D Algorithm Theoretical Basis Document Volume III BA Merging

Transcription

1 D Algorithm Theoretical Basis Document Volume III BA Merging (ATBD III) BA Merging, version 2 Project Name fire_cci Contract N /10/I-NB Project Manager Arnd Berns-Silva Last Change Date 02/11/2014 Version 2.3 State Final Author Kevin Tansey / Andrew Bradley / Marc Padilla Document Ref: Fire_cci_Ph3_UL_D3_6_3_ATBD_III_v2_3 Document Type: Public

2 Project Partners Prime Contractor/ Scientific Lead - UAH - University of Alcalá de Henares (Spain) Project Management - GAF AG, (Germany) System Engineering Partners - GMV - Aerospace & Defence (Spain) - DLR - German Aerospace Centre (Germany) Earth Observation Partners - ISA - Instituto Superior de Agronomia (Portugal) - UL - University of Leicester (United Kingdom) - DLR - German Aerospace Centre (Germany) Climate Modelling Partners - IRD-CNRS - L Institut de Recherche pour le Développement - Centre National de la Recherche Scientifique (France) - JÜLICH - Forschungszentrum Jülich GmbH (Germany) - LSCE - Laboratoire des Sciences du Climat et l Environnement (France) Distribution Affiliation Name Address Copies ESA-ECSAT Stephen Plummer (ESA ECSAT) Stephen.Plummer@esa.int electronic copy Project Team Emilio Chuvieco, (UAH) Itziar Alonso-Canas (UAH) Stijn Hantson (UAH) Marc Padilla Parellada (UAH) Dante Corti (UAH) Arnd Berns-Silva(GAF) Christopher Sandow (GAF) Stefan Saradeth (GAF) Jose Miguel Pereira (ISA) Duarte Oom (ISA) Bernardo Mota (ISA) Gerardo López Saldaña (ISA) Kevin Tansey (UL) Andrew Bradley Oscar Pérez (GMV) Luis Gutiérrez (GMV) Ignacio García Gil (GMV) Andreas Müller (DLR) Martin Bachmann (DLR) Kurt Guenther (DLR) Martin Habermeyer (DLR) Eric Borg (DLR) Martin Schultz (JÜLICH) Angelika Heil (JÜLICH) Florent Mouillot (IRD) Philippe Ciais (LSCE) Patricia Cadule (LSCE) Chao Yue (LSCE) Emilio.chuvieco@uah.es itziar.alonsoc@uah.es Hantson.stijn@gmail.com padilla.marc@gmail.com dante.corti@uah.es arnd.berns-silva@gaf.de christopher.sandow@gaf.de stefan.saradeth@gaf.de jmocpereira@gmail.com duarte.oom@gmail. bmota@isa.utl.pt gerardolopez@isa.utl.pt kjt7@le.ac.uk a.bradley@imperial.ac.uk operez@gmv.com lgutierrez@gmv.com igarcia@gmv.com andreas.mueller@dlr.de martin.bachmann@dlr.de kurt.guenther@dlr.de martin.habermeyer@dlr.de eric.borg@dlr.de m.schultz@fz-juelich.de a.heil@fz-juelich.de florent.mouillot@ird.fr philippe.ciais@cea.fr patricia.cadule@lsce.ipsl.fr chaoyuejoy@gmail.com electronic copy D3.6.3 ATBD III - BA Merging Page II

3 Summary This document is the Algorithm Theoretical Basis Document (ATBD) Volume III (version 2) for the fire_cci project. The document reviews existing methods of merging remotely sensed and outlines the method used to produce the merged burned area data sets considering the user requirements and BA data that is to be merged. Affiliation/Function Name Date Prepared UL Kevin Tansey, Andrew Bradley 12/12/2013, 24/01/ /08/2014 Reviewed UAH/GAF AG Emilio Chuvieco, Arnd Berns-Silva 13/08/2014 Authorized UAH/ Prime Contractor Emilio Chuvieco Accepted ESA/ Project Manager Stephen Plummer Signatures Name Date Signature Signature of authorisation and overall approval Signature of acceptance by ESA Emilio Chuvieco Document Status Sheet Issue Date Details /05/2012 First Document Issue /08/2012 Addressing comments according to CCI_FIRE_EOPS-MM pdf /08/2013 Upgrade and final algorithm development /12/2013 Addressing comments according CCI-FIRE-EOPS-MM pdf /01/2014 Addressing comments according CCI-FIRE-EOPS-MM pdf /08/2014 Minor update in section Document Change Record # Date Request Location Details /08/2012 UL Whole document Updating and adding further remarks /08/2013 UL Section 3 Fig. 3 Fig. 4 Section 3.3 Updating and rephrasing Updating annotation Updating annotation Updating considerations on algorithm design Sections 4.2.1, 4.4 Updating and rephrasing Section 6 Section Table 1 Fig. 7 Section 6.2 Section Table 2 Updating and rephrasing Upgrading and further remarks Updating table attributes Updating workflow chart Updating and further remarks Further remarks on reference data sets Updating output attributes D3.6.3 ATBD III - BA Merging Page III

