its technica capabiities. Whie digita aeria photography tends to have a higher spatia resoution by at east a factor of 4, both Ikonos and QuickBird have the advantage of a broader spectra coverage with a near Infrared waveband (usefu to distinguish vegetation types). The key issue is whether the avaiabiity of a near infrared waveband counterbaances the ower spatia resoution for the discrimination of individua trees from the background vegetation. A review of pubished scientific studies on utiising primariy Ikonos data in forest environments suggests main groupings of resuts: Firsty, those studies concerned with using utra-high resoution imagery together with more traditiona (eg SPOT or Landsat) ower spatia resoution sateites; secondy, specific cassification agorithms are deveoped or adapted for the mapping of forests or woodands from Ikonos data; and thirdy, Ikonos is used as a support to fied observations for the extraction of biophysica parameters. In the first category, Ikonos is used over sma areas for the caibration or vaidation of Landsat or SPOT in tropica forest environments where the acquisition of fied observations is particuary difficut. Qi et a. (2002) have used Ikonos imagery to vaidate the identification of seective ogging practices in the Braziian Amazon region from Landsat TM imagery. Souza et a. (2003) impemented a spectra mixture modeing approach to map forest degradation in the Eastern Amazon from SPOT imagery using again Ikonos as a vaidation data set. However, in both cases the authors do not state how accurate the identification of seective ogging or degraded forest is from Ikonos imagery. In another instance (Asner and Warner, 2003), Ikonos data are used to assess the amount of shadow occurring in the Amazon Forest to assess its impact on Landsat TM measurements as part of the Large Scae Biosphere Experiment. In other cases, Ikonos data are compared against other sensors for forest cassification. Surprisingy, Ikonos does not aways perform best compared with ower resoution sensors. The study of Thenkabai et a. (2004) compared the Hyperion hyperspectra sensor, the Advanced Land Imager (ALI), Landsat Enhanced Thematic Mapper (ETM) pus and Ikonos imagery and demonstrated a superiority of the hyperspectra sensor in terms of identification of the African rain forests and use and and cover casses in Cameroon. Nevertheess, Viering et a. (2002) assess the reationship between LIDAR, Ikonos and ground based measurements for the gathering of inventory data of a Ponderosa pine forest and concude that Ikonos can distinguish between tree canopy coverage and exposed understory grasses. In a mountainous region, severa sensors are assessed by Schwartz et a. (2003) for their capabiity at characterising forest damage after a disastrous storm event. The best cassification accuracy is obtained by manua interpretation of Ikonos imagery foowed by an automatic cassification of SPOT 4 imagery. The reative ack of marked differences between the cassification of traditiona sateite imagery and Ikonos data is probaby due to the fact that traditiona cassification agorithms are not adapted to the increased textura information contained in an Ikonos image. Hence, a significant number of studies have concentrated on deveoping cassification techniques specificay based on the textura anaysis of Ikonos imagery. Bugnet et a. (2003) obtained cassification accuracies of around 90% in the automatic mapping of woodands in urban environments from the textura anaysis of Ikonos panchromatic imagery. Perkins et a. (2000) have deveoped a cassification agorithm adapted to feature cassification of Ikonos imagery based on neura networks. Schwartz et a. (2002) make a series of comparisons of traditiona agorithms such as maximum ikeihood against object based agorithms such as the one impemented in the ecognition software package of Definiens for Ikonos and SPOT imagery and find that athough ecognition does not bring any significant advantages for SPOT, the accuracy of the Ikonos derived ecognition cassification of storm osses is significanty higher. 19

28 However, in a reated study, Schwartz et a. (2003) concude that the manua interpretation of Ikonos imagery provides the best cassification resuts. Read et a. (2003) found that the visua interpretation of Ikonos imagery was capabe of identifying individua trees as we as some ogging management features in Centra Amazonia. However, Asner et a. (2002) demonstrates that quantitative information on tree crown dimensions in Eastern Amazonia coud not reiaby be derived from Ikonos imagery. This is by Cark et a. (2004b) who show that Ikonos imagery can provide data on tree size, ocation, mortaity and growth by comparing a 1988 aeria photo with a 2000 Ikonos image of the La Seva Bioogica Station, Costa Rica. Change anaysis is aso performed successfuy by Cark et a. (2004a) to quantify tropica rainforest tree mortaity by comparing a 2000 Ikonos image of the same La Seva Bioogica Station with a 2002 QuickBird image. The above shows that utra-high resoution imagery coud potentiay provide the information required for SCM and that significant advances exist in the deveopment of automatic information extraction agorithms for these data. It shoud be noted however, that many of the pubications reviewed sti represent preiminary research findings for specific environments different to those of Scotand. 3.6 Operationa issues Under the assumption that image data from a specific sensor can meet stated accuracies for mapping woodand areas, the key operationa requirements become focused on cost and coverage. It has ony been feasibe within this study to provide genera indicative costs, since the quoted costs provided by companies often reate to purchase of a singe image and purchase of mutipe images often entais some degree of cost saving. In addition, costs are normay for purchase of compete scenes, rather than for coverage of arge numbers of individua sites. Tabe updates or confirms the recent assessments reported in DEFRA (2003). Tabe Typica costs for specific sensor data Airborne Ortho-photo Hyper-spectra LIDAR SAR Cost ( per km 2 ) Cost (SSSI M)* Sateite Ikonos QuickBird OrbView SPOT-5 Cost ( per km 2 ) Not yet avaiabe 0.60 Cost (SSSI 1000)* Not yet avaiabe 6 * Assuming area of 10073km 2 With respect to coverage, currenty no singe source of sensor data provides fu coverage of Scotand. Digita ortho-photography is expected to be avaiabe in time for the next round of SCM beginning in The utra-high resoution sateite sensors have not yet been in operation for a sufficient time period to provide compete coverage individuay. However, it is expected that this situation wi change in the medium term. Airborne hyper-spectra, SAR and LIDAR sensors are sti operated primariy within a research framework and data coected cover reativey modest study areas. 20

29 4 IMAGE DATA PROCESSING AND INFORMATION EXTRACTION 4.1 Overview Regardess of the source, spatia, tempora and/or spectra resoution of digita image data used, some form of interpretation has to be undertaken for these to provide usefu information. Athough there is no absoute connection between the process chosen to perform this operation and the imagery avaiabe there has historicay been a we-deveoped ink between source and interpretation method chosen. Aeria photographic imagery, with a history of hard-copy usage ong predating computer deveopments, is suited to visua interpretation (API) and as such has traditionay been manuay interpreted and cassified. Athough computer technoogies and GIS have increasingy been expoited for image correction, anaysis and storage, their use has amost excusivey been under-pinned by manua interpretation and digitisation. The more recent deveopment of sateite sensors which generate digita image data directy has occurred in parae within computer image anaysis deveopments. Sateite imagery has therefore been traditionay expoited using automatic cassification procedures. High-resoution sateite muti-spectra imagery is in any event difficut to interpret and cassify manuay to any extent. Automated methods are aso justified in terms of the arge area extent of most studies using these data sources. Reativey recent deveopments such as utra-high resoution sateite imagery and muti-spectra airborne sensors somewhat confound this estabished generaisation. However, in genera the sateite/automated versus aeria photography/manua cassification divide is sti apparent and in particuar automated cassification procedures are rarey appied to the cassification of aeria photography across arge areas. However, the automated processing of high-resoution coour imagery hods considerabe promise. Combined with the ready avaiabiity of ortho-corrected high-resoution aeria photography in digita form the exporation of such computer cassification procedures is perhaps the most significant area of change over the past five years. Such deveopments may provide the basis of a repeatabe method for habitat monitoring. Image segmentation or cassification routines can be considered as ying within two categories depending on the format of the resutant outputs. Pixe cassifiers anayse every pixe in an image, assigning it to a cass using spectra and, sometimes, additiona non-spectra information. The output from the process is a raster of, typicay, the same number of pixes as the origina image. Newer object or vector cassifiers use an initia stage that anayses each pixe, using spectra and maybe additiona information to assign it to a cass, subsequent to this vector objects (poygons) are derived from the boundary extents of continuous bocks of pixes of the same cass. During this process additiona object attributes, such as shape, may be defined to refine the fina cassification. Both techniques have merit in vegetation anayses. However, it shoud be said that currenty reativey few studies of the newer object based cassifiers exist and especiay for the Scottish environment. 4.2 Pixe cassifiers There are two primary strategies for automated pixe cassification from which a variants are derived. These are generay referred to as supervised and unsupervised methods and use either a priori or post priori cass determination respectivey. Both have been used at various times for ecoogica studies. Both use spectra information, ie refectance vaues or DNs for each band. 21

30 Unsupervised cassification identifies natura groupings within the spectra data. These are however unikey to correspond to the desired casses on the ground and hence wi need to be reconcied as best as possibe, a process often invoving the merging of casses. Supervised cassification starts from the constraint that spectra signatures of the desired casses are determined a priori by the use of pixes (training data) known to represent these casses through, typicay, the use of fied observation. Hence unknown pixes are forced into a desired cass. Whie this superficiay seems to offer the best approach it is not without probems. Perhaps the most serious of these are probems associated with unanticipated casses within the area of the study. The reative merits of both these processes are summarised in Tabe Tabe Comparison of cassification strategies Supervised pre-defined casses serious cassification errors detectabe defined casses may not match natura casses casses based on information categories seected training data may be inadequate a priori cass training is time-consuming and aborious ony pre-defined casses wi be found Unsupervised unknown casses no cassification errors natura casses may not match desired casses casses based on spectra properties derived custers may be unidentifiabe post priori custer identification is time-consuming and aborious unexpected categories may be reveaed Unsupervised cassification The objective of cassification is to automaticay categorize pixes in an image into and cover casses or themes. Mutispectra data are usuay used in cassification because different features manifest different signatures based on their spectra refectance and emittance properties. Athough a arge number of different data processing and custering techniques have been deveoped the essence of a unsupervised protocos is the same. Pixes are identified as beonging to groups or custers using spectra information ie the DN vaues for each band considered in the anaysis, hence these are a spectra custering methods. As this process may use information from more than three bands custering can be visuaised as occurring in mutidimensiona space. The main area of variance between differing unsupervised cassifiers is the way in which custers are defined. Hierarchica custering: Either an aggomerative or a divisive process is empoyed. Aggomerative hierarchica custering begins with individua objects and custers that are most simiar first. Successive mergers continue unti a objects are members of one custer. Once the process is compete the points at which custers were merged during the process are examined to decide the most appropriate number and form of custers to be used. Divisive methods are the opposite. An initia group is divided into two custers, the most dissimiar first. Subgroups undergo successive divisions unti there are as many custers as objects. In this case the history of division of casses is used to identify the actua custers to be in the fina cassification. 22

