User Guide for the MODIS Vegetation Continuous Field
Contents
1. 305 337 O clear 1 bad 6 DOY 353 017 O clear 1 bad 7 DOY 033 045 O clear 1 bad Table 1 Description of bit field for Quality sds in VCF product Essentially each bit in the field represents 1 of the 8 input surface reflectance files to the model If the value for that time period had only bad data as defined above the bit is turned on indicating that data for that time period was bad The user should take this information into consideration when evaluating the usefulness of data for a given pixel If the data are bad for 2 or more of the 8 time periods the user should be cautious with the value as it may be biased due to the poor inputs This layer can be used in conjunction with the 4 sds which identifies those bad data pixels which were cloudy 4 3 Percent_tree_cover_SD The percent tree cover SD layer provides the standard deviation SD of the 30 models that were used to generate the pixel value in the percent tree cover data layer This information can be used to determine the level of agreement between the models in the production of the tree cover value Values in this field represent percent cover and can be read as the value shown Obviously a lower value indicates better agreement between the independent models A fill value of 100 indicates that there was not sufficient data to make an accurate prediction and the pixel contains a reasonable value derived from other sources Valid2. Although these results are preliminary we are encouraged by the overall improvement in the VCF tree cover product with available ground based validation data Site Field data Old VCF New VCF Maryland SERC 1 29 16 34 SERC 2 48 61 51 SERC 3 33 40 50 SERC 4 59 61 46 SERC 5 69 40 57 GB 1 67 74 59 GB 2 69 66 68 GB 3 33 74 37 RMSE 19 27 9 47 Mean Absolute Error 14 37 7 87 Mato Grosso Explorada 1 64 49 Explorada 2 80 78 Moth 63 76 Disturbed 64 74 Logged 2 72 79 Logged 55 79 Ik log 50 80 Tower 0 RMSE 10 46 Mean Absolute Error 9 40 Table 3 Validation data from field sites in Maryland United States and Mato Grosso South America 7 Accessing and citing the data The VCF data are provided free of charge The data can be accessed from the Land Processes DAAC https Ipdaac usgs gov Ipdaac products modis products table under the product name MOD44B in MODIS tile format and in alternate formats from the Global Land Cover Facility http landcover org Data should be cited as follows Full Citation Example Townshend J R G M Carroll C Dimiceli R Sohlberg M Hansen and R DeFries 2011 Vegetation Continuous Fields MOD44B 2001 Percent Tree Cover Collection 5 University of Maryland College Park Maryland 2001 digital data accessed MM DD YYYY from lt URL gt The highlighted portion should change to reflect the data year and
3. Range 0 10 000 scaled by 100 Fill Value 100 4 4 Cloud The cloud layer is an 8 bit packed bit layer which clarifies the Quality layer to give the user an indication that the bad data refers to cloudy input data The bit field is described in table 2 below Bit Input layers State O DOY 065 097 O clear 1 cloudy 1 DOY 113 145 O clear 1 cloudy 2 DOY 161 193 O clear 1 cloudy 3 DOY 209 241 O clear 1 cloudy 4 DOY 257 289 O clear 1 cloudy 5 DOY 305 337 O clear 1 cloudy 6 DOY 353 017 O clear 1 cloudy 7 DOY 033 045 O clear 1 cloudy Table 2 Description of bit field for Cloud sds in VCF product As with the Quality layer each bit in the field represents of the 8 surface reflectance inputs The cloud is provided as a clarification to the quality because cloudy data is likely to depress the tree cover value in the model This information is provided to help the user understand potential reasons why values seen in the percent tree cover data layer are high or low and provides the user a tool for determining whether they should trust the result 5 Results Annual results for the VCF product using MODIS Terra data from 2000 to 2009 have been produced for percent tree cover These results figure 1 show expected patterns of tree cover extent There remain some minor confusion with some cropped areas high latitude mountain shadows
4. and some wetlands but overall the output is substantially better than the previous 500m version in spatial detail and coherence In the image pairs in figure 2 the image on left is the old 500m product and the image on right is the new 250m product Both are shown in a 250m grid to emphasize the improvement in spatial detail Figures 2 a and b show improvements in the representation of the ridge and valley system in southern Pennsylvania in the US Figures 2 c and d show clearings in southern Mato Grosso state in Brazil where the new VCF shows values approaching 0 tree cover in the clearings and the old VCF product showed values between 10 and 30 in many cases Finally figures 2 e and f show agricultural areas in southern Brazil The old 500m product showed these areas as having between 10 and 25 tree cover where the new 250m product indicates that the tree cover is near 0 and the forested areas are highly fragmented Figure 2 Image pairs showing the Collection 3 500m VCF on the left and the new Collection 5 250m VCF on the right Darker green color indicates denser tree cover 6 Validation A limited amount of validation has been performed using field data from two sites in Maryland and three sites in Brazil South America Table 3 Initial results show that the new C5 VCF product is substantially more accurate compared to ground based measurements of canopy cover with as much as a 50 improvement in RMSE between the two versions