4 # Date Request Location Details Section 6.3 Section Table 3 Updating and further remarks Further remarks on reference data sets Updating output attributes # Date Request Location Details 2.0 UL Section 6.4 Fig. 8, 9, 10, 11 Fig. 13, 14, 15, 16, /12/2013 UL, A. Bradley Section 1 Section 2 Section 3 Section 3.1 Section 3.3 Section 4 Section 4.1 Section 4.2 Section Section 6.1 Section 6.2 Section 6.3 Section /01/2014 UL Section 1 Section Annex 1 Annex 2 Upgrading and further remarks Updating workflow chart Introducing examples of data at the global scale Rephrasing and restructuring Executive summary Rephrasing Rephrasing and further remarks; Figure 1 updated; Figure 3 annotation updated; Restructuring, amending and further remarks on main considerations in algorithm design Rephrasing Restructuring and further remarks Adding further remarks; Rephrasing Rephrasing and adding further remarks; Figure 7 updated; Rephrasing and further remarks; Figure 8 updated; Updated; Previous chapters Processing flow and Data availability discarded; Updated Including additional sensors within CCI Phase 2 Confidence level for Algorithm 2 corrected Rephrased and updated; Introducing Table 4; Rephrased and updated; /08/2014 UL Section Including notice on single sensor burned area D3.6.3 ATBD III - BA Merging Page IV

5 Table of Contents 1 Executive summary Introduction Purpose of this document Applicable Documents Background to the burned area merging algorithm Challenges for burned area merging Product specification Main considerations in algorithm design Review of approaches to create EO data products Non-fire products Single sensor products Multi-sensor products Fire and burned area products Single sensor fire and burned area data sets Multi-sensor fire and burned area data sets Method summary Stratification and global processing Gap filling and time series extension Combining spectral data Blending products Contrasts between burned area and other EO products Uncertainty and confidence in fire products The Merging Algorithm Development Merging strategies Merging solution Algorithm Requirements: PIXEL Data required Data input Data output Algorithm Requirements: GRID Data required Data input Data output References Appendix Appendix D3.6.3 ATBD III - BA Merging Page V

6 List of Tables Table 1: The resolution of the BA output for each sensor source and algorithm used for burned area detection...13 Table 2: Attributes of each pixel based burned area algorithm product...16 Table 3: Data and attributes required for the GRID product...17 Table 4: Error matrix where p ij is the proportion of area in cell (i, j) List of Figures Figure 1: Temporal overlaps for sensors suitable for detecting burned area (1, ATSR-2. 2 AATSR, 3 VGT-1, 4 VGT-2, 5 MERIS FRS.)...3 Figure 2: Four day composite of AATSR data (22-25/06/2008) (provided by Kurt Guenther/DLR)...3 Figure 3: Conceptual overlay of 3 sensors on the same day with varying swath widths. Almost complete coverage is possible. Green represents the surface, light grey outlines and grey strips represent the MERIS swath, dark grey outlines and grey strips represent SPOT VGT swath width and yellow outlines and yellow strips represent AATSR swath width....4 Figure 4: Conceptual demonstration showing the potential variability in BA detection between sensors for two different land covers. The level of uncertainty is represented by the grey tones in each swath, the dark grey patches represent low uncertainty and the lighter grey patches have higher uncertainty...4 Figure 5: The 22 unit stratification for Globcover based on global vegetation characteristics (from Bontemps et al. 2010)...6 Figure 6: Stratified areas used for the GBA2000 product. These units tended to be based areas of interest specific to the scientific groups in the consortium rather than representation of global environmental conditions (from Tansey et al. 2002)...8 Figure 7: Spatial resampling in the merging process. Data processing without MERIS follows the dashed arrows Figure 8: Processing flow for the merging algorithm...15 D3.6.3 ATBD III - BA Merging Page VI

7 List of Abbreviations AATSR Advanced Along Track Scanning Radiometer AIRS Atmospheric InfraRed Sounder ATBD Algorithm Theoretical Basis Document ATSR Along Track Scanning Radiometer AVHRR Advanced Very High Resolution Radiometer BA Burned Area BAE Burned Area Estimate DMSP Defence Meteorological Satellite Programme DMSP OLS DMSP Operational Line Scan System EO Earth Observation ETM Enhanced Thematic Mapper FRS Full Resolution FRS Full Resolution Swath GBA2000 Global Burned Area 2000 GFED Global Fire Emissions Database GLC2000 Global Land Cover 2000 GMS S-VISSR Geostationary Meteorological Satellite, Stretched-Visible InfraRed Spin Scan Radiometer JERS Japanese Earth Resources Satellite NIR Near InfraRed MERIS Medium Resolution Imaging Spectrometer MODIS Moderate Resolution Imaging Spectroradiometer NDVI Normalised Difference Vegetation Index OLS Operational Line scan System OMI Ozone Monitoring Instrument PSD Product Specification Document PVASR Product Validation and Algorithm Selection Report PVP Product Validation Plan RR Reduced Resolution RS Remote Sensing SeaWiFS Sea viewing Wide Field of View Scanner SPOT Système Pour l'observation de la Terre VGT VEGETATION, Earth observation sensor owned by CNES on SPOT SST Sea Surface Temperature TM Thematic Mapper TRMM Tropical Rainforest Measuring Mission TRMM MI Tropical Rainforest Measuring Mission Microwave Imager TRMM VIRS Tropical Rainforest Measuring Mission Visual and InfraRed Scanner URD User Requirements Document VIRS Visual and InfraRed Scanner VIS Visible D3.6.3 ATBD III - BA Merging Page VII