31 Non-hierarchica custering: This is designed to group items rather than variabes and identifies natura custers within muti-spectra space. The number of custers (K) may usuay be specified in advance or as part of the custering procedure. This starts with either initia partition of items into groups, or an initia set of seed points. Many agorithms for unsupervised custering exist, perhaps the most commony depoyed, being part of ERDAS Imagine, is a variant of K-means custering caed the Iterative Sef-Organizing Data Anaysis Technique (ISODATA). Simiar operators are aso avaiabe from other software suppiers such as IDRISI. Once the agorithm is chosen operator contro of the process of unsupervised cassification is normay restricted to determining how many casses are to be initiay separated. Operationay this vaue is generay set higher than might finay be required to aow interactive foding or combination of casses to occur to create a better match with the desired cassification. Whatever process is used during the initia extraction and subsequent reduction of casses, unsupervised cassification requires a post priori interpretation of their on the ground meaning. Hence with this technique the agorithm chosen decides the cass to which each pixe is assigned and the operator then has to decide what ground cass this equates with. This woud usuay be done by reference to ground data. Foowing this process error-checking by use of additiona ground data can be undertaken Supervised cassification Supervised cassification techniques require the definition of a set of training data or spectra signatures for each cass that is to be extracted. The areas chosen for these shoud be representative of the fu range of spectra variances shown by each cass. This is a compex and invariaby time-consuming process. A number of strategies may be used to obtain training data. Tabe Summary of reative merits of pixe cassification strategies Method Advantages Disadvantages Seed pixe quick can underestimate cass variance Interactive poygon good interactive contro can overestimate cass variance sow Digitised poygon can use ground recorded training can overestimate cass variance data extremey accuratey sow Raster ayer quick requires thematic data may be aows iterative processing time consuming as a pre-process Once satisfactory cass signature data are obtained the cassification process itsef can be undertaken. The spectra characteristics of each pixe in the image are examined and compared to the cass signatures. Assignment to a cass is conditiona on meeting the criteria of either a parametric or nonparametric decision rue or agorithm. Parametric rues utiise continuous decision space and hence a pixes are assigned to a cass. Nonparametric decision space contains discrete signature boundaries and hence many pixes may ie outside of these and be uncassified. Many such agorithms exist. 23

32 4.2.3 Sub-pixe cassification Even with the advent of higher resoution image data the probem of mixed pixes sti remains. This is a particuar issue with regard to the form, type and number of casses being extracted and coud be a particuar issue for ecoogica survey. Severa targets of interest are often found within one pixe, however ony a singe category is assigned to each. To address this issue a number of methods have been expored as aternatives to the conventiona hard cassification techniques. In genera they aim at the cassification within one pixe, thus they are termed sub-pixe cassifiers. As an exampe LMM (inear mixture mode) typifies the approach adopted. Starting from the pure signatures for each cass (end members) the spectra refectance of each pixe is then assumed to be a inear combination of the spectra of these end members weighted by their respective area proportions within the pixe (Ichoku and Karniei 1996). Another series of methods that have been expored in an attempt to improve accuracy of cassification, especiay arising from mixed pixes, are those using fuzzy sets. With this concept pixes may have fuzzy membership of more than one cass expressed as the degree of its membership to each cass (vaues range between 0 and 1). The fuzzy cassifiers produce images showing the degree of membership of pixes to stated categories. Training data for fuzzy cassification need not be homogeneous as is desirabe for conventiona hard cassifiers. Commony used fuzzy set based approaches are fuzzy c-means custering (FCM) (Bezdek et a.,1984), the possibiistic c-means custering (PCM) (Krishnapuram and Keer 1996) as we as the method of fuzzy supervised cassification introduced by Wang (1990). A third approach adopted to try to minimise error is the use of neura network cassifiers borrowed from artificia inteigence research. Training data together with a known and-cover cass (the input ayer) are fed into the neura network system (the hidden ayer). Agorithms inside the network try to match training data with the known cass spectra patterns and produce an output ayer together with errors of non-matching neura nodes. The procedure reiterates trying to minimize errors and may repeat severa times. For the cassification of pixes neura networks have proven to be more accurate than conventiona methods (Foschi and Smith 1997). Skidmore et a. (1997) criticay discussed the use of neura network cassification, pointing out that accurate meaningfu resuts require good training data sets; otherwise outputs wi not be very reiabe. The cassification procedure needs the adjustment of various parameters, which increases the compexity of the whoe system and seems to imit its usefuness. Computationa demands of neura network cassification are aso strikingy high. Efficient computing coud ony be reaized on expensive, speciaized parae processing machines. None of the various pixe-based cassification methods seem abe to consistenty deiver reiabe, robust and accurate resuts, whie neura network protocos are compex and ikey to be to sow at this time. 4.3 Object cassifiers Object cassification differs from pixe cassification in one major way, the extraction of vector objects, features or poygons as an end point of the cassification. This aows criteria such as feature size, shape, perimeter ratios etc to form part of the process. Additiona information such as spatia context can ikewise 24

33 be incuded. In this way features such as roads, buidings etc in the image can, to a greater or esser extent, be identified as objects. As such features coud incude, for exampe, a woodand habitat type these techniques appear to offer great potentia. In addition once an object is defined it has the same spatia characteristics of any vector object and hence can be anaysed using GIS techniques and spatia database functions. A number of new software packages offer this form of anaysis incuding ecognition (Definiens Imaging GmbH), Feature anayst (Visua Learning Systems, Inc) and FeatureXTR (Hitachi Software Goba Technoogy). The faciity to extract vector objects from raster imagery without manua digitisation is an important potentia advantage of object based cassifiers. However, the avaiabe evidence is not definitive with respect to the expected accuracy improvements and ease of integration compared to other technoogies. Whist the abiity of these techniques to extract objects is not in doubt, the abiity to extract those of interest to the user certainy is. As with a cassification agorithms it is not the abiity to define casses that is crucia, it is the abiity to define functionay usefu ones. This remains whether the output is pixe or object. The number of confounding objects identified that require extensive time-consuming, post-cassification ceaning needs to be evauated. As yet there appears to have been no particuar attempt to use such techniques for ecoogica mapping of the type required for conservation monitoring. However, these cautions aside, vector cassifiers appear to offer one of the few possibiities for achieving the semi-automated cassification necessary to deiver rapid and cost-effective monitoring of the type required by SNH. As such there is reason to beieve these methods are worthy of further testing against the criteria set by SNH for CSM, in particuar using aeria photography and/or utra-high resoution sateite data. 4.4 Data fusion Data fusion refers to the integration of data from a number of sources to produce a higher vaue or more usefu output compared to use of any singe data source (Wad, 1999). Originay, data fusion invoved the merging of higher resoution panchromatic sensor outputs with spectra data from ower resoution mutispectra sensors to produce pan-sharpened imagery. However, the reaisation that feature cassification accuracy might be improved by using more than spectra information within a singe pixe has ed to the deveopment of more compex ideas about fusion of data from different sources. The majority of recent work has focused on the fusion of mutispectra image data with higher spatia resoution panchromatic images coected on the same patform (Ranchin et a., 2003; Sun et a., 2003). Exampes incude Landsat ETM+ 15m panchromatic with 30m muti-spectra, SPOT 5 HRG 5m panchromatic with 10m muti-spectra, athough fusion between different sensors on different patforms is aso possibe, for exampe aeria photography with Landsat TM (Schetseaar, 2001) and Landsat TM with IRS panchromatic (Teggi et a., 2003). In the majority of data fusion appied to remotey sensed data the main aim is to obtain a muti-spectra image with enhanced spatia detai whie preserving the radiometric (equivaent to the coour baance) properties of features within the image. A variety of agorithms have been appied to this genera probem (Tapiador and Casanova 2002 and 2003), and new agorithms deveoped (Laporterie and Fouzat, 2003; Wu et a., 2004), with the aim of overcoming the recognised probems of coour distortion and dataset and operator dependency (Zhang, 2002). 25

34 Evidence exists that cassification can be improved by use of fused panchromatic and muti-spectra data. Sun et a. (2003) integrate spectra, spatia and structura information and demonstrate improved and-use cassification compared to conventiona methods based ony on muti-spectra data. Ray (2004) compared on basis of cassification accuracy six different methods of data merging using Indian Remote Sensing Sateite (IRS) panchromatic and Linear Imaging Sef-scanning Sensor (LISS) III muti-spectra data for a predominanty agricutura area. Fusion of muti-tempora data can aso improve cassification accuracies. Mohan et a. (2000) used fuzzy ogic operators to combine the cassification resuts of two dates of IRS-1A LISS II images and achieve an overa accuracy of cassification of 95%. Megani and Serpico (2003) demonstrate a method of integrating tempora contextua information into the cassification process and appy this to muti-tempora, muti-sensor data. Lombardo et a. (2003) describe a method for the fusion of muti-tempora singe waveband synthetic aperture radar (SAR) images with a singe muti-spectra image. Ufarsson et a. (2003) appy data fusion and feature extraction to muti-sensor and geographica data and compare the resuts with existing feature extraction methods. In respect to the specific monitoring of woodands, McCombs et a. (2003) demonstrated an improved capabiity for tree identification accuracy and mean height estimation by fusing LIDAR and 0.61m 4-waveband muti-spectra data compared to use of the data separatey. However, methods of data fusion increasingy encompass other types of remotey sensed data, anciary information about the tempora evoution of and cover (Largouet and Cordier, 2001), post-cassification methods (Megani, 2004; Tapiador and Casanova, 2003) as we as cassifiers (Kumar et a., 2002; Smits, 2002). The existing research indicates an increasing capabiity to merge data and cassification approach to produce improved accuracies of cassification. However, at the current time the focus appears to be on deveopment of method and few exampes provide cear indications of equivaent performance for argescae monitoring of woodand environments. 4.5 Use of cassifiers for vegetation mapping When interpreting resuts from the appication of cassification protocos on differing sources of digita data a number of points shoud be given consideration. Pubished iterature invariaby represents academic exporation of methodoogies, as such, the successes reported may represent considerabe effort, often of months, on a singe scene. Sedom have these been repeated in a manner to mimic an actua use of the method in a work-pace situation. Derived casses are often far fewer, broader and generic than woud be usefu for SCM. Generay the fewer casses resoved the greater the accuracy so this wi infate success compared to what might be anticipated from utiising the same method for such a task. Reported accuracies are frequenty overa accuracies ie an average accuracy across a casses. Hence they are open to being infated by one or two casses of high accuracy but no interest in the scene (eg water). It is therefore possibe the casses of interest are ony resovabe at far ower accuracies. 26

