System data analysis of Greek sites

Transcription

1 Project Title: BIO_SOS Biodiversity Multisource Monitoring System: from Space TO Species Contract No: FP7-SPA Instrument: Thematic Priority: Start of project: 1 December 2010 Duration: 36 months Deliverable No: D7.2 System data analysis of Greek sites Due date of deliverable: Actual date: submission 30 September November 2013 Version: 1st version of D7.2 Main Authors: Richard Lucas (P11), Vito De Pasquale (PKI), Fifamè N. Koudogbo (P7), Cristina Tarantino (P1), Maria Adamo (P1), Palma Blonda (P1).

2 Project ref. number Project title BIO_SOS: Biodiversity Multisource Monitoring System: from Space to Species Deliverable title System data analysis of Greek sites Deliverable number D7.2 Deliverable version Version 1 Previous version(s) Contractual date of delivery NA 30 September 2013 Actual date of delivery 12 November 2013 Deliverable filename Nature of deliverable Dissemination level BIO_SOS _D7.2_System_data_analysis_GR_v1 R PU Number of pages 22 Workpackage WP 7 Partner responsible Partner 13 (PKH) Author(s) Richard Lucas (P11), Vito De Pasquale (PKI), Fifamè N. Koudogbo (P7), Cristina Tarantino (P1), Maria Adamo (P1), Palma Blonda (P1). Editor EC Project Officer Richard Lucas (P11) Florence Beroud Abstract The classifications for all sites in Greece were undertaken primarily using Worldview-2 imagery and were based on the Food and Agricultural Organisation (FAO) Land Cover Classification System (LCCS) categories, which were then translated subsequently to General Habitat Categories (GHCs)

3 and then to Annex I classes. The method was developed using open source software that utilised a relatively simple set of rules. Whilst rule-based classifications have been adopted within BIO_SOS, the inputs can equally be achieved by extracting the appropriate information from more traditional classification approaches to produce the LCCS classification with these then translated to GHCs and to Annex I. Keywords BIO_SOS, EODHaM, LCCS map, GHC map, Annex I map

4 Signatures Written by Responsibility- Company Date Signature Richard Lucas Main author (P3) Verified by Dimitris Sykas Responsible for D7.2 (P13) Fifamè N. Koudogbo WP7 Leader (P7) Approved by Palma Blonda Coordinator (P1) Fifamè N. Koudogbo QAP (P7) BIO_SOS FP7-SPACE GA Page 4 of 22

5 Table of Contents 1. Executive summary EO data pre-processing List of the used EO imagery GR1 - Ekvoles Kalama (Kalamas Delta) GR2 - Elos Kalodiki (Kalodiki Fen) GR3 Stena Kalama (Kalamas Gorge) Pre-processing steps Orthorectification Radiometric correction Image co-registration Ancillary data for classification map production GR1 Ekvoles Kalama (Kalamas Delta) GR2 Elos Kalodiki (Kalodiki Fen) GR3 Stena Kalama (Kalamas Gorge) Ancillary data for GHC map production GR1 Ekvoles Kalama (Kalamas Delta) GR2 Elos Kalodiki (Kalodiki Fen) GR3 Stena Kalama (Kalamas Gorge) Data for Annex I map production GR1 Ekvoles Kalama (Kalamas Delta) GR2 Elos Kalodiki (Kalodiki Fen) GR3 Stena Kalama (Kalamas Gorge) GR1 Ekvoles Kalama (Kalamas Delta) images classification LCCS map Spectral indices and thresholds Output maps GHC map Annex I map Discussion on data requirements EO data requirements Ancillary data acquisition BIO_SOS FP7-SPACE GA Page 5 of 22

6 4.3 Open issues Conclusions Appendix Abbreviations and Acronyms BIO_SOS FP7-SPACE GA Page 6 of 22