8 1 Executive summary Detecting fire extent, otherwise known as burned area mapping, is a useful component for the mapping and modelling of emissions (including greenhouse gases and aerosols), ecosystem disturbance and disaster management amongst other reasons. The ESA Fire_cci project aims to provide the climate modelling community and other scientific users with a consistent error characterised long term global record of the extent and timing of fire disturbance monitored from satellite platforms. This document describes the scientific basis behind the merged product following a review of methods used to create merged thematic products from Earth observation data. Individual burned area algorithms applied to satellite data often show inconsistent performance as the algorithms cannot account for all the different and varying environmental factors controlling and characterising the presence of burn scars. The extent of burned area mapping is also limited by the spatial and temporal characteristics of the satellite. To overcome these challenges the Fire_cci project joins data from different satellite missions with short life spans into a long time series, and improves the confidence of burn detection by merging different burned area products. The Phase I merging algorithm accepts burned area output from two different burned area detection algorithms implemented on five satellite data sources (ATBD II). These data sources are data from the following satellite sensors, the Along Track Scanning Radiometer-2 (ATSR-2), Advanced Along Track Scanning Radiometer (AATSR), the SPOT VEGETATION-1 (VGT-1) and SPOT VEGETATION-2 (VGT-2) instruments, and the Medium Resolution Imaging Spectrometer at Full Resolution full Swath (MERIS FRS). With minor modifications the processing chain can be extended to accommodate a wider range of sensors if required. In accordance with the URD (Schultz et al. 2011) and PSD (Chuvieco et al. 2013) the basis behind the production of two merged burned area data sets, the monthly pixel product (resolution of 1/120 then 1/360 when higher resolution MERIS data are available) and the half monthly grid product (resolution 0.5 degree) are described. During processing multiple observations of the same burn recorded by more than one BA source are accepted. Observations of burns that are only recorded by one BA source are screened to eliminate data with low confidence thresholds set according to the performance of each BA source over different land cover types and low uncertainty according to the position of a burned pixel relative to other BA patches. For the pixel product the algorithm calculates the earliest possible observation of a burned area, which sensors have detected a burned area, the highest confidence levels, and the land cover burned. These results are then aggregated to produce the grid product, with half monthly statistics on the sum of burned area, the standard error of the burned area, the fraction of the surface observed, the number of burned patches and the sum of each land cover type burned. This document refers to the production of a global time series of BA between 2005 and 2009 for Phase I of the Climate Change Initiative programme. Phase II will increase the length of the time series ( ) with the addition of data from PROBA-V, the Sentinel mission and potentially AVHRR (Metop) and further development of the system for processing beyond and potentially going backwards (AVHRR) with a capacity for frequent reprocessing. D3.6.3 ATBD III - BA Merging Page 1

9 2 Introduction 2.1 Purpose of this document This document has been written to inform the users of about the theoretical basis and processing stages of the v2 ESA Fire_cci merged product. It is an update of and supersedes the information in ATBD III v1. The document informs the user about existing scientific developments for merging satellite data products, and how these findings have contributed to the development of the fire merged product. The processing chain is expanded so the user has an understanding behind the calculations and assumptions for the product, strengthening the user interpretation for their particular application of this product. Following the CCI Statement of Work [AD-1], this document is the final issue of the ATBD III. 2.2 Applicable Documents [AD-1] ESA Climate Change Initiative (CCI) Phase 1, Scientific User Consultation and Detailed Specification, Statement of Work, EOP-SEP/SOW/ /SP, v1.4, 2009, D3.6.3 ATBD III - BA Merging Page 2

10 3 Background to the burned area merging algorithm 3.1 Challenges for burned area merging The specific aim of the Fire_cci project is to develop a long term, consistent time series of burned area globally. The main issues in providing this product are spatial and temporal satellite coverage, different spatial resolution of data and product consistency due to the variable performance of burned area algorithms on global satellite data. Temporal coverage is provided the following European satellite sensors; the Along Track Scanning Radiometer-2 (ATSR-2), Advanced Along Track Scanning Radiometer (AATSR), SPOT VEGETATION-1 (VGT-1), SPOT VEGETATION-2 (VGT-2) and the MEdium Resolution Imaging Spectrometer (MERIS) at Full ReSolution (FRS), which are suitable sources to detect burned area (see Pereira et al. 2013). There is a temporal overlap of these missions but not all of the missions exist at the same time (Figure 1) T1 T2 T3 T4 T6 T7 T8 Figure 1: Temporal overlaps for sensors suitable for detecting burned area (1, ATSR-2. 2 AATSR, 3 VGT- 1, 4 VGT-2, 5 MERIS FRS.) In T1 (~1995 to ~1998) only data from ATSR-2 is available whereas in T4, a short time period (~2002 to ~2003), data are available for ATSR-2 AATSR, VGT-1, VGT-2 and MERIS FRS. The longest period of data overlap is during T7 (~2005 to ~2010) when AATSR, VGT-2 and MERIS data are available. Figure 2: Four day composite of AATSR data (22-25/06/2008) (provided by Kurt Guenther/DLR) Spatial coverage is variable because of the technical constraints of the sensors and sensor platforms. Swath widths and intervals between repeat overpass times will determine how well the Earth surface is observed. On the one hand the 2200 km swath of SPOT VGT provides almost complete daily global coverage with the exception of a few gaps at the equator whereas the 512 km swath of ATSR cannot make a complete global mosaic even after 4 days (Figure 2). Spatial coverage can be improved by combining data from the different missions to reduce gaps in daily coverage of individual missions. For example the MERES FRS instrument has a swath of 1150 km, VGT has a swath of 2200 km and AATSR has a swath of 512 km. If we consider the daily coverage for each sensor there is a high probability of achieving total or near total coverage (Figure 3). Total coverage is more likely at the poles where swaths converge (Figure 2). D3.6.3 ATBD III - BA Merging Page 3