35 4.5.1 Unsupervised custering Thomson et a. (1998) appied unsupervised custering to 4m resoution CASI imagery and ground radiometry to distinguish 10 casses of intertida habitat over an area of 3km 2. A comparison with supervised maximum ikeihood cassification reveaed that unsupervised custering performed amost as we, but was far ess abour intensive. This method was further refined (Thomson et a. 2003) by incuding Airborne Thematic Mapper (ATM) data in the process. Using unsupervised custering (ERDAS Imagine ISODATA agorithm) to produce 50 casses initiay that were subsequenty merged to 13, overa accuracy reported was 70%. It shoud be noted however that of the 13 casses, ony five referred to vegetated habitat so eight of the casses were reativey simpe in structure. Up to 78% accuracy was reported by Backburn and Miton (1997) using unsupervised custering on 2m resoution CASI data from a 10km 2 area of the New Forest. The anaysis was preceded by principa components anaysis (PCA) to provide information on canopy structure used to obtain eight casses, which were described by comparison with a co-registered aeria photograph. The number of casses was reduced to two, canopy and gap, for investigation of gap properties within the study area. The appication of spatia statistics were used to describe the ecoogica properties of woodand of various ages within the study area. A possibe way to extract textura information from a scene in a form that might assist cassification is provided by Song and Woodcock (2002) who used unsupervised cassification of 1m resoution IKONOS panchromatic imagery of forest canopy with the purpose of estabishing successiona phase by remote sensing. Two casses were derived representing sunit canopy and shadow. Semi-variogram anaysis was then used to estabish the scae of variation in the scene, and this was reated to predictions of the spatia characteristics of forest canopies derived from a disc scene mode of forest canopy spatia characteristics, a canopy structure mode, and a canopy refectance mode. Areas of od growth, mature and young forest stands were separabe on the basis of variogram si height (differences in spatia variance of refectance properties of canopy). Od growth has the highest si with young growth the owest, representing increasing heterogeneity with stand age. These differences are most pronounced with pixes of 4m, rather than 1m or 15m. It is ikey however, that the non-uniform age-structure of most woodand woud severey hamper the use of this method. Townsend and Wash (2001) report the use of tempora series of Landsat TM images throughout a singe year to expoit the phenoogica variabiity of forest communities for generating a and cover cassification over a wide geographica scae. Unsupervised hierarchica custering was used to generate the cassified images of the forest community. A fuzzy cass membership approach was used which cosey mimics the natura variabiity and transitiona nature of the communities in the study area. The muti-tempora approach gave accuracies, according to the fuzzy cassification scheme, of 92%, though this measure is not comparabe to traditiona estimates of accuracy Supervised, maximum ikeihood cassification Kadmon and Harari-Kramer (1999) achieved an overa accuracy of 0.8 (r 2 ) (three derived casses) using a supervised maximum ikeihood cassifier on panchromatic aeria photography. The pixe size was 0.75m and the work undertaken as part of a tempora tree cover change study in a 5km 2 area of Mediterranean maquis. 27

36 Gong et a. (2003) used maximum ikeihood cassification with a variety of image processing techniques on airborne muti-spectra imagery with the overa aim of distinguishing oakwoods from vineyards. The most successfu cassification, as determined by t-tests of the Kappa coefficients, was based on the use of a texture image of homogeneity obtained from the near infrared image band, NDVI and brightness generated through orthogonaization anaysis. This method averaged an overa accuracy of 81% for six frames of images tested. By using post-cassification morphoogica processing, the overa accuracy was increased to 87%. Supervised maximum ikeihood cassification of 4m resoution CASI imagery of coasta reefs and seagrass beds was tried by Mumby and Edwards (2002) as part of a study to test the accuracy and cost-effectiveness of various data sources for thematic mapping over an area of 60km 2. Of the instrumentation tested CASI was the most successfu at determining 13 fine-habitat casses, with an overa user accuracy of 81%. As a hyper-spectra instrument, Mumby and Edwards (2002) suggest that CASI may be more reiabe than sateite instruments when a arge number of casses are to be distinguished. Thomson et a. (2003) aso used 4m CASI data from a coasta environment (satmarsh) to map 10 casses of intertida biotopes, with the aim of ong-term measurement of coasta processes by remote sensing. Spectra signatures of vegetation were obtained by ground radiometry using photographs of 0.4m quadrats. Overa correspondences ranged between 74% and 91%, athough it is not cear which figure appies to supervised cassification and which to unsupervised. Of particuar pertinence to this review Brown et a. (2003) used supervised maximum ikeihood cassification on 2m pixe CASI imagery with data merging from LIDAR on a variety of coasta habitats: satmarsh, shinge, mudfats, saine agoons, and sand dunes, with the aim of deveoping remote sensing techniques for marine SAC monitoring, incuding vegetation change. The addition of LIDAR improved cassification accuracy. The overa accuracy using maximum ikeihood cassification was reported at 66%. However, it shoud be noted that whie the accuracies of cassification of mud and water were extremey good, identification of many actua vegetated zones such as the Atripex community, for exampe, is reativey inaccurate. In an interesting attempt to ink the theories and practice of cassification used by remote sensing speciaists and ecoogists Thomas et a. (2002) used supervised maximum ikeihood cassification on CASI imagery, together with TWINSPAN and canonica anaysis to define ecoogica communities on the ground, over a 5km 2 area of northern peatand. It was hoped that ordination/custering using a Two-way Indicator Species Anaysis (TWINSPAN) and correspondence anaysis woud give a natura set of groupings into which the image cassification woud fa. However, spectra separabiity performed better. TWINSPAN produced a cassification accuracy of 32% (kappa), and correspondence anaysis 40%. The probem appears to be at east party the difficuty in defining discrete communities on the ground in what is more propery seen as a muti-dimensiona continuum. Mumby and Edwards (2002) performed a comparison of the abiity of a variety of remote sensing patforms to distinguish fine community casses in subittora habitats. IKONOS with a simiar number of wavebands to oder sateite sensors ike those on Landsat and SPOT (see beow), but a much better resoution, coud be expected to perform best. Indeed a supervised, maximum ikeihood cassification with this imagery achieved a 41% accuracy in separating 13 communities. Interesting from the perspective of conservation monitoring, they concude that IKONOS coud detect the edge of habitat patches quite we, but coud not categorize the habitats correcty. Images coud however be cassified into four broad habitat casses with 75% accuracy. Landsat and SPOT imagery, using the same supervised maximum ikeihood cassifier performed even more 28

37 poory in fine-scae habitat cassification achieving 15 30% accuracy, with the four broader habitats this was better with between 30 60% being achieved. One important point was beieved to be that that the scae of heterogeneity within broad habitats was ess than the resoution of the image. The measure of texture used to improve cassification represented inter-habitat heterogeneity rather than intra-habitat heterogeneity. This study is perhaps a good test of cassification accuracy using the type of habitat casses most ikey to be assessed during CSM ie not very spectray distinct rather than the more usuay tested broad habitat using medium resoution imagery. Luque (2000) used supervised maximum ikeihood cassification, foowed by a 3x3 pixe smoothing agorithm and threshoding, on a time series of images taken over 19 years. Images of 450,000ha of a forested andscape were cassified into seven rather broad casses to assess community changes over the 19 years of the images. Radiometric normaisation was used to detect changes over the sequence of images, and accuracy was assessed by spot-comparison with expert-interpreted aeria photography. Accuracy is described as 90%, but this does not aow for the confounding effect of the same error rate over a series of images. Taking this into account, accuracy woud be coser to 66%, which is quite ow considering the measured change over the duration of the images (8%). Goud (2000) used supervised maximum ikeihood cassification of Landsat TM imagery, couped with normaised difference of vegetation index in the Hood River region of Centra Canada to provide estimates of regiona variation in pant species richness. The estimates of species richness agree to within +/ 15 species over 60% of the area and +/ 30 species over 93% of the area. However, when one considers that the species richness of an area varied between 69 and 109, estimations of an accuracy of +/ 30 cover the entire range of observed vaues. Sun et a. (2003) appied a maximum-ikeihood cassification to mutispectra imagery from IRS. Probabiistic reaxation was then performed on the thematic map to refine the cassification with neighbourhood information. Edges extracted from the higher resoution panchromatic imagery were incuded in the cassification. A modified region-growing approach was used to improve image cassification. The cassification accuracy improved when panchromatic imagery was used for edge-detection to 75% (kappa) from 52% for unrefined maximum ikeihood cassification of the mutispectra data Spectra separabiity Mason et a. (2003) used endmember mutispectra anaysis for remote sensing of skyark habitat correates on farmand at a fine scae and more broady (5km 2 and 140km 2 ) with ATM. By considering ony the pure cases (endmembers) of bare ground or competey vegetated pixes, using red and IR bands, they were abe to generate measures of %cover of vegetation. Vegetation height was derived from LIDAR, which was used to segment the area under consideration into poygons. Texture was measured using standard deviation of pixes in a 10m window. These methods combined gave an accuracy of 80% (r 2, correation of measured against estimated vegetation height); cassification accuracy ranged from 63 80%. One of the important ecoogica correates of skyark habitat was crop height. The measurement of this was subject to an average error of c.15cm. AVIRIS imagery with a resoution of 2m was used by Dennison and Roberts (2003) to suppy data for mutipe endmember mixed spectra anaysis (MESMA) to investigate the effect of seasona changes on species mapping. The pure components of mixed spectra are identified using EAR (endmember average root mean square error). Caibration data was taken from spectra anaysis of pure cass poygons at east 40x40m. 29

38 The method was moderatey successfu with 61 70% (kappa), but ony identified c.50% of pixes to a particuar cass, eaving the rest undefined. There is considerabe between-year variation in endmember spectra, seeming in most cases to exceed variation between species. To overcome this Dennison and Roberts used dua endmembers but images show arge variabiity in the distribution of ong-ived perennias variation ony due to mode inaccuracy. Aspina (2002) aso used aeria hyperspectra sensor data to map the distribution of three species of Popuus in a riparian area of c.25km 2. Aspina used a variety of spectra separation caed mixture tuned matched fitering (MTMF). Vaidation used ogistic regressions to turn matched fiter outputs into probabiity of presence, and the three species of Popuus are separabe at c.90% probabiity. This procedure maps the probabiity of presence of individua species, rather than mapping discrete areas of habitat/community. Using a ogistic regression provides a continuum of probabiity rather than a definitive yes or no, which may more accuratey refect the way that communities grade into one another in the fied. Shanmugam et a. (2003) assessed remote sensing techniques for mapping dune communities to NVC cassification by spectra separabiity using Transformed Divergence. Spectra were obtained by CASI. Sub-pixe cassification was performed with a inear mixture mode using constrained east squares and a fuzzy-c means mode with different weight matrices: Eucidean, Diagona and Mahaanobis Norms. The site had aready been surveyed on the ground into 26 NVC categories, but it was ony possibe to reiaby separate casses if the number was reduced by merging sub-nvc categories. Accuracy ranged from 50 92%, depending on methodoogy and number of casses. Thomas et a. (2002), as we as using TWINSPAN and canonica anaysis with supervised maximum ikeihood cassification, aso used spectra separabiity to define casses. Indeed, the Jeffries-Matsusita measure of spectra separabiity outperformed the other two methods. In other words, casses grouped on the basis of ow spectra separabiity more accuratey refected what was actuay present on the ground than groupings based on ecoogica characteristics of the communities. However, accuracy was sti ow: 55% when combining a three techniques Artificia neura networks Brown et a. (2003) in their study of remote sensing techniques for marine SAC monitoring (see above), used a muti-ayer perceptron in addition to the supervised maximum ikeihood method of cassification. The neura network was 10% better than simpe maximum ikeihood cassification (76% against 66%). Shanmugam et a. (2003), in addition to the two sub-pixe modes (see above, under spectra separabiity), aso appied a Muti-Layer Perceptron Neura Network with a back propagation agorithm. Foody and Cuter (2003) used a feed forward neura network to estimate biodiversity indices in Landsat TM imagery at a scae of 225km 2 in a Bornean forest to assess the use of remote sensing to measure biodiversity. A Kohonen sef-organising feature matrix was used to cassify the resutant data, and discriminant anaysis to determine cass separabiity. A high eve of agreement was reported (96%), but the ony diversity index that gave a correation between fied and remotey-sensed data was Shannon entropy; indices ike species richness and Hi s ratio were not correated. Muchoney and Straher (2002a) used the Gaussian ARTMAP artificia neura network agorithm with AVHRR on an extensive scae (2*106km 2, 1km pixe size). Cassification accuracy was 71% for pixe-based cassification. Poygon-based cassification was aso investigated (see beow). 30