7 List of Figures Figure 2-1 a) Pre and b) Peak flush images for the Kalamas Delta, Greece Figure 2-2 Examples of a) cultivation (defined using an existing thematic layer), b) individual trees and c) urban areas delineated through feature extraction. Areas are highlighted in magenta. D) Worldview-2 image from the peak flush period Figure 3-1 LCCS Level 3 classification of the Kalamas Delta, Greece Figure 3-2 LCCS classification of the Kalamas Delta based on the EODHaM system Figure 4-1 a) Pre- and b) peak-flush imagery over a subset of the Kalamas Delta test site Figure 4-2 LCCS Classification for the Kalamas Delta subset (Figure 4-1) highlighting the diversity of classes obtained BIO_SOS FP7-SPACE GA Page 7 of 22

8 List of Tables Table 2-1 List of available EO data over GR1 site Table 2-2 Worldview-2 spectral bands characteristics Table 3-1 Thresholds used for differentiating woody and photosynthetic vegetation BIO_SOS FP7-SPACE GA Page 8 of 22

9 1. Executive summary The classifications for all sites in Greece were undertaken primarily using Worldview-2 imagery and were based on the Food and Agricultural Organisation (FAO) Land Cover Classification System (LCCS) categories, which were then translated subsequently to General Habitat Categories (GHCs) and then to Annex I classes. The method was developed using open source software that utilised a relatively simple set of rules with these based on: a) Pixel level classification of vegetation phenology (evergreen or permanently green, winter deciduous/senescent, summer deciduous/senescent and continuously low productivity) and the extent of woody vegetation. b) Classification of vegetated (photosynthetic, non-photosynthetic, non-submerged aquatic) and non-vegetated surfaces (including urban and water) and separation between managed/cultivated/artificial and semi-natural and natural vegetation, with the latter either achieved through small feature extraction (i.e., for urban areas) or through reference to existing cadastral layers. Once defined, the 8 categories associated with LCCS Level 3 were mapped through cross tabulation. c) Classification of Level 4 categories which, for vegetated surfaces, focused on lifeform (woody trees and shrubs, graminoids) and non-lifeform categories relating to water and bare surfaces. Within these categories, a further division was provided based on their particular characteristics including, for example, leaf type, phenology, height, cover and stratification in the case of woody vegetation. Similar sub-level classifications of other categories were also achieved using a rule-based approach. For all sites, the classification was undertaken using pre- and peak flush imagery. The accuracy in the classification of LCCS categories and GHCs was difficult to establish because of the large number of classes generated and different interpretations from the remote sensing and field point of view. However, the approach provided a consistent, scale independent approach that could be applied to any site although expert knowledge was desirable in the generation of detailed classifications. BIO_SOS FP7-SPACE GA Page 9 of 22

10 2. EO data pre-processing 2.1 List of the used EO imagery GR1 - Ekvoles Kalama (Kalamas Delta) For the Kalamas Delta, the available imagery is listed in Table 2-1. Both pre- and peak flush datasets are displayed in Figure 2-1 a) and b), respectively. Sensor Date of acquisition Flush period 4 April 2013 Pre WV-2 16 June 2013 Peak Table 2-1 List of available EO data over GR1 site a) b) Figure 2-1 a) Pre and b) Peak flush images for the Kalamas Delta, Greece BIO_SOS FP7-SPACE GA Page 10 of 22