11 Figure 3: Conceptual overlay of 3 sensors on the same day with varying swath widths. Almost complete coverage is possible. Green represents the surface, light grey outlines and grey strips represent the MERIS swath, dark grey outlines and grey strips represent SPOT VGT swath width and yellow outlines and yellow strips represent AATSR swath width. This combination of sensor data can also reduce day to day data loss e.g. cloud cover (Figure 2 blue areas), haze and bad quality data flagged as unsuitable for processing. Furthermore mission overlap increases the probability of multiple detection of the same burn to create a more robust Fire_cci product because uncertainty and confidence of detection of each burn is improved. Burned area algorithms are prone to inconsistent performance over different vegetation types because current burned area algorithms cannot account for all the different and varying environmental factors controlling and characterising the presence of burn scars. With a series of different burned area products observation of the same burn from several different sources has the potential to improve the extent, the timing and certainty in the detection of burned area (Figure 4). In the example BA 1 and BA 2 are burns in two different vegetation cover types denoted by the green and yellow colours. Burned area products created from sensors 3, 2 and 5 perform differently when compared and perform differently in BA1 and BA2 vegetation cover types. By overlaying the different data sets these results can be combined to give a single product which much more closely resembles the extent of the actual burned area. More frequent repeat overpasses of one sensor can also improve on finding the earliest date of the burn. Figure 4: Conceptual demonstration showing the potential variability in BA detection between sensors for two different land covers. The level of uncertainty is represented by the grey tones in each swath, the dark grey patches represent low uncertainty and the lighter grey patches have higher uncertainty It is not always possible to have the same sensor combinations observing a burned area as described in figure 4 however the merging process needs to be consistent and use the same method. Although the reduction in uncertainty is less when the full complement of sensors is unavailable different sensor combinations can still improve the uncertainty and representation of the burned area. The aim of this merging algorithm is to take advantage of the different spatial and temporal coverage and characteristics of the source data and to improve burned area uncertainty by combining available combinations of burned area products. There is also an additional challenge to meet the user D3.6.3 ATBD III - BA Merging Page 4

12 requirements set out in the Product Specification Document (Chuvieco et al. 2013) briefly outlined in section Product specification There are a number of factors taken into account that have led to the specification of this product mainly through the results of the user survey (questionnaire to climate modellers and other fire product users) which is summarised in the User Requirements Document (Schultz et al. 2011). Not all of the requirements can be met due to technical constraints of the satellites and sensors e.g. GCOS indicate that a higher resolution than is currently provided for fire detection is needed. Other recommendations from GOFC-GOLD that products are provided at their native resolution are also considered and applied where possible. Considering these needs and recommendations the final merged products are at PIXEL resolution (1/120 then 1/360 with MERIS) and GRID (0.5 degree) resolution. The final product will also be released in user friendly outputs, the PIXEL product as a GeoTiff file and the GRID product in the NetCDF format adhering to the conventions of the modelling community. The temporal reporting accuracy requirements are daily, but cloud coverage and sensor temporal resolution will reduce this to ±3.5 days as stated in the PSD (Chuvieco et al. 2013). The BA PIXEL product will be accumulated in monthly temporal composites, storing the detection day of each pixel, sensor combinations, confidence levels and land cover burned. The BA GRID product accumulated as 15 day (bi-monthly) time steps recording the sum of burned area, error in burned area estimate, fraction of surface observed, number of patches, and a summary of area burned for each of 13 different land covers. 3.3 Main considerations in algorithm design To meet the specifications of the PSD (Chuvieco et al. 2013) the main considerations in algorithm design are that the algorithm can: process different resolutions of BA data process a number of different combinations of BA data sets whose coverage will vary spatially and temporally from each other detect and correct double counting of BA between products between time steps provide a check and correction based on the performance of the algorithm when only one product detects a burn Specific considerations for the PIXEL product are that the algorithm can: assign the earliest date of burn detection record the combinations of BA data sets recording the burns assign the highest confidence level assign a BA to a land cover from a land cover reference source output the results as monthly compilations in a GeoTiff format at the best resolution of the sensor combinations Specific considerations for the GRID product are that the algorithm can: Aggregate burned area and land cover statistics from the PIXEL product into each 0.5 degree cell Calculate the standard error in burned area Calculate the latitudinal variation in area per 0.5 degree cell Provide information regarding how well a cell was observed during the aggregation time period Output the results as half monthly compilations in a NetCDF format for use by the climate research group and wider climate community D3.6.3 ATBD III - BA Merging Page 5

13 4 Review of approaches to create EO data products This section reviews and classifies different processing approaches applied to remote sensing data to produce thematic products, techniques that could be replicated or hybridised to the Burned Area (BA) merging. There are few examples of global multi- sensor burned area products to learn from so the review also draws from the experience of other single and multi-sensor thematic products. A brief review is made of some non-fire products, followed by a review of some products from the Earth observation fire community. The methods of production are then summarised and the final section then outlines major differences of fire data sets, which need to be considered in the creation of the Burned Area Product. 4.1 Non-fire products Single sensor products Globcover (Bicheron et al. 2008) is a land cover map derived from the Medium Resolution Imaging Spectrometer (MERIS). This is not a merged product but the Earth was divided or stratified into 22 specific areas to: account for regional variation such as cloud cover and regional vegetation characteristics and minimise reflectance variability to improve classification techniques (Figure 5). In each area the land cover classification procedure then followed four steps: 1) general classification, 2) a classification based on temporal changes, 3) clustering of step 2) into manageable groups and, 4) class labelling based on decision rules and knowledge of land cover experts. The advantage of this method is that each sub area is processed in different and appropriate ways. Figure 5: The 22 unit stratification for Globcover based on global vegetation characteristics (from Bontemps et al. 2010) Global land cover data sets have also been combined to produce a new product by recalibrating one data set to another data set. This was applied to maps of vegetation characteristics produced by spectrally unmixing Advanced Very High Resolution Radiometer (AVHRR) data and a global classification of AVHRR data into land cover types. The new map was produced by rescaling and merging the existing mixture model product with the existing land cover classification (DeFries et al. 2000). A common method used to merge data from the same sensor is referred to as image fusion (Pohl and van Genderen 1998) that can increase the accuracy and utility of the image data. The method uses the original post processed pixel data and increases the spatial resolution of the multispectral bands by splitting and weighting the spectral content of the multispectral band with the higher resolution D3.6.3 ATBD III - BA Merging Page 6