39 4.5.5 Poygon-based cassifiers It shoud be noted that feature or object cassifiers are reativey new and as such have not been expoited to any great extent in situations anaogous to those envisioned for SNH appications. Muchoney and Straher (2002b) compared a poygon based cassification with a pixe-based approach. The pixe-based cassification was more accurate (71% against 49% agreement). However, in the resuts it was apparent that the pixe-based scheme overstated accuracy due to spatia correation, whie the poygon-based scheme understated it uness a subpopuations of cass types are represented in the training data. It is aso worth noting that the cassification scheme used 42 features for cassification, many of which were co-inear. Burnett et a. (2003) used ecognition and caimed it was more robust than pixe cassifiers. Burnett and Baschke (2003) appied an object based approach to andscape anaysis. However, none of these can be said to present a significant improvement in accuracy over pixe techniques. 31

40 5 CONCLUSIONS AND RECOMMENDATIONS Surveying vegetation for conservation purposes presents a number of generic issues that must be considered when deveoping a monitoring methodoogy. These are discussed in the foowing sections and provide the context for the concusions and recommendations made. 5.1 Identification of trees and woodand boundaries Community and habitat cassification types are frequenty artefacts imposed on essentiay continuous phenomena for the convenience of the ecoogists rather than organisms. They rarey form discrete units in a andscape, but graduay merge as species are sowy repaced by others, an ecocine, eg the transition of banket bog and upand heath priority habitats. Even where reative contributions of a particuar species is part of the definition, eg proportion of Ericaceae etc in dwarf-shrub heath broad habitat cass, the scae at which these proportiona assessments are to be made is sedom cear. It is often difficut therefore, or even impossibe, to draw objective, non-arbitrary boundaries between different named communities, even when the observer is in the fied. These issues remain in the remote sensing domain and are a particuar probem with a forms of manua interpretation. For monitoring and reporting area attributes of specific features the definition of boundaries becomes crucia, as does the abiity to detect changes in them over time. If boundaries are probematic to determine it foows differing observers wi record them in differing positions, even on the same site. Hence the crucia issue in tempora comparison is whether changes in vegetation, or changes in observer practice, are being detected. Discussion of the roe of any method of mapping community or habitat types must consider these issues if reiabe data are to be gathered and any proposa for a remote sensing soution shoud seek to repace such judgements with automatic repeatabe methods. In the specific case of woodand extent monitoring, these probems are epitomised by the simpe question, What is a wood? Remote sensing methods for identifying trees and tree extent shoud be possibe and therefore aso the area of trees. However, this is sedom going to be the same as assessing the extent of a wood. Woods may exist as discrete entities within a we defined boundary, in particuar where they abut agricutura and and the wood is marked with a wood-bank, fence, ditch ine or road. In this case mapping the extent of a woodand edge within a GIS by using other andscape criteria is possibe. Subsequent assay of tree cover within that enveope wi aso be practica. The term wood however is sedom used to describe a andscape that consists of tota canopy cover, and few woodand SSSIs wi exist in this state. Instead woods consist of trees at varying densities, with differing sized gaps between them. In a naturay wooded andscape a discrete edge is aso rarey apparent, instead trees become increasingy ess dense, the ratio of gap to canopy increases, unti a that remains are individua trees standing within an open andscape of other community types. Identification of the boundary of such a woodand ie where a wood ends, is ceary not a trivia probem and setting the task for an ecoogist to decide where it occurs in the fied woud be far from easy. In reaity then woodand is covered by trees with a continuum of densities, from cose stands to, in the extreme case, wood-pasture where they are very much the minority component in the wooded andscape. Even where a discrete edge may be visibe, say at an ecotone, the issue of where to draw a boundary eg at the trunk, canopy edge, etc sti exists. 32

41 In summary, a remote sensing method for identifying trees, which are recognisabe entities, is most ikey. The possibe deineation of areas of woodand, whose extent is more determined by perception, history or ega definition than by vegetation type, wi be far more probematic. 5.2 Additiona issues In conducting this review the foowing issues have aso been noted to be important: Athough the exampe case presented here is for woodand, SCM wi require simiar data to be gathered for other habitats and sites, hence a genera method that coud be used for monitoring more than just woodand woud be beneficia. Future methods shoud be as rigorous and fixed as possibe to ensure both ong-term compatibiity and to remove probems associated with identifying changes in practice rather than changes in vegetation. However, they shoud be fexibe enough to aow expoitation of new data sources as these become avaiabe in the future. Revised monitoring methods wi have to be depoyed in a way that ensures they are as compatibe as possibe with baseine data from previous reporting rounds, or that, at a minimum, previous data can be adapted for use within a new protoco. It may therefore be necessary, as a prerequisite to commencing future cyces of monitoring using a new protoco, for existing data from previous surveys, monitoring, management pans etc to be integrated to form a database of baseine data. Data acquired today wi become the ony source of estimation of past species/habitat distributions in the future. Therefore we shoud aso seek to ensure that the monitoring process used is adaptabe to foreseeabe future requirements, and that the method of deriving area estimates is as fixed, repeatabe and transparent as possibe. More importanty data shoud be coected in ways that enabes ready re-anaysis in the future if required. Any new method shoud become a fixed standard for the future thereby ensuring a high degree of compatibiity through the years to aow ong-term changes to be identified. Such a method for monitoring shoud aso seek to be integra to inventory mapping programmes, routine agency reporting activities, management panning etc to ensure ong-term compatibiity between datasets. The ease of integration of any recommended techniques is an important issue since not a SNH can be expected to be trained extensivey in any given method. However, the nature of the materia reviewed is not suited to carification of this issue. 5.3 Avaiabiity of remote sensing data to meet SNH area target monitoring requirements The interaction of pixe size, boundary effects and cassification accuracy in determining the feasibiity of deriving estimates of area at a scae appicabe to Scottish terrestria sites has been assessed in this review. The resuts presented are reveaing and show ceary that for bocks of habitat of the size of interest for SNH, the earier generations of remote sensors are simpy not adequate. Indeed the majority of the existing iterature on the appication of remote sensing techniques to vegetation survey and monitoring are generay deaing with sensor resoutions at east ten times too coarse to resove reaisticay the targets set for most SNH habitats. For exampe one 30m pixe, (representative of typica high resoution sateite systems) represents approximatey 5% of a 2ha site. 33

42 In respect to sensor data therefore, airborne systems producing digita ortho-photographs, hyperspectra or muti-band SAR data can provide suitabe data for woodand monitoring. The ikey avaiabiity within a twoyear time frame of ortho-photography at 50cm resoution is significant in this context. However, the data coverage of both muti-band SAR and hyperspectra images are insufficient for the purposes of woodand monitoring and the costs of acquisition and processing are currenty prohibitive. These data coud be augmented by the use of data from utra-high resoution sateite systems such as IKONOS or QuickBird where coverage exists. 5.4 Avaiabiity of cassification routines to meet SNH area target monitoring requirements Within the context of the earier comments, it is important for the purposes of this review to separate the issue of accurate cassification from that of consistent boundary determination, especiay for compex woodand sites. With respect to cassification methodoogies and the abiity to correcty identify the main feature of interest, trees, the reported accuracies of most automated cassification and segmentation agorithms are typicay <80% and this is usuay ony achieved with broad generic casses. Where attempts to differentiate casses of the compexity more reevant to conservation monitoring, CSM, priority habitat reporting etc are made, accuracy is typicay beow 50%. This is ceary insufficient for the purpose of conservation monitoring and somewhat beow that achievabe using manua interpretation by trained interpreters. The new generation of object cassifiers appear to hod considerabe promise and they are currenty the subject of much research. In particuar, the use of segmentation agorithms might provide an advantage over pixe based cassifiers in that the parameters of the segmentation shoud be abe to be extended across image scenes. However, there remains the issue of cassification based upon the ambient DN signature and at this point, these techniques have not been shown in the iterature to yied consistenty more accurate, robust and repeatabe cassification resuts. This may possiby be because very few studies have yet expored their use fuy using casses typicay assessed for conservation survey. 5.5 Summary of comments, suggested strategies and techniques Summary comments To identify overa changes in area extent of woodand SSSI (in this case 5%) wi require the abiity to discern changes in the area of stands of trees that are far smaer than the tota extent of the SSSI. The traditiona high resoution sateite sensors are not capabe of satisfying the stated monitoring requirements. The ony source of digita image data with the both the coverage and resoution ikey to achieve the monitoring requirements over the next two CSM rounds wi be ortho-corrected aeria photography. The advantages of this form of digita data over traditiona hard copy non-rectified photography shoud not be under-estimated. With further testing, it is aso possibe that utra-high resoution imagery may be usabe to assess SNH targets over arger areas, athough coverage remains an issue. In the future hyper-spectra imagery may become useabe, cost effective and avaiabe over arger areas. If so this shoud bring associated benefits in cassification accuracy. 34