11 2.1.2 GR2 - Elos Kalodiki (Kalodiki Fen) No Very High Resolution (VHR) data over this site was available for the project GR3 Stena Kalama (Kalamas Gorge) No Very High Resolution (VHR) data over this site was available for the project. 2.2 Pre-processing steps The WorldView-2 high-resolution commercial imaging satellite was launched in 2009 and operative in The satellite orbit is sun-synchronous orbit with a period of minutes and an altitude of approximately 770 km. WorldView-2 acquires 11-bit data in nine spectral bands covering panchromatic, coastal, blue, green, yellow, red, red edge, NIR1, and NIR2. See Table 2-2 for details. At nadir, the collected nominal ground sample distance is 0.46 m (panchromatic) and 1.84 m (multispectral). Commercially available products are resampled to 0.5 m (panchromatic) and 2.0 m (multispectral). The nominal swath width is 16.4 km. WorldView-2 products used were radiometrically corrected images (Digital Numbers, DN i ) with no geometrical correction. Therefore in the pre-processing phase of the workflow two main steps have been performed: the geometric distortion introduced in the images by the process of acquisition and by the land topography are removed. the DN values are converted in Top-of-Atmosphere reflectance in order to have a better comparison between the images Orthorectification The orthorectification process removes the geometric distortion inherent in imagery caused by sensor orientation, topographic relief displacement, and systematic errors associated with imagery. This correction is performed specifying a sensor model which transforms object space coordinates to image map coordinates. For WorldView-2 a general sensor model has been used. It is realized by Rational Polynomial Functions (RPF), where the exterior and interior orientations are implicitly encoded in the form of rational polynomial coefficients (RPC) using third order polynomials for nominator and denominator. The terrain displacements are taken into account using an accurate digital elevation model derived by LiDAR acquisition with a spatial resolution of 8 m. The model parameters are refined in order to improve the geometric accuracy of the orthoimages fitting a set of ground control points (GCP). A GCP consists of a 3-dimensional ground coordinate (x, y, h) and its estimated accuracy. Marking the position of a GCP in the satellite image consequently gives a constraint on the model parameters. A number of such measurements give rise to a system of constraint equations that can be solved by least-squares adjustment. The GCPs in an image have been collected manually pointing in an image viewer using a reference image. The latter is a high quality orthophotos of the area Radiometric correction The values of VW2 pixels images are a function of how much spectral radiance enters the telescope aperture and the instrument conversion of that radiation into a digital signal. BIO_SOS FP7-SPACE GA Page 11 of 22

12 Spectral Band Center Wavelength ( m) Effective Bandwidth ( m) Spectral Irradiance (Wm -2 m -1 ) Panchromatic Coastal Blue Green Yellow Red Red Edge NIR NIR Table 2-2 Worldview-2 spectral bands characteristics Therefore, they are unique to WorldView-2 and should not be directly compared to imagery from other sensors in a radiometric/spectral sense. Instead, image pixels should be converted to top-of-atmosphere spectral radiance at a minimum. Top-of-atmosphere spectral radiance is defined as the spectral radiance entering the telescope aperture at the WorldView-2 altitude of 770 km. The conversion from radiometrically corrected image pixels to spectral radiance uses the following general equation for each band of a WorldView-2 product: L i band K band DN band i, band, (1) where L i,band are the Top Of Atmosphere (TOA) Spectral Radiance image pixels (Wm -2 sr -1 m -1 ), K band is the absolute radiometric calibration factor (Wm -2 sr -1 count -1 ) for a given band, DN i,band are the radiometrically corrected image pixels value (count) and band is the effective bandwidth ( m) for the given band (Table 2-2). The TOA Reflectance can be obtained using the following equation: i, band L Esun i, band band d 2 ES cos( ) s (2) where Esun band is the band averaged solar spectral irradiance (Wm -2 m -1 ) at a given Earth-Sun distance (Table 2-2), s is the solar zenith angle and d ES is the Earth-Sun distance (AU) for a given image acquisition. Top-of-atmosphere reflectance does not account for topographic, atmospheric, or BRDF differences effects Image co-registration When using multiple images, the spatial alignment among the images needs to be assured through the so called co-registration process. Almost 10 Ground Control Points, recognized on special stable artificial structures (e.g., vertices for buildings or road crossing points), have been used with a sub-pixel error and a linear warping and Nearest Neighbour resampling. BIO_SOS FP7-SPACE GA Page 12 of 22