5. User Guide for the MODIS Vegetation Continuous Fields product Collection 5 version 1 PI John Townshend Co I Matthew Hansen Co I Mark Carroll Co I Charlene DiMiceli Co I Robert Sohlberg Co I Chengquan Huang University of Maryland 1 Table of Contents LT ee Sa aaah a Baa Nel aac had olan eel Fh EA 2 2 A acy sch syle sk Suse aay eas adh cade vhs handed g ewe ad despa pce 3 Bo Algorithmi sonar dapd od Pa is 4 3l Traning Data A hh ae et silt aed heals eas Nate i ae ta 4 Bede DEA MS A 4 Sr Dita Sa 5 34 File Naming CONVE a tb 5 Arn A EAE E AA sue oda E aoa 6 Al Percent O 6 42 QUIN A A A A Gat owed 6 A Percent tree cover AAA A A 7 mis E 6 bc eae eee eae ee T E Se a OE Re gO ESR ROS Re SR NC Re Teen ET re eee ere Urey ote 7 Ds IR SST E SE S EN EEE AE arte elt EEN fal Mace RI Raat 7 O Mad eintrat aneian as T a A id 10 7 Accessing and citing the data A sade nn adie ete eee aan 10 Se Reference A A O A N 11 2 Introduction Characterization of the land surface from satellite data has been performed for over three decades The Vegetation Continuous Fields VCF product is a global representation of the Earth s surface as gradations of three components of ground cover percent tree cover percent non tree vegetation and percent bare figure 1 Carroll et al 2011 Hansen et al 2000 2002 2003 2005 Each pixel is shown as a sub pixel mixture of cover with each of the three components expressed as a percentage of ground cove
6. ciences 10 27 pp 9439 9444 Hansen M Townshend J Stehman S Mayaux P and Morisette J in preparation Recommendations on the validation of vegetation continuous fields cover maps Report from a joint CEOS WGCV and GOFC GOLD workshop on validation of vegetation continuous fields products October 27 28 2005 Brookings South Dakota Hansen M Townshend J DeFries R and Carroll M 2005 Estimation of tree cover using MODIS data at global continental and regional local scales International Journal of Remote Sensing 26 19 4359 4380 Hansen M C DeFries R S Townshend J R G Carroll M Dimiceli C and Sohlberg R A 2003 Global Percent Tree Cover at a Spatial Resolution of 500 Meters First results of the MODIS Vegetation Continuous Fields Algorithm Earth Interactions 7 7 007 Hansen M C Sohlberg R Dimiceli C Carroll M DeFries R S and Townshend J R G 2002 Towards an operational MODIS continuous field of percent tree cover algorithm Examples using AVHRR and MODIS data Remote Sensing of Environment 83 1 2 303 319 Hansen M C DeFries R S Townshend J R G and Sohlberg R 2000 Global land cover classification at 1km spatial resolution using a classification tree approach International Journal of Remote Sensing 21 1331 1364 Masek J G Friedl M Loveland T Brown de Colstoun E Townshend J Hansen M and Ranson K J 2006 ESDR Community White Pa
7. data layer used 8 References Carroll M Townshend J Hansen M DiMiceli C Sohlberg R Wurster K 2011 Vegetative Cover Conversion and Vegetation Continuous Fields In Ramachandran B Justice C O Abrams M eds Land Remote Sensing and Global Environmental Change NASA s Earth Observing System and the Science of ASTER and MODIS Springer Verlag Chang J Hansen M C Pittman K Carroll M amp DiMiceli C 2007 Corn and soybean mapping in the united states using MODIS time series data sets Agronomy Journal 99 6 1654 1664 DeFries R Hansen M Townshend J R G Janetos A C and Loveland T R 2000 Continuous Fields 1 Km Tree Cover College Park Maryland The Global Land Cover Facility DeFries R Field C B Fung I Justice C O Matson P A Matthews M Mooney H A Potter C S Prentice K Sellers P J Townshend J Tucker C J Ustin S L and Vitousek P M 1995 Mapping the land surface for global atmosphere biosphere models toward continuous distributions of vegetation s functional properties Journal of Geophysical Research 100 20 867 20 882 Hansen M Stehman S Potapov P Loveland T Townshend J DeFries R Pittman K Arunarwati B Stolle F Steininger M Carroll M and DiMiceli C 2008 Humid Tropical Forest Clearing from 2000 to 2005 Quantified Using Multi temporal and Multi resolution Remotely Sensed Data Proceedings National Academy of S