43 There is itte evidence that any commony depoyed, or reported, method of automated image cassification or segmentation can achieve the accuracy required to meet the stated targets uness imagery of <1m resoution is utiised. Visua interpretation methods have been demonstrated to be abe to achieve on average the required eves of interpretation. The potentia of the new generation of poygon cassifiers to match this has yet to be tested. Boundary determination is an important issue that remains to be resoved. Consistency in determination requires that rigorous quaity assurance procedures be foowed. This appies whether the method of cassification is manua or digita. Procedures based upon the definition of a minimum mapping unit and tree density provide a basis for impementation using a visua interpretation method. However, the recent research suggests a number of aternative approaches might be empoyed to provide a more automated method of boundary pacement Summary strategies and techniques Reiabe tempora comparison wi ony be achieved by approaches to monitoring that concentrate on compatibiity of outputs, irrespective of data source, and a future forms of survey shoud be judged against this abiity. Such a strategy shoud be fexibe enough to take advantage of new technoogies as they become avaiabe whie stabe enough to aow vaid statistica comparison with earier data. Pre-operationa assessment of any change to the existing system shoud be undertaken and shoud be conducted across a range of suitabe sensor data and physica environments so that the resuts can indicate the eve of repeatabiity and robustness of the approach. This shoud be assessed against agreed targets. The resuts of this review suggest that the ony system depoyabe over the very short-term, and ikey to achieve the criteria and standards required by SNH CSM requirements, woud be a monitoring programme based on a we-executed and rigorous visua interpretation using ortho-photography. Such a system woud need under-pinning by defined procedures of interpretation and make use of the recent deveopments in on-screen digitising and visuaisation. This shoud be centraised for the initia interpretations and the resuts sent to fied officers for fina confirmation. SNH comments about such a system are acknowedged and a range of techniques, recenty deveoped and subjected to imited testing, may be abe to address specific points about time and cost constraints. In addition they hod promise in their potentia to address issues of arbitrary boundary pacement. These techniques can be combined and coud feasiby be used in a more automated system of monitoring. In particuar the foowing components coud be investigated: a hybrid image samping protoco using geostatistica anaysis to give confidence predictions to boundary estimates; an object or vector based cassifier using casses of interest to a conservation monitoring programme. It is aso feasibe that these techniques wi provide an approach that can be integrated within a system of visua interpretation and provide a degree of semi-automated interpretation. It must be noted that these techniques have not to date been vaidated as part of any system of nationa monitoring and therefore a monitoring system cannot be impemented now using these techniques operationay. However, there is sufficient evidence to suggest that these techniques shoud be investigated further and coud form part of a mid to onger-term strategy. 35

44 The increasing avaiabiity of sateite borne imagery with utra-high spatia resoution and muti-spectra capabiities may provide a suitabe aternative to the use of ortho-photography as a data source. These shoud be incorporated into any system that is deveoped. It is aso suggested that as part of deveoping a monitoring strategy the improvement of fied methods to enabe accurate spatia data coection shoud form part of a new surveys. In particuar it is fet a ecoogica data coected by SNH in the future shoud be coected in ways that enabe integration with such a strategy. 36

45 6 REFERENCES Asner, G.P. and Warner, A.S. (2003). Canopy Shadow in Ikonos Sateite Observations of Tropica Forests and Savannas. Remote Sensing of Environment 87, Asner, G.P., Paace, M., Keer, M., Pereira, R., Siva, J.N.M. and Zweede, J.C. (2002). Estimating Canopy Structure in an Amazon Forest From Laser Range Finder and Ikonos Sateite Observations. Biotropica 34, Aspina, R.J. (2002). Use of ogistic regression for vaidation of maps of the spatia distribution of vegetation species derived from high spatia resoution hyperspectra remotey sensed data. Ecoogica Modeing 157, Bezdek, J., Ehrich, R. and Fu, W. (1984). FCM: the fuzzy c-means custering agorithm. Computers and Geosciences 10, Backburn, G.A. and Miton, J.E. (1997). An ecoogica survey of deciduous woodands using airborne remote sensing and geographica information systems (GIS). Internationa Journa of Remote Sensing 18, Backburn, G.A. (2002). Remote sensing of forest pigments using airborne imaging spectrometer and LIDAR imagery. Remote Sensing of Environment 82, Bowman, D.M., Wash, A. and Mine, D.J. (2001). Forest expansion and grassand contraction within a Eucayptus savanna matrix between 1941 and 1994 at Litchfied Nationa Park in the Austraian monsoon tropics. Goba Ecoogy and Biogeography 10, Brandtberg, T., Warner, T.A., Landenberger, R.E. and McGraw, J.B. (2003). Detection and anaysis of individua eaf-off tree crowns in sma footprint, high samping density LIDAR data from the eastern deciduous forest in North America. Remote Sensing of Environment 85, Brown, K., Hambridge, C. and Matthews, A. (2003). The deveopment of remote sensing techniques for marine SAC monitoring. Engish Nature Research Reports, No. 552, Engish Nature. Bugnet, P., Cavayas, F. and Gagnon, L. (2003). Automatic Mapping of Woodands in Urbanized Environments Based on Panchromatic Ikonos Images Case of the Montrea Metropoitan Region. Canadian Journa of Remote Sensing 29, Burnett, C. and Baschke, T. (2003). A muti-scae segmentation/object reationship modeing methodoogy for andscape anaysis. Ecoogica Modeing 168, Burnett, C., Aaviksoo, K., Lang, S. and Langanke, T. and Baschke, T. (2003). An object-based methodoogy for mapping mires using high resoution imagery. Abstracts: Ecohydroogica Processes In Northern Wetands Tainn, 30 June 4 Juy Cameron, A.D., Mier, D.R., Ramsay, F., Nikoaou, I. and Carke, G.C. (2000). Tempora measurement of the oss of native pinewood in Scotand through the anaysis of orthorectified aeria photographs. Journa of Environmenta Management 58, Chang, C.I. and Du, Q. (2004). Estimation of number of spectray distinct signa sources in hyperspectra imagery. IEEE Transactions on Geoscience and Remote Sensing 42,

46 Chang, C.I., Du, Q., Sun, I-L. and Athouse, M.L.D. (1999). A Joint Band Prioritization and Band- Decorreation Approach to Band Seection for Hyperspectra Image Cassification. IEEE Transactions on Geoscience and Remote Sensing 37, Cark, D.B., Castro, C.S., Avarado, L.D.A. and Read, J.M. (2004a). Quantifying Mortaity of Tropica Rain Forest Trees Using High-Spatia-Resoution Sateite Data. Ecoogy Letters 7, Cark, D.B., Read, J.M., Cark, M.L., Cruz, A.M., Dotti, M.F. and Cark, D.A. (2004b). Appication of 1-M and 4-M Resoution Sateite Data to Ecoogica Studies of Tropica Rain Forests. Ecoogica Appications 14, DEFRA (2003). UK Emissions by Sources and Removas by Sinks due to Land Use, Land Use Change and Forestry Activities. DEFRA Contract EPG1/1/ pp. Dennison, P.E. and Roberts, D.A. (2003). The effects of vegetation phenoogy on endmember seection and species mapping in southern Caifornia chaparra. Remote Sensing of Environment 87, Dobson, M.C., Pierce, L.E. and Uaby, F.T. (1999). Evauation of SAR Sensor Configurations for Terrain Cassification and Forest Biophysica Retrievas. Drake, J.B., Dubayah, R.O., Cark, D.B., Knox, R.G., Bair, J.B., Hofton, M.A., Chazdon, R.L., Weishampe, J.F. and Prince, S.D. (2002). Estimation of tropica forest structura characteristics using argefootprint LIDAR. Remote Sensing of Environment 79, Dunn, C.P., Sharpe, D.M., Guntenspergen, G.R., Stearns, F. and Yang, Z. (1990). Methods for anaysing tempora changes in andscape pattern. In: Quantitative Methods in Landscape Ecoogy (M.G. Turner and R.H. Garner, Eds.), Springer-Verag, New York, Engish Nature (2002). Objective setting and condition monitoring within woodand Sites of Specia Scientific interest. Engish Nature Research Reports 472. Engish Nature. Peterborough. Foody, G.M. and Cuter, M.E. (2003). Tree biodiversity in protected and ogged Bornean tropica rain forests and its measurement by sateite remote sensing. Journa of Biogeography 30, Foschi, P.G. and Smith, D.K. (1997). Detecting sub-pixe woody vegetation in digita imagery using two artificia inteigence approaches. Photogrammetric Engineering and Remote Sensing 63, Gong, P., Maher, S.A., Biging, G.S. and Newburn, D.A. (2003). Vineyard identification in an oak woodand andscape with airborne digita camera imagery. Internationa Journa of Remote Sensing 24, Goud, W. (2000). Remote sensing of vegetation, pant species richness, and regiona biodiversity hotspots. Ecoogica Appications 10, Goya, S.K., Seyfried, M.S. and O Nei, P.E. (1999). Correction of Surface Roughness and Topographic Effects on Airborne SAR in Mountainous Rangeand Areas. Remote Sensing of the Environment 67, Green, D.R., Cummins, R., Right, R. and Mies, J. (1993). A methodoogy for acquiring information on vegetation succession from remotey sensed imagery. In: Landscape Ecoogy and GIS (R. Haines-Young, D.R. Green and S. Cousins, Eds.), Tayor and Francis, London,

47 Herod, M., Hochschid, V., Kenke, M. and Müer, A. (1999). Cassification and anaysis of mutifrequent and mutipoarimetric airborne E-SAR data for hydroogica appications in Proceedings of AK Interpretation of Remote Sensing Data, Environmenta Assessment and Monitoring, Apri 1999, Hae, Germany. Hi, R.A., Smith, G.M., Fuer, R.M. and Veitch, N. (2002). Landscape modeing using integrated airborne muti-spectra and aser scanning data. Internationa Journa of Remote Sensing 23, Hoekman, D.H. and Quiñones, M.J. (2000). Land cover type and biomass cassification using AirSAR data for evauation of monitoring scenarios in the Coombian Amazon. IEEE Transactions on Geoscience and Remote Sensing 38, Hoopainen, M. and Wang, G. (1998). The caibration of digitized aeria photographs for forest stratification. Internationa Journa of Remote Sensing, 19, Houhouis, P.F. and Michener, W.K. (2000). Detecting wetand change: A rue-based approach using NWI and SPOT-XS data. Photogrammetric Engineering And Remote Sensing 66, Howard, J.A. (1991). Remote sensing of forest resources. Chapman and Ha, London. pp 420. Hudak, A.T., Lefsky, M.A., Cohen, W.B. and Berterretche, M. (2002). Integration of LIDAR and Landsat ETM+ data for estimating and mapping forest canopy height. Remote Sensing of Environment 82, Ichoku, C. and Karniei, A. (1996). A Review of Mixture Modeing Techniques for Sub-Pixe Land Cover Estimation. Remote Sensing Reviews 13, Jager, G. and Benz, U. (2000). Measures of cassification based on fuzzy simiarity. IEEE Transactions on Geoscience and Remote Sensing 38(3): Kadmon, R. and Harari-Kremer, R. (1999). Studying ong-term vegetation dynamics using digita processing of historica aeria photographs. Remote Sensing of Environment 68, Kerekes, J.P. and Landgrebe, D.A. (1989). Simuation of optica remote sensing systems. IEEE Transactions on Geoscience and Remote Sensing, 27, Krishnapuram, R. and Keer, J. (1996). The possibiistic c-means agorithm: insights and recommendations. IEEE Transactions on Fuzzy Systems 4, Kumar, S., Ghosh, J. and Crawford, M.M. (2002). Hierarchica fusion of mutipe cassifiers for hyperspectra data anaysis. Pattern anaysis and appications, 5, Laporterie, F. and Fouzat, G. (2003). The morphoogica pyramid concept as a too for muti-resoution data fusion in remote sensing. Integrated Computer-Aided Engineering 10, Largouet, C. and Cordier, M.O. (2001). Improving the andcover cassification using domain knowedge. AI Communications 14, Lefsky, M.A., Cohen, W.B., Acker, S.A., Parker, G.G., Spies, T.A. and Harding, D. (1999a). LIDAR remote sensing of the canopy structure and biophysica properties of Dougas-fir western hemock forests. Remote Sensing of Environment 70,