13 2.3 Ancillary data for classification map production GR1 Ekvoles Kalama (Kalamas Delta) Much of the low-lying areas within the Kalamas Delta were cultivated, often with tree (e.g., olives) but also herbaceous crops. A difficulty in the classification of cultivated and managed areas was that these were highly variable spectrally because of the turnover of cropping and different levels of productivity in the fields at the times of the observations. For this reason, an agricultural survey that delineated most of the field units was used, although this did not encompass all of the cultivated area and hence some omissions were evident (Figure 2-2 a)). The individual trees and clusters of trees within the cultivated area but also in the mountains were extracted as separate features (Figure 2-2 b)), although smaller trees were not identified. These were combined subsequently with a woody mask, generated using a threshold of the blue/green ratio to map the extent of woody vegetation. A digital terrain model (DTM) was also available for the Delta. The feature extraction was conducted to isolate buildings and also road infrastructure (Figure 2-2 c)). The worldview image from the pre-flush period is also shown for reference. a) b) BIO_SOS FP7-SPACE GA Page 13 of 22

14 c) d) Figure 2-2 Examples of a) cultivation (defined using an existing thematic layer), b) individual trees and c) urban areas delineated through feature extraction. Areas are highlighted in magenta. D) Worldview-2 image from the peak flush period GR2 Elos Kalodiki (Kalodiki Fen) Classifications were not generated for GR2 because very high resolution (VHR) optical data were not acquired GR3 Stena Kalama (Kalamas Gorge) Classifications were not generated for GR3 because very high resolution (VHR) optical data were not acquired. 2.4 Ancillary data for GHC map production GR1 Ekvoles Kalama (Kalamas Delta) For GHC production, ancillary data layers were not available GR2 Elos Kalodiki (Kalodiki Fen) See GR3 Stena Kalama (Kalamas Gorge) See BIO_SOS FP7-SPACE GA Page 14 of 22

15 2.5 Data for Annex I map production GR1 Ekvoles Kalama (Kalamas Delta) For Annex I production, ancillary data layers were not available GR2 Elos Kalodiki (Kalodiki Fen) See GR3 Stena Kalama (Kalamas Gorge) See BIO_SOS FP7-SPACE GA Page 15 of 22

16 3. GR1 Ekvoles Kalama (Kalamas Delta) images classification 3.1 LCCS map Spectral indices and thresholds At the pixel level, differentiation of several categories of vegetation was needed because of the diversity occurring when observed during the pre and peak flush periods, including photosynthetic vegetation, non-photosynthetic vegetation and non-submerged aquatic vegetation. However, the ruleset was simplified by identifying only those areas that were photosynthetic in the pre and the peak images and the differences in these areas was then used to define the extent of evergreen vegetation (including permanent grasslands), winter deciduous (e.g., woodlands) and summer deciduous (e.g., senescent grasslands). Areas of low productivity were characterized as having a low level of photosynthetic activity in both the pre and peak season imagery. The extent of woody vegetation was also determined with reference to the canopy height model (CHM) derived from the LiDAR data although outside of the LiDAR mask, the Forest Discrimination Index (FDI) provided best differentiation. PIXEL LEVEL Woody vegetation Photosynthetic vegetation PeakBG > 0.05 & PeakWBI < 0.95 PreNDVI > 0.5 PeakNDVI > 0.2 Table 3-1 Thresholds used for differentiating woody and photosynthetic vegetation Output maps The pre- and peak flush imagery were classified to provide maps of the extent of photosynthetic vegetation with an NDVI threshold used for the pre and peak flush images with these capturing also the lower productivity vegetation with the wetlands areas on the coast. The extent of woody vegetation was also defined at the pixel level using a threshold of the blue/green ratio for the peak image, although the Water Band Index (WBI) also had to be used to separate areas of shadow and wetlands. At Level 1-3, the extent of urban infrastructure was established through feature extraction with roads determined through reference to existing data layers (OS Mastermap). Whilst the extent of the cultivated area could be defined through classification of large objects, only part of the area was identified and hence existing cadastral information was necessarily used. Areas of open water were mapped using the Water Band Index (WBI) from both the pre and the peak flush periods. These nonvegetated areas were then combined. For discriminating photosynthetic vegetation based on small objects, the Forest Discrimination Index (FDI) from both the pre and peak periods was used although areas of non-photosynthetic vegetation were defined using the Plant Senescence Reflectance Index (PSRI). Areas of non-submerged aquatic vegetation were also mapped using a threshold of the pre-flush red reflectance (< 8.5 %), the pre-flush FDI and the slope derived from the DEM (< 5 m) as well as a DTM upper limit of 20 m. This avoided confusion with shadowed areas in the more mountainous regions. All remaining areas were assigned to a non-vegetation category. As with other sites, a cross-tabulation was performed to generate the 8- BIO_SOS FP7-SPACE GA Page 16 of 22