14 panchromatic band. Sensors that have this capability include Ikonos, QuickBird, SPOT and Landsat TM/ETM. There are many image processing techniques that are applied to do this e.g. intensity, hue and saturation, principal components analysis, and the Brovey transformation. In an early demonstration of image fusion Cliché et al. (1985) used aerial imagery to simulate image fusion of SPOT Panchromatic and multispectral bands Multi-sensor products Image fusion is not confined to single sensor data and researchers have also progressed to fuse data from multiple sensors, for example Meng et al. (2010) combine Ikonos and QuickBird and QuickBird and OrbView-3 images. Image fusion of multiple sensor data can have the advantage of including a broader range of spectral information that is then enhanced spatially. The Global Land Cover map 2000 (GLC2000) is mostly created from SPOT S1 VGT data products although some of the regional groups used a multi-sensor approach, e.g. South America (Eva et al. 2004) and Africa (Mayaux et al. 2004). The South American map combined Along Track Scanning Radiometer (ATSR), SPOT VGT, Japanese Earth Resources Satellite 1 (JERS-1) and the Defence Meteorological Satellite Programme-Operational Line Scan System (DMSP-OLS) and the African map used the same sensors without the ATSR data. Each sensor was considered to best represent a particular set of land cover types owing to the technical characteristics of each sensor, ATSR to discriminate for humid forest cover, SPOT VGT to discriminate seasonal forest, JERS-1 radar backscatter for water dominated regions and DMSP-OLS to detect night lights for urban areas. The ATSR and JERS-1 best performing regions were stitched into the SPOT VGT classification, and the DMSP-OLS data was used to identify training sites to improve urban area classification in the SPOT VGT data. In the Ocean Colour community there are examples of merged products from the NASA Ocean colour MEaSUREs (Making Earth Science data records for Use in Research Environments) project, which linked together data from a constellation of satellites. Rather than using a secondary classification product the Ocean Colour MEaSUREs product used data from individual channels of the Sea viewing Wide Field of View Scanner (SeaWiFS), Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua and MERIS using the L3 output data filtered with the highest quality flags (Maritorena et al 2002, Maritorena and Siegel. 2005). A further development was made with GLOBCOLOUR where the merging generally involves averaging pixel values at the same locations to defined resolutions, and then aggregating to a specific time interval creating averages/ weighted averages and errors. The product vastly improves spatial and temporal coverage of individual sensors (Maritorena et al. 2010) InfraRed and microwave Sea Surface Temperature (SST) data sets have been merged, for example Guan and Kawamura (2004) used AVHRR, Geostationary Meteorological Satellite, Stretched-Visible InfraRed Spin Scan Radiometer (GMS S-VISSR), Tropical Rainforest Measuring Mission Microwave Imager (TRMM MI) and Tropical Rainforest Measuring Mission Visible and InfraRed Scanner (TRMM VIRS) data using objective analysis based on the Gauss-Markoff theorem. To merge the product the order of each data set was prioritised according to spatial and temporal coverage then merging the best pair of correlated data sets. A linear minimum least square estimator was applied under certain criteria; otherwise the mean of the selected products was taken as the new SST. The outcome was that low resolution microwave could see through the clouded areas and the high resolution infra-red picked up detailed thermal surface features. There is now a coordinated project to produce high resolution seas surface temperatures (Group for High Resolution for Sea Surface Temperature (GHRSST) who merge satellite and other seas surface temperature data sets together for scientific use ( (last accessed 20/10/2013)). A number of regional centres merge regional SST data (e.g. satellite and buoys) to a defined set of specifications and then transfer the data to a global compilation centre which calculates a global daily SST at depth plus two hourly estimates of SST at the surface. In the domain of atmospheric chemistry there are also examples of data merging. Using total column ozone measurements Nirala (2008) developed a method to merge combinations of MODIS Terra and Aqua, Atmospheric InfraRed Sounder (AIRS) and Ozone Monitoring Instrument (OMI) using means, maximum likelihood, interpolation and the cumulative semivariogram technique. As well as increasing the spatial coverage it was possible to determine the best correlations between the different combinations of sensor data. D3.6.3 ATBD III - BA Merging Page 7