48 Lefsky, M.A., Harding, D., Cohen, W.B., Parker, G. and Shugart, H.H. (1999b). Surface LIDAR remote sensing of basa area and biomass in deciduous forests of eastern Maryand, USA. Remote Sensing of Environment 67, Lombardo, P., Oiver, C.J., Peizzeri, T.M. and Meoni, M. (2003). A new maximum-ikeihood joint segmentation technique for mutitempora SAR and mutiband optica images. IEEE Transactions on Geoscience and Remote Sensing 41, Luque, S.S. (2000). Evauating tempora changes using Muti-Spectra Scanner and Thematic Mapper data on the andscape of a natura reserve: the New Jersey Pine Barrens, a case study. Internationa Journa of Remote Sensing 21, Mathus, T.J. (2002). Review of Remote Sensing in Commercia Forestry. Unpubished report prepared for the Forestry Commission. Forestry Commission. 141p. Mason, D.C., Anderson, G.Q., Bradbury, R.B., Cobby, D.M., Davenport, I.J., Vandepo, M. and Wison, J.D. (2003). Measurement of habitat predictor variabes for organism-habitat modes using remote sensing and image segmentation. Internationa Journa of Remote Sensing 24, McCombs, J.W., Roberts, S.D. and Evans, D.L. (2003). Infuence of fusing LIDAR and mutispectra imagery on remotey sensed estimates of stand density and mean tree height in a managed oboy pine pantation. Forest science 49, Means, J.E., Acker, S.A., Harding, D.J., Bair, J.B., Lefsky, M.A., Cohen, W.B., Harmon, M.E. and McKee, A. (1999). Use of Large-Footprint Scanning Airborne LIDAR To Estimate Forest Stand Characteristics in the Western Cascades of Oregon. Remote Sensing of Environment 67, Megani, F. (2004). Cassification of mutitempora remote-sensing images by a fuzzy fusion of spectra and spatio-tempora contextua information. Internationa Journa of Pattern Recognition and Artificia Inteigence 18, Megani, F. and Serpico, S.B. (2003). A Markov random fied approach to spatio-tempora contextua image cassification. IEEE Transactions on Geoscience and Remote Sensing 41, Mikkoa, J. and Peikka, P. (2002). Normaization of bi-directiona effects in aeria CIR photographs to improve cassification accuracy of borea and subarctic vegetation for poen-andscape cassification. Internationa Journa of Remote Sensing 23, Mohan, B.K., Madhavan, B.B. and Gupta, U.M.D. (2000). Integration of IRS-1A L2 data by fuzzy ogic approaches for anduse cassification. Internationa Journa of Remote Sensing 21, Muchoney, D. and Straher, A. (2002a). Regiona vegetation mapping and direct and surface parameterization from remotey sensed and site data. Internationa Journa of Remote Sensing 23, Muchoney, D. and Straher, A. (2002b). Pixe- and site-based caibration and vaidation methods for evauating supervised cassification of remotey sensed data. Remote Sensing of Environment 81, Mumby, P.J. and Edwards, A.J. (2002). Mapping marine environments with IKONOS imagery: enhanced spatia resoution can deiver greater thematic accuracy. Remote Sensing of Environment 82,

49 Næsset, E. (2002). Predicting forest stand characteristics with airborne scanning aser using a practica two-stage procedure and fied data. Remote Sensing of Environment 80, Paterson, M., Lucas, R.M. and Chishom, L. (2001). Differentiation of seected Austraian woodand species using CASI Data. Proceedings, IGARSS Symposium, Sydney, Austraia. Sydney, Austraia (9 13 Juy 2001). Perkins, S., Theier, J., Brumby, Sp., Harvey, N., Porter, R., Szymanski, J.J. and Boch, J.J. (2000). Genie: a Hybrid Genetic Agorithm for Feature Cassification in Muti-Spectra Images. 4120, Proceedings of the Society of Photo-Optica Instrumentation Engineers (Spie). Popescu, S.C., Wynne, R.H. and Neson, R.F. (2002). Estimating pot-eve tree heights with LIDAR: oca fitering with a canopy-height based variabe window size. Computers and Eectronics in Agricuture 37, Qi, Jg., Wang, Cz., Matricardi, E. and Skoe, D. (2002). Improved Seective Logging Detection With Landsat Images in Tropica Regions. IGARSS 2002: IEEE Internationa Geoscience and Remote Sensing Symposium and 24th Canadian Symposium on Remote Sensing, Vos I-Vi, Proceedings Remote Sensing: Integrating Our View of the Panet, June 2002, Ranchin, T., Aiazzi, B., Aparone, L., Baronti, S. and Wad, L. (2003). Image fusion the ARSIS concept and some successfu impementation schemes. ISPRS Journa of Photogrammetry and Remote Sensing 58, Ray, S.S. (2004). Merging of IRS LISS III and PAN data evauation of various methods for a predominanty agricutura area. Internationa Journa of Remote Sensing 25, Read, J.M., Cark, D.B., Venticinque, E.M. and Moreira, M.P. (2003). Appication of Merged 1-M and 4-M Resoution Sateite Data to Research and Management in Tropica Forests. Journa of Appied Ecoogy 40, Riano, D., Meier, E., Agower, B., Chuvieco, E. and Ustin, S.L. (2003). Modeing airborne aser scanning data for the spatia generation of critica forest parameters in fire behavior modeing. Remote Sensing of Environment 86, Schetseaar, E.M. (2001). On preserving spectra baance in image fusion and its advantages for geoogica image interpretation. Photogrammetric Engineering and Remote Sensing 67, Shanmugam, S., Lucas, N., Phipps, P., Richards, A. and Barnsey, M. (2003). Assessment of remote sensing techniques for habitat mapping in coasta dune ecosystems. Journa of Coasta Research 19, Schwarz, M., Steinmeier, C. and Waser, L. (2002). Detection of Storm Losses in Apine Forest Areas by Different Methodica Approaches Using High-Resoution Sateite Data. Proceedings of 21st annua symposium of the European Association of Remote Sensing Laboratories May 2001, Schwarz, M., Steinmeier, C., Hoecz, F., Steber, O. and Wagner, H. (2003). Detection of Windthrow in Mountainous Regions With Different Remote Sensing Data and Cassification Methods. Scandinavian Journa of Forest Research 18,

50 Skidmore, A.K., Turner, B.J., Brinkhof, W. and Knowes, E. (1997). Performance of a neura network: Mapping forests using GIS and remotey sensed data. Photogrammetric Engineering and Remote Sensing 63: Smits, P.C. (2002). Mutipe cassifier systems for supervised remote sensing image cassification based on dynamic cassifier seection. IEEE Transactions on Geoscience and Remote Sensing 40, Smits, P.C., Deepiane, S.G. and Schowengerdt, R.A. (1999). Quaity assessment of image cassification agorithms for and-cover mapping: a review and a proposa for a cost-based approach. Internationa Journa of Remote Sensing 20, Song, C.H., Woodcock, C.E. and Li, X.W. (2002). The spectra/tempora manifestation of forest succession in optica imagery The potentia of mutitempora imagery. Remote Sensing of Environment 82: Song, C.H. and Woodcock, C.E. (2002). The spatia manifestation of forest succession in optica imagery The potentia of mutiresoution imagery. Remote Sensing of Environment 82, Souza, C., Firestone, L., Siva, L.M. and Roberts, D. (2003). Mapping forest degradation in the Eastern Amazon from SPOT 4 through spectra mixture modes. Remote Sensing of Environment 87, Sun, W.X., Heidt, V., Gong, P. and Xu, G. (2003). Information fusion for rura and-use cassification with high-resoution sateite imagery. IEEE Transactions on Geoscience and Remote Sensing 41, Tapiador, F.J. and Casanova, J.L. (2002). An agorithm for the fusion of images based on Jaynes maximum entropy method. Internationa Journa of Remote Sensing 4, Tapiador, F.J. and Casanova, J.L. (2003). Land use mapping methodoogy using remote sensing for the regiona panning directives in Segovia, Spain. Landscape and Urban Panning 62, Tayor, J.C., Brewer, T.R. and Bird, A.C. (2000). Monitoring andscape change in the nationa parks of Engand and Waes using air photo interpretation and GIS. Internationa Journa of Remote Sensing 21, Teggi, S., Cecchi, R. and Serafini, F. (2003). T.M. and IRS-1C-PAN data fusion using mutiresoution decomposition methods based on the a trous agorithm. Internationa Journa of Remote Sensing 24, Thenkabai, P.S., Smith, R.B. and De Pauw, E. (2000). Hyperspectra vegetation indices and their reationships with agricutura crop characteristics. Remote Sensing of the Environment 71, Thenkabai, P.S., Encona, E.A., Ashton, M.S., Legg, C. and De Dieu, M.J. (2004). Hyperion, Ikonos, Ai, and Etm Pus Sensors in the Study of African Rainforests. Remote Sensing of Environment 90, Thomas, V., Treitz, P., Jeinski, D., Mier, J., Lafeur, P. and McCaughey, J.H. (2003). Image cassification of a northern peatand compex using spectra and pant community data. Remote Sensing of Environment 84, Thomson, A.G., Eastwood, J.A., Yates, M.G., Fuer, R.M., Wadsworth, R.A. and Cox, R. (1998). Airborne remote sensing of intertida biotopes. Biota Marine Poution Buetin 37, Thomson, A.G., Fuer, R.M., Yates, M.G., Brown, S.L., Cox, R. and Wadsworth, R.A. (2003). The use of airborne remote sensing for extensive mapping of intertida sediments and satmarshes in eastern Engand. Internationa Journa of Remote Sensing 24,