17 class LCCS Level 3 classification that described the extent of cultivated/managed and artificial areas (including water and urban areas) and natural and semi-natural terrestrial and aquatic vegetation (Figure 3-1). Terrestrial vegetated Aquatic vegetated Cultivated Natural bare Artificial bare Natural water Figure 3-1 LCCS Level 3 classification of the Kalamas Delta, Greece Fort the Level 4 classification, a wide range of land covers were evident because of the broad distribution of cultivated and semi-natural/natural terrestrial and aquatic systems. The extent of woody vegetation was defined by considering the distribution of trees at the object level but also the blue/green ratio during the peak flush period. Trees and shrubs were able to be differentiated to a certain degree by using a threshold of the Forest Discrimination Index (FDI). All remaining areas were associated with herbaceous vegetation, with graminoids identified using a combination of the pre-flush PSRI and NDVI to identify senescent and hotosynthetic vegetation respectively in both the semi-natural and cultivated areas. Estimates of cover were generated by referencing the number of vegetated pixels occurring within each of the small objects. Needle-leaved vegetation was differentiated using a threshold of the pre-fluh blue/green ratio whilst evergreen and deciduous phenologies were separated by considering the proportion of green or photosynthetic pixels BIO_SOS FP7-SPACE GA Page 17 of 22

18 within each object. No canopy height model (CHM) was available for the area. Different descriptors of water and bare ground were also provided. The classification at Level 4 provided a good representation of the extent of the woody and non-woody vegetation. It is represented in Figure 3-2. Aquatic vegetation was also identified although some confusion with emergent aquatic and terrestrial vegetation was evident towards the landward margins. Areas of bare ground (primarily sand) were differentiated. Figure 3-2 LCCS classification of the Kalamas Delta based on the EODHaM system 3.2 GHC map No GHC map was generated for this site. 3.3 Annex I map No Annex I map was generated for this site. BIO_SOS FP7-SPACE GA Page 18 of 22

19 4. Discussion on data requirements 4.1 EO data requirements The classification of the Kalamas Delta was undertaken to Level 3 and subsequently to Level 4 such that lifeforms, water and bare ground were described in good detail. Additional categories could have been defined but this requires extensive user interaction and knowledge. For example, knowing whether water is tidal or otherwise or the persistence in the absence of multiple image sets requires a good level of user knowledge without the need for satellite-based observations. As with all of the sites in the Mediterranean and northern Europe, images acquired during the pre and peak flush periods are required as a minimum. The use of additional images in the transition periods would allow better discrimination of some categories (e.g., annual/perennial graminoids/forbs or different water seasonality and durations) but the requirements on the rule-base become greater. Better discrimination of cultivated land would also be improved using more image dates. Worldview-2 data is also preferred over other VHR datasets with just visible and near infrared channels as the 8-band dataset allows for the calculation of key indices (e.g., PSRI), which facilitates discrimination of some vegetated categories and photosynthetic states. 4.2 Ancillary data acquisition The feature extraction of urban infrastructure and individual trees or clusters of trees allows better mapping of artificial surfaces but also semi-natural/natural woodlands and cultivated tree crops (e.g., olive groves). Whilst the use of cadastral thematic layers is not desirable, this was necessary in this case because of the diversity of reflectance values associated with the cultivated fields. 4.3 Open issues As with all sites, knowledge of the area is essential when considering the classification to Level 4 and beyond but the rules can be inserted within the EODHaM system to provide a full LCCS classification but discussions with land managers, conservationists and other interested parties are needed in order to optimise the classification. The assessment of error is also compromised by the high level of detail obtained in the classification, particularly if LiDAR data are available which allow combinations of height to be included with cover, thereby leading to multiple categories. Pre- and peak-flush images of a subset area are shown in Figure 4-1 a) and b). The corresponding LCCS map is displayed in Figure 4-2. The level of detail provided by the classification shows multiple categories of aquatic and terrestrial (particularly forest vegetation) as well as bare surfaces. BIO_SOS FP7-SPACE GA Page 19 of 22