15 4.2 Fire and burned area products Many multi-sensor approaches using fire and burned area data sets are not merged products and data from different sensors are more commonly used as comparative and complimentary information to understand particular (environmental) problems at localised study locations. Eva & Lambin (1998) give examples of the problems encountered whilst calibrating fine resolution data with coarse resolution data for BA estimates in Africa, Stolle et al. (2004), compared the performance of different fire data sets with different fire regimes detected by BA maps in Indonesia and, Bradley and Millington (2006) combined several active fire and burned area estimates to understand burning in specific Andean eco-regions. On the global scale there are some global multi-sensor data sets that are specifically produced to suit the global modelling community. Products include BA data for single sensors and products with statistics that include the results of BA detected by several sensors. These data sets are reviewed in more detail Single sensor fire and burned area data sets GLOBSCAR applied two different algorithms K1 using Near InfraRed (NIR) and Thermal channels, (Piccolini 1998), and E1 using Visible (VIS), NIR and Thermal, (Eva and Lambin 1998), on ATSR-2 sensor data and produced a burned area product for year As the performance of each algorithm varied over different land cover types the two results were then combined using AND in the merging, thus resulting in high omission errors because this did not allow for one of the sensors if it produced good results (Simon et al. 2004). The processing behind the Global Burned Area 2000 (GBA2000) product (Tansey et al. 2004) involved stratification of the globe into geographical areas where a different algorithm or combinations of algorithms were used to detect BA (Figure 6). Although GBA2000 overcame the disadvantage of applying a single algorithm to a global data set the product is constrained to the technical limitations of one sensor, e.g. overpass times and the spectral range of (SPOT VGT). Figure 6: Stratified areas used for the GBA2000 product. These units tended to be based areas of interest specific to the scientific groups in the consortium rather than representation of global environmental conditions (from Tansey et al. 2002) Multi-sensor fire and burned area data sets GLOBCARBON included a burned area estimate created from ATSR-2 and SPOT VGT data between 1998 and The GLOBSCAR algorithm (Simon et al. 2004) was applied to the ATSR-2 data and two of the GBA2000 algorithms, IFI (Ershov and Novik 2001) and UTL (Silva et al. 2002), were applied to the SPOT VGT data. The burned area output was then combined with active fire data from three separate sources Fire-M3, TRMM fires and the ATSR-World Fire Atlas as a reference source for the end user to make an informed decision in the burn confidence. The results were summarised into files consisting of: (1) a pixel level output of the date of detection of a burn with a separate line recording every hit by every sensor; (2) a pixel level output recording the date of detection of a burn with just one line even if there were multiple hits at the same location; (3) a D3.6.3 ATBD III - BA Merging Page 8

16 1 km monthly global product recording which algorithms hit a particular location and which fire data base showed a coincident fire and (4) a 10 km and 0.5 degree aggregated product recording the proportion of burned area and dispersion of burned area per grid cell where two or more detections were found. The Global Fire and Emissions Database-3 product (GFED-3) is a multi-sensor product (at 0.5 degree resolution) from 1997 through 2008, the core data being derived from the MODIS Burned Area Estimate (BAE) product. This data set is gap filled with a burned area estimate from MODIS active fire data and the time series is extended using ATSR and TRMM VIRS active fire data (Giglio et al. 2010). Where there are gaps in the MODIS time series burned area is estimated from the number of active fire counts in the same grid cell. The relationship between fire counts and burned area is derived from (i) a local regression where there is enough data for correlation in a grid square or (ii) where there is not enough data in (i) the estimate of burned area is done via a regression tree which considers variations in vegetative cover for 14 geographical areas of pooled burned area and active fire data. ATSR and VIRS fire counts are calibrated when there is a coincidence in the timing of MODIS burned area data. There is a lower count of active fires from ATSR and VIRS so a regression with MODIS fire counts is used to infer how much burned area there is using the ATSR and VIRS fire counts. With this estimate it is possible to extend the burned area estimate backwards prior to MODIS active fire records, however VIRS is limited to tropical regions because of the TRMM platform orbit. Where there are missing observations in grid cells the procedures (i) and (ii) are repeated for ATSR and VIRS. In terms of quality VIRS takes precedence over ATSR, then because VIRS coverage is absent at the poles ATSR data is used anyway. 4.3 Method summary Stratification and global processing From the examples global processing generally follows two modes, either processing the whole (global) data set or stratifying the data into small units. Global processing is susceptible to variable performance of the sensor(s)/algorithm(s) in different biomes, but is less complex to apply. The alternative, stratification requires the use of ancillary data to split the region of interest (e.g. the Earth) into geographic areas with similar physical characteristics that may be by climatic conditions, vegetation composition and boundaries such as biome interfaces and mountain ranges. The use of decision trees at this level can optimise processing even further. The processing (and/or multi-sensor merging) is then specific to or can be adjusted to suit each of the stratified areas improving the overall performance of a product Gap filling and time series extension This method selects a primary sensor to produce the bulk output of the product, possibly because it has the most complete time series or because the sensor has the best overall performance. Where there are spatial or temporal gaps in the product, because of no coverage or missing data and where data does not pass data quality thresholds, a secondary sensor is used to produce the product and fill the gaps. The use of the secondary sensor may be selected to continue the time series forwards if the mission lifetime of the primary sensor is reached or backwards if the secondary sensor began before the lifetime of the primary sensor Combining spectral data This method uses the original pixel data from several sensors with spectrally different or spectrally similar bands or channels producing a dataset that enhances the spectral information of the target by increasing the number of channels. The extra channels cover more regions in the electromagnetic spectrum and allow construction of more sophisticated detection algorithms. This is a process that is applied in the pre-processing stages and not retrospectively on a remote sensing product. Image fusion also combines spectral data but enhances the original spectral detail of the existing image data, e.g. panchromatic sharpening of multispectral bands. In principle a high resolution BA data set may be able to sharpen the confidence levels of lower resolution BA data sets, the difficulty is though D3.6.3 ATBD III - BA Merging Page 9