51 Townsend, P.A. and Wash, S.J. (2001). Remote sensing of forested wetands: appication of muti-tempora and muti-spectra sateite imagery to determine pant community composition and structure in south-eastern USA. Pant Ecoogy 157, Ufarsson, M.O., Benediktsson, J.A. and Sveinsson, J.R. (2003). Data fusion and feature extraction in the waveet domain. Internationa Journa of Remote Sensing 24, Varekamp, C. and Hoekman, D.H. (2001). Segmentation of high-resoution InSAR data of a tropica forest using Fourier parameterized deformabe modes. Internationa Journa of Remote Sensing 22, Viering, La., Rowe, E., Chen, Xx., Dykstra, D. and Viering, K. (2002). Reationships Among Airborne Scanning LIDAR, High Resoution Mutispectra Imagery, and Ground-Based Inventory Data in a Ponderosa Pine Forest. Proceedings of the Internationa Geoscience and Remote Sensing Symposium (IGARSS), 2002, Wad, L. (1999). Definition and terms of reference in data fusion. Int. Arch. Photogrammetry Remote Sens. 32, Wang, F. (1990). Fuzzy supervised cassification of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing 28, Wikie, D.S. and Finn, J.T. (1996). Remote Sensing of Natura Resources. Coumbia University Press, New York. Wittmann, F., Anhuf, D. and Junk, W.J. (2002). Tree species distribution and community structure of centra Amazonian varzea forests by remote-sensing techniques. Journa of Tropica Ecoogy 18, Wu, D.R. and Linders, J. (1999). A new texture approach to discrimination of forest cearcut, canopy, and burned area using airborne C-band SAR. IEEE Transactions on Geoscience and Remote Sensing 37, Wu, L.X., Sun, B., Zhou, S.L., Huang, S.E. and Zhao, Q.G. (2004). A new fusion technique of remote sensing images for and use/cover. Pedosphere, 14, Xu, G. (2003). Information fusion for rura and-use cassification with high-resoution sateite imagery. IEEE Transactions on Geoscience and Remote Sensing 41, Zhang, Y. (2002). A New Automatic Approach for Effectivey Fusing Landsat 7 as we as IKONOS Images. Geoscience and Remote Sensing Symposium, IGARSS Zhang, C. and Frankin, S.E. (2002). Forest structure cassification using airborne mutispectra image texture and kriging anaysis. Geoscience and Remote Sensing Symposium, IGARSS Zimbe, D.A., Evans, D.L., Carson, G.C., Parker, R.C., Grado, S.C. and Gerard, P.D. (2003). Characterizing vertica forest structure using sma-footprint airborne LIDAR. Remote Sensing of Environment 87,

52 7 APPENDIX 1 Minimum object sizes in which a 5% change can be detected using pixe cassification techniques 7.1 Introduction The accuracy of estimating an object s size is affected by two sources of error. The first of these is due to the presence of party-occupied pixes around the object s edge, which may be fasey excuded from, or fasey incuded in, the object. these are the source of rounding error, a set of possibiities here termed edge effects. The second source of error is due to the degree of noise or genera cassification inaccuracy in an image. 7.2 Methods Edge effects The importance of edge effects varies according to an object s size and shape. Smaer, thinner objects, because of their higher perimeter/area ratio, are more subject to edge effects than arger, rounder objects. To quantify this, the size of edge effects were estimated for a variety of object shapes and sizes. Note this assumes no spectra confusion between objects. The rea edge of an object passes through a number of pixes, some of which are cassified as part of the object, and some as part of an adjacent object, or the background. This simuation assumes that each pixe through which the object s edge passes has an even chance of being incuded as part of the object or being excuded from the object. The presence of corner squares resuts in a consistent overestimation of the size of an object (by 1 pixe), however this does not affect the accuracy of this simuation. Using a binomia distribution, where the number of trias is equa to the number of edge squares, and the probabiity of success is 0.5, confidence imits on the number of edge squares incuded in the object can be generated. If the range of this estimate fas within a 5% range in the object s size, it is at east 95% ikey that a 5% change in the object s rea size can be discriminated. Noise effects The effect of image cassification error is greatest for sma objects and diminishes for arge objects. This is a consequence of sampe size: arge sampe sizes provide a more accurate estimation of a popuation mean than sma sampe sizes. To quantify this effect a uniform error-rate was appied to each pixe in objects of varying size, ie each pixe was given the same chance of being wrongy cassified. The resutant uncertainty in measuring the object s size was estimated by repeating this process (1000 times) to create confidence intervas for estimates of the object s size. If the variation we note due to noise is ess than the range of ±5% of the object s true size, then a 5% change in the object s size can be reiaby detected. Combining edge effects and noise To estimate the combined importance of edge effects and noise, the foowing procedure was used. A random observation from the frequency distribution of noise error for an object/error rate was added to a random observation from the frequency distribution of error due to edge effects for the same object. The net error was cacuated by summation. This procedure was repeated (1000 times) to generate a new set of confidence imits for tota error. As before, if the error range is greater than 5% of the object s area, a change of 5% magnitude can not be reiaby detected. 44

53 7.3 Resuts Tabe beow demonstrates the error associated with edge effects, noise, and edge effects and noise combined for three exampe objects. A change in the size of the objects can be reiaby measured when the error on the area estimate is ess than the size of the change. For exampe, it is possibe to reiaby detect a 5% change in a square 400 pixes in area if 95% of the error due to noise and/or edge effects is within ±20 pixes. Thus, considering noise error ony, a 5% change in a square 400 pixes in area can be reiaby detected if cassification accuracy is 90%, but not if it is 80%. Tabe Edge effect and noise error for exampe objects H and W (pixes) Rectange, 20x10 Square, 20x20 Square, 26x26 Area (pixes) Toerance (5% of area) ±10 ±20 ±34 95% confidence imits of area estimate (pixes), by source of error Edge effect ony ±6 ±7 ±8 Noise error, by cassification accuracy 60% ±13 ±27 ±35 80% ±11 ±21 ±30 90% ±8 ±16 ±21 95% ±6 ±12 ±16 99% ±3 ±6 ±7 Noise error and edge effect combined, by cassification accuracy 60% ±14 ±30 ±35 80% ±13 ±24 ±30 90% ±11 ±18 ±25 95% ±9 ±15 ±18 99% ±8 ±10 ±13 Tabe considers the minimum object sizes in which a 5% change in size can be reiaby detected (95% of the time) for a range of object sizes, shapes, and over a range of imaging resoutions. For exampe, an image with 2m resoution can reiaby detect changes of 5% in objects of 2500m 2 or arger if image cassification is 80% or better, when both edge effects and noise error are taken into account. Note the projections presented here for cassification errors (noise), and edge effect and noise combined, assume a square object. This represents a best case ie sma perimeter/area ratio. Rea word objects such wooded areas wi most ikey have more compex geometry and thus a far higher perimeter/area ratio. Hence the detection accuracy wi decine markedy and the area extent for which a 5% can be detected increase significanty. 45

54 Tabe Minimum object sizes (areas, in m 2 ) in which a 5% change in size can be detected for different image resoutions, by error type N.B. anaysis for noise and noise/edge effect combined assume a square object and as such are a best case scenario Error type Pixe size (m) No error Edge effect ony, for square object and object one pixe thick Sq. object Linear object Noise error ony, by cassification accuracy, % Noise error and edge effect combined, by cassification accuracy, %, square object Appying this anaysis to the particuar case of woodand bocks of within SNH SSSI woodand inventory is iuminating and ceary iustrates the issue of sensor resoution when considering a requirement to detect sma changes in area extents see Tabe overeaf. Note the information used here is obtained from SNH woodand SSSI vector data beieved to represent wooded area within SSSI boundaries. 46

55 Tabe Percentage of SNH SSSI woodands in which a 5% change in size can be detected for different image resoutions, by error type Error type Pixe size (m) No error Edge effect ony Noise error ony, by cassification accuracy, % Noise error and edge effect combined, by cassification accuracy, %, square object Note that this anaysis assumes the boundary between two habitat casses is a ceary demarked ine. Whie this serves to iustrate the issues attention is aso drawn to the reaity of community boundaries discussed under

56 8 APPENDIX 2 Layman s summaries of key remote sensing technoogies 8.1 Aeria digita ortho-photographs What are they? A digita ortho-photograph is an aeria photograph that has been modified to make it map-ike in the sense that it is amost a bird s eye view of the andscape. An additiona benefit is that a pixes can be referenced using map coordinates and readiy compared with other data in a GIS. The majority of existing digita ortho-photographs are derived from scanned hard-copy stereo aeria photography. Digita mapping cameras are ikey to become a more common source of data in the near future. Basic properties? Much ike optica images, digita ortho-photos can be described in terms of their spatia and spectra resoution properties, athough scae is an important property for the source imagery if in hardcopy format. Typica pixe sizes range from 10 50cm. The majority of avaiabe ortho-photographs are in natura coours (equivaent to a three channe optica image) as iustrated beow. The production of ortho-photographs requires a digita terrain mode be avaiabe to remove distortions in the spatia position of features compared to their true position in a map. For ow and moderate reief areas, high geometric quaity can be achieved, but in high and extreme reief areas some dispacements wi remain. Regardess of terrain effects however, the boundaries of ta objects (trees, buidings) wi aso show some dispacement. (Ortho-photography suppied and used by kind permission of Simmons Aerofims Ltd.) These data provide the basis for high quaity, spatiay detaied cartographic mapping of and cover, as shown beow for the area covered by the above ortho-photograph. (Land cover map courtesy of Cranfied University) The pros and the cons Digita ortho-photographs provide very detaied (10 50cm spatia resoution) images in natura coours within which the textura properties (a key eement for visua interpretation) of andscape features are we resoved. The discrimination of woodand from open habitats and areas of open water is generay straightforward and boundaries can be deineated quicky and accuratey visuay by trained interpreters. Feature recognition is aso better where fase-coour data are avaiabe. Surveys of arge-areas ( km 2 ) require many separate images that are taken at different times of day and on different days. Changes in the ighting conditions and changes in the ground conditions make digita processing for cassification of andscape features a major chaenge. For arge areas on-screen interpretation and digitisation can be time-consuming and aborious. The use of compression formats for the suppy of digita data can degrade the information content of the images. 48

57 Appendix 2 (continued) 8.2 Passive optica sensors (1 of 2) What are they? Passive sensors carried onboard sateites record either soar energy refected by the earth s surface (ie optica images) or energy that has been emitted by the earth itsef (ie therma images). Such images are normay described in terms of their resoution characteristics: spatia, spectra, radiometric and tempora. Passive images from optica and therma sensors have been avaiabe since the eary 1970s and since that time there has been continuing improvement in resoution characteristics with every new generations of earth observation (EO) sateite. Resoution properties The basic properties describing a remote sensed digita images are those of resoution: spatia, spectra, radiometric and tempora. The spatia resoution of a sensor is most simpy defined in terms of equivaent width on the ground of a singe pixe in the image. One of the main technica deveopments in the fied of remote sensing has been the increased number of sateite patforms that provide utra-high resoution images with pixe sizes between 1m and 10m, compared to the more traditiona high resoution systems with 20m and 30m resoution. For comparison it shoud be considered that superhigh resoution images (pixe sizes <1m) are aso becoming avaiabe from airborne patforms. These images provide more detaied information about earth surface features, as can be seen in the series of images beow of an area of moorand, woodand and agricuture that iustrate how the eve of detai avaiabe varies with the spatia resoution. The radiometric resoution of an imaging system describes its abiity to discriminate very sight differences in received energy, in effect, the number of tones distinguishabe in each band. The finer the radiometric resoution of a sensor the more sensitive it is to detecting sma differences in refected or emitted energy. The direct consequence of these improvements in radiometric resoution is an increased abiity to measure reativey sma changes in energy from target features. Athough the radiometric resoution wi have a direct infuence on the abiity to discriminate or identify specific features within an image, the radiometric resoution associated with a modern sensors is not a imiting factor in this context. 49