20 a) b) Figure 4-1 a) Pre- and b) peak-flush imagery over a subset of the Kalamas Delta test site Loose and shifting sands Gravels or mud Graminoids on flooded land Open (40-65 %) Graminoids Needle-leaved deciduous shrublands (thicket) on flooded land Broadleaved deciduous open (40-65 %) shrubland (thicket) on flooded land Broadleaved evergreen open (40-65 %) shrubland (thicket) on flooded land Needle-leaved deciduous open (40-65 %) shrubland (thicket) Needle-leaved open (40-65 %) trees Graminoid crops Herbaceous vegetation on flooded land Perennial waterbodies Figure 4-2 LCCS Classification for the Kalamas Delta subset (Figure 4-1) highlighting the diversity of classes obtained BIO_SOS FP7-SPACE GA Page 20 of 22

21 5. Conclusions Focusing on the Kalamas data, a classification of LCCS categories at Level 3 and Level 4 was obtained using the EODHaM system and using standard datasets (e.g., NDVI, PSRI and WBI). These classes were translated to GHCs with ambiguities resolved and also to Annex I classes. A full classification of LCCS Level 4 was not achieved because of the complexity of the landscape and requirement for greater interaction with interested parties particularly in relation to water dynamics and cultivation cycles. However, the classification provided a detailed overview of the landscape which was consistent with those applied at other sites as part of the BIO_SOS project. Further refinement of the classification can be achieved by integrating data from more than two dates (i.e., the pre and peak season) and by also providing more advanced rules for the delineation of cultivated areas (e.g., through descriptions of trees within cultivated areas). BIO_SOS FP7-SPACE GA Page 21 of 22

22 6. Appendix Abbreviations and Acronyms ASCII AU BIO_SOS BNR BRDF CHM CORINE DN DSM DTM EODHaM FAO FDI GCP GHC GR GRR LC/LU LCCS LIDAR LO m NDVI NIR OA OSM PSRI RPC RPF SO TOA VHR WBI WV-2 American Standard Code for Information Interchange Astronomical Unit Biodiversity Multi-Source Monitoring System: From Space To Species Blue/NIR Ratio Bidirectional Reflectance Distribution Function Canopy Height Model Coordination of Information on the Environment Digital Number Digital Surface Model Digital Terrain Model EO Data for Habitat Monitoring Food and Agricultural Organisation Forest Discrimination Index Ground Control Points General Habitat Categories Greek site Green/Red Ratio Land Cover/Land Use Land Cover Classification Scheme Light Detection and ranging Large Object meter Normalized Difference Vegetation Index Near-Infra Red Overall Accuracy OpenStreet Mastermap Plant Senescence Reflectance Index Rational Polynomial Coefficient Rational Polynomial Function Small Object Top of Atmosphere Very High Resolution Water Band Index Worldview-2 BIO_SOS FP7-SPACE GA Page 22 of 22