17 that this method would require BA data that gives low or zero confidence levels where there are no burns, currently BA data sets are discontinuous and do not provide information outside the BA Blending products Comparing and reporting different products does not give a new multi-sensor product, it does however demonstrate how confidence and uncertainty can be estimated for BA products. For example, GLOBCARBON provided summary statistics that included information for several sensors and Boschetti et al. (2006) illustrated how levels of agreement could be graded between data sets. 4.4 Contrasts between burned area and other EO products Products such as global land cover, ocean colour and SST maps have a gradual variation in pixel values in files containing continuous data fields. It is easier to merge and mosaic continuous fields from different sensors as the object being measured is generally persistent year on year and the spatial variability in sensor performance can be assessed easily. To merge data from different sensors it is possible to match pixels at corresponding locations and derive an approximated value often by averaging. When there are obvious gaps in the data field the blank pixels can be flagged as such or gap filled with data from another sensor if available. Burned area data sets do not share these characteristics as burns are more discrete spatial entities, discontinuous fields that vary from year to year in terms of extent and location that sometimes can only be observed in a certain time frame. There is high probability of missed detection and it is often difficult to know with complete certainty if a sensor has actually observed the surface when a burning event had taken place and if the fire or BA algorithm is good enough to detect all fires or BA anyway. The implications for a BA merged product are that as well as providing a measure of confidence in the area detected as burned there needs to be some information about how certain the areas between burns were not burnt, or indeed if these areas were actually observed. This is not a strong element in current burned area data sets as indicated in section 4.2, however a measure of confidence and uncertainty in burn detection and records of non- observation may assist this. D3.6.3 ATBD III - BA Merging Page 10

18 5 Uncertainty and confidence in fire products Uncertainty in fire related products may be due to omission and commission, multiple observations of the surface (at high latitudes) and non-observation of the surface (see Pereira (2003) for details). There are also additional uncertainties in the pre-processing such as atmospheric correction, masking of water, snow, cloud, cloud shadow and topographic shadow. A simple way to measure the performance of a remote sensing product is to make an inter comparison with other relevant products. Inter comparisons have been made between burned area products and fire data (Roy et al. 2008), between day (MODIS Terra) and night time (MODIS Aqua) fires (Giglio et al. 2006) and on a smaller scale Duncan et al. (2010) has demonstrated how confidence levels can be graded by comparing field observations of BA with BA derived from Landsat TM images. The performance of a product can also be quantified by validating representative land cover types with higher resolution data, for example the L3JRC product correlated SPOT VGT BA against BA on Landsat scenes (Tansey et al. 2008). Measures to quantify missed observations include reporting the numbers of unmapped observations of the surface due to snow, cloud cover, unknown land surface status and missing data due to instrument or data processing failures (Roy et al. 2008). These errors could also cause BA products to have lagged time stamps. Merged products present their own additional uncertainties. In the GFED3 product uncertainty was calculated for all the regression procedures and correction factors between sensors. Uncertainty estimates were: (a) a top down estimates based on MODIS v Landsat validation (MODIS BA v residual of burned area) and (b) regression uncertainties between the burned area and fire counts for either individual cells or nodes of the regression tree. These uncertainties are then summed (Giglio et al. 2010). The use of multiple sensors to gap fill and extend the time series also means that an uncertainty factor is reported that accounts for all combinations of sensors, including the fire count regressions from the MODIS and ATSR/VIRS data. Overall the final data set has an uncertainty measure that varies over space and with time. D3.6.3 ATBD III - BA Merging Page 11

19 6 The Merging Algorithm 6.1 Development The method summary in section 4.3 shows that there are several ways that remote sensing data can be combined, and during the development of the burned area product between ATBD III v0 and ATBD III v1 four potential merging strategies were proposed and tested where possible. These strategies are (i) Blanket overlay, (ii) Core global dataset (iii) Core regional mosaic and, (iv) Land cover mosaic. The four proposed methods of merging are outlined noting their relative merits Merging strategies (i) Blanket overlay This involves an overlay or blending of all products, and the most certain burns are where there is greatest agreement. This follows the model of GLOBCARBON. This approach decreases the uncertainty in the detection of a burned area, i.e. burned area mapped from three sources is more likely than a burn mapped from one source. The sensor combinations are recorded, a statistical average of the confidence levels is made, the earliest date of detection can be identified, and the land cover burned can be identified on a pixel by pixel basis. A drawback of this method is that it is possible that there may be omission/commission errors from certain products where algorithms under/over perform for certain vegetation types. This may be problematic where only one sensor has detected a burn or the coverage is such that only one sensor has actually observed the surface. Without comparison of another sensor or details of non- observation of the surface it is impossible to reduce the level of uncertainty. This method is feasible but is always susceptible to greater uncertainty when there is an incomplete spatial or temporal overlap between the source data sets (Figure 1 and Figure 2). (ii) Core global dataset This approach uses a single BA data set as the core dataset, e.g. the best globally performing algorithms were selected from GBA2000 to create L3JRC, and where there are gaps in the monthly time series or areas of low quality, BA can be used from another dataset, e.g. the basis of GFED3. The advantage of this method is that it ensures some continuity / stability in the global BA product. If the product is gap filled these lower quality data areas can be flagged and reported with a different set of uncertainty measures. As there may be several products detecting the same burn even though the best performing algorithm may have the best confidence level, other results can be used to assess the earliest date of detection of the burn. The disadvantage of this method is that the core algorithm could still have inconsistent performance globally and may still be outperformed by a gap filling data set in certain areas. One solution is to vary the core data set at the regional level (e.g. GLOBCOVER) (see (iii)). (iii) Core regional dataset This approach considers the regional performance of each BA product and the merged product becomes a global stratification of a sub-regional mosaic, a nested version of (ii). This follows the model of GFED3 and GBA2000 and ensures regional continuity /stability of the product. Likewise other products can be used to gap fill and support the BA detection similar to the approach in (ii). As with (ii) this method would rely heavily on knowledge of sensor performance from a retrospective analysis such as a validation. A simple option would be to use the sub regional data areas defined in the Product Specification Document (Chuvieco et al. 2013) such as North America. However difficulties can arise when a sub region is divided across characteristic vegetation zones using an arbitrary boundary (e.g. between Europe and Asia). If different sensors are used then there may be visible errors in BA either side of the boundary because of contrasting sensor performance. For this reason it is more feasible to define sub regions according to environmental factors e.g. using a land cover or ecoregion map (see (iv) below). D3.6.3 ATBD III - BA Merging Page 12