58 Appendix 2 (continued) 8.2 Passive optica sensors (2 of 2) Resoution properties (continued) The fundamenta property of the energy sensed is waveength, and this is normay used as a basis for separating the three fundamenta forms of remote sensing. Passive optica sensors record the magnitude of eectro-magnetic radiation (EMR) that is refected from the surface, and is commony expressed as digita number (DN). Whie surfaces exhibit continuous variation of refectance with waveength (as shown beow), passive sensors can ony sampe a finite number of waveengths. A simpe categorisation of sensors can be identified in terms of the number of bands coected for each image into: singe band (1 band); muti-spectra (<10 bands); hyper-spectra ( bands). It is generay true that the discrimination of vegetation casses is improved by acquiring images across a wider range of wave-engths and this woud impy a significant advantage in respect of hyper-spectra sensors. However recent research tends to indicate that from a cassification point of view there is unikey to be a significant advantage in acquiring data with more than bands. Consequenty, athough many hyper-spectra sensors are capabe of recording separate images for upwards of 100 wavebands, typicay this is ony undertaken for very specific types of research. Normay data are recorded in separate wavebands. It is generay acknowedged that therma remote sensing can provide ony imited discrimination of vegetated surfaces and therefore therma sensing devices are not discussed here. A remote sensing patforms have a tempora resoution or revisit time, ie the time between repeat coverage of the same area. The typica range for specific sateite sensors is between a few days and 10 s of days. Notionay, these are far shorter than monitoring periods and therefore the tempora resoution can appear unimportant. However, for both passive and active optica sensors the presence of coud severey inhibits the coection of imagery. For airborne systems tempora resoution is not fixed and can be determined based on ogistica issues. The pros and the cons Passive optica sensors can record directy in digita form mutipe-waveband images that can resove features as sma as 1m in size. They are capabe of coecting images at waveengths not captured by standard photography thereby enhancing the abiity to discriminate features. There has been a significant increase in the number of sateite sensors, with improved spectra and spatia resoution and imited hyper-spectra capabiity is avaiabe from space athough at high spatia resoution. A greater range of hyper-spectra devices are avaiabe from aeria patforms. Sateite patforms have an advantage compared to aeria patforms in that the ighting conditions at the time of image coection are very stabe. This makes the subsequent processing ess compex than those from aeria patforms. However, when the area of study remains fixed and is arge (1000s km 2 ) obtaining compete coverage of the area from a singe data source is difficut where coud is prevaent. For passive aeria sensors, the radiometric and geometric processing remains a considerabe chaenge when the area of study becomes requires mutipe (10s) of fight ines of data to be acquired. 50

59 Appendix 2 (continued) 8.3 Active sensors (1 of 2) What are they? Active sensors iuminate the earth s surface themseves and then record the properties of the energy returned by the surface to the sensor. These are more compex than passive optica devices and not a produce an image directy. The main types of active sensor are SARs (Synthetic Aperture Radars) and LIDARs (Light Detection and Ranging). SAR sensors produce images anaogous to passive optica and therma sensors, in which each pixe records the amount of energy returned to the sensor, but at microwave (ie mm to cm) waveengths. LIDAR sensors generate a series of point based estimates of the distance to the objects beow the sensor, and from this derive estimates of the height of these features. SAR properties By convention SAR systems are characterised using a etter to designate the waveength (cm) or frequency (Hertz) ranges of the EMR signa generated. The majority of SAR systems operate at X-band ( cm), C-band ( cm), L-band ( cm) or P-band ( cm). SARs are capabe of both day and night time operation and, as they operate at mm to cm waveengths, are reativey itte affected by coud and rain. SARs systems are carried on both sateite and aircraft patforms, but ony aeria sensors are capabe at present of being used operationay for woodand monitoring. Increasingy systems depoyed are capabe of muti, rather than singe, band data coection, aowing improved discrimination of surface features. A recent improvement of SAR technoogy is the capabiity to coect images using different poarisation of the microwave energy recorded. In some new SAR systems both the signa generated and that received can be separated into vertica and horizonta poarisations. This aows increased surface textura properties to be distinguished and again this process shoud aow increased differentiation of surface features. Note however that not a systems in use currenty record this poarisation information. SAR pros and the cons Exampe image from the E-SAR sensor (Copyright Definiens Imaging GmbH) Recent airborne SAR sensors can provide images with spatia resoutions of better than 5m, with 2m pixes a common operationa specification. Increasingy systems are becoming avaiabe that are capabe of simutaneousy acquiring images at more than a singe waveband and with different poarizations, thereby increasing the number of useabe images. In addition the deveopment of advanced processing techniques provides a basis for estimating the height of features as we as providing images. There remain, however, certain technica chaenges specific to the processing of SAR images. SAR pixe data are sensitive to a variety of surface characteristics other than the simpe presence of vegetation and other surface covers. For vegetated surfaces SAR images are sensitive to moisture content and structure. Where the emphasis is on feature recognition in regions with moderate to extreme terrain variation, the most significant confounding infuence can be the sensitivity to topography. SAR images aso contain a specific source of random noise caed specke that can ead to a 3 4 fod reduction in the actua compared to the nomina spatia resoution. This reduction in the effective spatia resoution wi be especiay significant where ony a singe microwave waveband image is avaiabe for feature mapping. 51

60 Appendix 2 (continued) 8.3 Active sensors (2 of 2) LIDAR properties LIDAR refers to a range of instruments that transmit a series of aser (ight ampification by stimuated emission of radiation) puses at near infrared waveengths towards a target. These systems are currenty ony avaiabe on aeria patforms. The signas returned from the ground surface are sensed and changes in the properties of the ight are used to determine one or more properties of the target. The simpest LIDAR instrument uses the time deay of the response to determine accuratey the distance to the feature being imaged. When the instrument is carried on an aeria patform (aircraft or heicopter) and the position of the patform and orientation monitored using GPS and an Inertia Measuring Unit (IMU), then terrain height can be determined. The technique is sometimes caed airborne aser terrain mapping or ALTM. The data from a LIDAR system occur as a series of points and these are processed to form a reguar image of height data. For many appications it is the height of the terrain in the absence of any surface features (eg trees, buidings) that is required and therefore the data may be re-processed to remove these features. The image beow eft shows an area with a bridge, vegetation and surrounding structures and on the right the terrain with the bridge retained but vegetation and other structures removed. (Copyright 2004 Optech Incorporated) LIDAR data have been coected across a range of forested environments mainy with a view to estimating the structura properties of woodand and forest regions. Current sensor technoogy provides a basis for identification of individua trees and the measurement precision appears adequate for the purpose of separating trees from shrub and grass. LIDAR pros and the cons LIDAR systems differ from the majority of other remote sensing systems in that the property measured is height and therefore an image is a derived rather than directy sensed product as for passive optica sensing and SAR sensing. A range of LIDAR systems are avaiabe and severa are currenty depoyed within the UK and data can be acquired with a 1 2m spatia resoution, height accuracies of around 20cm and across a swath of around 1km. For woodand appications these data are capabe of resoving individua tree crowns. LIDAR technoogy is sti deveoping and remains a reativey costy form of remote sensed data coection. Increasingy these data are coected in combination with other sensor data. 52

61 Appendix 2 (continued) 8.4 Pixe cassifiers What are they? Pixe cassifiers are used to assign ground cover (aso caed thematic) casses (eg deciduous woodand, arabe crop etc) to each pixe of a remote sensed image. Mathematica rues are used to match the spectra profie (or signature) of an individua pixe with the typica spectra profie that has been determined by a training process for a range of possibe casses. A pixes are assigned to a singe cover cass or are eft unassigned if their spectra profies are particuary unusua, athough the exact treatment varies between methods. How do they work? A source of remote sensing data, such as SPOT High Resoution Visibe (pane A), is processed to aow other anciary data (eg ground cover or map data) to be matched in a GIS. Using a source of ground data (pane B), a sampe of pixes is seected and a spectra profie showing how the refectance changes with waveength generated for each thematic cass of interest (pane C). These signatures are used to train the software to recognise a the thematic casses based on the spectra profie of each individua DN. The software cassifies a the pixes in the image (pane D) to one of the thematic casses or eaves the pixe uncassified. The pros and the cons On the positive side, pixe cassifiers typicay provide moderate eves of cassification accuracy that are repeatabe and based on an objective mode of how spectra profies vary. The imitations of these methods are aso we known. They rey mainy on coour variation for the assignment of casses, but ceary cover types that have simiar spectra profies (eg barey and grass in pane C) cannot be reiaby identified (compare panes B and D). Pixes occurring across the boundaries of ground cover types have mixed signatures and are often incorrecty cassified. Aso, features such as agricutura fieds, areas of woodand, and towns are not represented as distinct area features but purey as individuay abeed pixes. 53

62 Appendix 2 (continued) 8.5 Object cassifiers What are they? Object cassifiers are used to identify ground cover (aso caed thematic) casses (eg deciduous woodand, arabe crop) but unike pixe cassifiers the casses are assigned directy to groups of pixes rather than individua pixes. Each group of pixes is considered a separate object or feature for the purposes of the cassification. Mathematica rues are used to match the object profie (or signature) of an individua pixe with the typica profie for a range of possibe casses. A objects are assigned a singe cover cass or are eft uncassified if the profies of the objects are particuary unusua. How do they work? A source of remote sensing data, such as an aeria photograph (pane 1), is processed to aow other anciary data (eg ground cover or map data) to be matched in a GIS. Prior to the cassification stage, the image is segmented (pane 2) that is sub-divided to produce a series of spatiay discrete objects. Each object is made up of bocks of pixes. This property aows for more object characteristics to be derived and assessed compared to pixe based cassifiers. For exampe, as we as coour variations and texture, which can be used within pixe cassifiers, shape, area, context, and information from other objects can aso be incorporated. Using a source of ground data, a sampe of objects is seected of known cover types. For each cover type, a profie is deveoped showing not ony how the refectance changes with waveength for each thematic object of interest, but aso a seection of the other properties. These profies are used to train the software to recognise a the thematic objects within the segmented image. The software cassifies a the objects in the image (pane 3) to one of the thematic casses or eaves the object uncassified. The exampes beow show the output from the ecognition software produced by Definiens Imaging GmbH. Copyright Definiens Imaging GmbH The pros and the cons Object cassifiers provide a more advanced method of cassification that resut in spatiay defined objects or features as a cassification output. The cassification process can make use of an extended set of object properties, for exampe shape and area, compared to pixe cassifiers, with the potentia for improved accuracies of cassification. Vector poygons are produced automaticay for each object thereby avoiding the need for manua digitisation. The achievabe eves of accuracy for and cover mapping are not we defined since these cassifiers have not been subject to the same degree of testing as traditiona cassification methods. A key stage in the appication of object cassifiers is the process of image segmentation. Severa factors contro this process and at present there appears insufficient knowedge on how best to seect vaues appropriate to different andscapes and types of image. Hence the segmentation and cassification process can become engthy and quite time-consuming. The vector objects may aso ony agree oosey with those required. 54