20 (iv) Land cover mosaic This method considers BA product performance in relation to land cover type. This follows the model of GFED3 and GBA2000 and the same principles of (iii) so would need to be qualified by the results of a validation programme. The globe would be stratified using a chosen land cover product but this may still carry some problems. A land cover mosaic may be hindered by uncertain boundaries between land cover types and the quality of assignment to a different land cover class which may also vary across boundaries (e.g. GLC2000). There is also the added difficulty in that interfaces between land cover types are often uncertain as they are commonly susceptible to human encroachment and fire (particularly forest boundaries). Using up to date land cover maps (annual if possible) may reduce these effects and keep up with the dynamics of land cover change Merging solution The algorithm has provision to process 10 by 10 degree global tiles of BA data at two different resolutions 1/120 and 1/360 (Table 1) where ATSR-2, AATSR, SPOT VGT are resampled to 1/120 and have been processed with algorithm 1 (Algo1) and the MERIS FRS data are resampled to 1/360 and have been processed with algorithm 2 (Algo 2), (Krauß et al. 2013). Table 1: The resolution of the BA output for each sensor source and algorithm used for burned area detection Sensor data Merge Algorithm BA algorithm product resolution ATSR-2 1 1/120 AATSR 1 1/120 SPOT VGT 1 1/120 MERIS Full Resolution (FRS) 2 1/360 The merging algorithm draws on the blanket merge and land cover map stratification methods (section 6.1.1), using all sensor data together where possible and applying corrections where only one BA data set records a burn. The correction thresholds are varies according to the land cover type ensuring the variable performance of algorithms is accounted for. All data are ingested into the merging algorithm and BA products are merged at their parent resolution when possible (Figure 7) to avoid resampling errors. Two stages of the merge are applied depending on MERIS availability. The primary merge which processes and fuses the 1/120 data is applied first and a secondary merge at 1/360 is applied when high resolution MERIS data are available. In the secondary merge the primary 1/120 merge data are resampled to 1/360 and the comparison and fusing is repeated at 1/360. (A)ATSR BA Res: 1/120 SPOT-VGT BA Res: 1/120 Merged BA Res: 1/120 Merged BA PIXEL product No MERIS: 1/120 with MERIS: 1/360 MERIS (FRS) BA Res: 1/360 Merged BA Res: 1/360 Aggregated BA GRID product Res: 0.5 Figure 7: Spatial resampling in the merging process. Data processing without MERIS follows the dashed arrows. D3.6.3 ATBD III - BA Merging Page 13

21 BA is recorded in the merged product when a BA pixel is coincident between more than one BA product. This avoids the need to know about sensor performance in a particular environment and avoids the problems of identifying the best performing BA product to use as a core data set or to prioritise in a stratified regional or stratified land cover mosaic. In these cases the uncertainty in timing is decreased by selecting the earliest date of detection and the confidence in BA detection is increased by selecting the highest internal confidence measure from the BA products. The land cover burned is recorded by extracting a land cover class from a land cover reference data set. When only one product has recorded a BA pixel the algorithm screens the data using a threshold on the confidence levels. By considering thresholds for each land cover type, this stage of the merging follows the land cover map stratification method (section 6.1.1). The confidence level thresholds are derived from a comparison of each individual BA product against a burned area reference data set (see Appendix 1). If there is MERIS data the confidence screen is applied in the secondary merge only, as it would be possible to screen out single sensor detections in the primary merge before the algorithm has chance to check if the MERIS BA product may have also detected that BA. After his step has been completed the earliest date of detection, sensor combinations, confidence level and land cover burned is revised for double counts with the previous month because up to 5% of burns can be double counted between months between sensors (Bradley et al. 2012). A second contextual check on the single sensor observations is then made. This procedure can clean burned area edges and removes single or small groups of outlying burns that are the most uncertain BA pixel detections. This is done by calculating uncertainty of the merged BA (see Appendix 2) and applying a threshold on the uncertainty where single sensor observations occur (see Appendix 2 (a)). The uncertainty thresholds are calculated in the same way as the confidence levels by comparing the uncertainty values to a reference data set over different land cover types. BA data can be eliminated from the product if it falls under the uncertainty threshold. The global 10 degree tiles are then mosaicked into the final PIXEL product global sub areas (see Chuvieco et al. 2013). Production of the GRID product requires the screened PIXEL product. The date of detection is used to divide the data into half months (~15 periods) and the date of detection is again used to calculate the sum of burned area, number of patches and sum of area burned per land cover as well as latitudinal area corrections not allowed for in the projection of the source data into 0.5 degree grid cells. A half monthly not valid pixels layer which comes direct from the BA algorithm is also used to calculate the proportion of the area observed in the 0.5 degree grid cells in each 15 day period. The sum of patches is then compared to a reference data set to give the standard error in the sum of burned area (see Appendix 2 (b)). Each 10 x 10 degree tile (now at 0.5 degree resolution) is mosaicked into the global data set Subsequent requirements from the users expressed a need for processing of single sensor burned area products into PSD-class output format. The necessary adaptations to the processing flow were made to ensure that this could be carried out by specifying at the command line what the user requirements are. A summary of the processing flow is detailed in Figure 8. D3.6.3 ATBD III - BA Merging Page 14