8. day composites into 8 composites and is retained during processing to be passed on to the end user Poor quality input data obviously results in poor quality outputs so this information is saved in two quality assurance layers in the final product see section 4 for full description 3 4 File Naming Convention The MODIS Vegetation Continuous Fields product is a standard MODIS product called MOD44B Filenames contain the product ID product date tileID collection number and production date time stamp ex MOD44B A2005065 h09v05 005 2011110122251 hdf and are produced in hdf eos file format with internal compression The product date refers to the start date of the annual period so a product with ID 2005065 was produced with data from 2005065 2006064 The start date of all MODIS VCF products is yyyy065 where yyyy refers to the 4 digit year This originally derived from the first full 16 day composite period in the MODIS data record which begins with day of year 2000065 However it has been continued because it relates better to seasons than the Gregorian calendar If the product were generated starting January 1 and ending with December 31 it would result in splitting the southern hemisphere summer between 2 product years which is less desirable 4 Data Layers In the Collection 5 version 1 release of the MODIS VCF product MOD44B there are 4 science data sets SDS 1 percent tree cover 2 QA bad data 3 QA cl
9. ning software The process employs bagging where 30 independent regression trees are created and the final result is the average of the 30 independent results The regression tree models are created using the M5 regression tree with pruning This process has been used in regression tree models with remote sensing data for over a decade and has been shown to produce more reliable results as compared to a single tree model Chang et al 2007 Hansen et al 2003 Once all 30 models are created they are applied to the MODIS data to yield 30 independent results The 30 independent results are averaged together to yield one result for any given pixel Standard deviation from the 30 results is retained in a QA layer for the end user to understand the amount of agreement between the independent models A global annual VCF data set can be produced in the MODIS Adaptive Processing System MODAPS the PI led processing system for MODIS products in about 5 days with the full ten year record able to be processed in under 6 weeks There are certain areas where persistent clouds or other factors preclude accurate prediction In those cases a reasonable value is inserted from a basemap created by the researchers These values are flagged with a value of 100 in the Percent_Tree Cover SD SDS The surface reflectance composites that form the basis of the VCF product contain a wealth of per pixel QA information This information is used to combine the 23 16
10. nted with new information 3 2 Data Inputs The inputs for the MODIS VCF product are a 16 day surface reflectance composite which includes MODIS bands 1 7 and brightness temperature from MODIS bands 20 31 32 the training data described above and the MODIS Global 250m Land Water Map The surface reflectance composites that are used are an intermediate product generated inside the MODAPS production facility for the VCF product for further information see Carroll et al 2011 There are 23 16 day composites for each year of data One year of 16 day composites are further composited to yield 8 composites per year in order to minimize clouds and as a data reduction step These 8 final composites are used in the data production algorithm 3 3 Data Algorithm The production algorithm for VCF runs in three parts sampling inputs under the training data creating models and applying the models to the output These three steps are accomplished with open source software Weka data mining software and custom software written in C programming language In order to run in the MODIS Adaptive Processing System MODAPS it was necessary to adopt the open source software in lieu of proprietary software that was used in previous versions of the VCF product The algorithm begins with step 1 creating 30 independent samples from the training data set and organizing them for step 2 the model creation step The model creation is performed by the Weka data mi
11. ocess was to develop an updated training data set The spatial resolution for MODIS data used in the VCF Collection 5 product is 250m The training data that were used for previous versions of the VCF product were developed in the late 1990 s and revised circa 2001 This data set represented a high quality data set at the time but relied heavily on Landsat 5 Thematic Mapper data from the 1980 s With the availability of the ortho rectified Landsat Geocover and globally available fine resolution data it was advantageous to create a completely new training data set that better matched the acquisition dates of the MODIS data 2000 to present 3 1 Training Data Training data were created by performing a discrete classification on the Landsat data into 4 classes of relative percent tree cover 0 25 50 80 These relative percentages were verified by overlaying the scenes with fine and ultra fine resolution imagery from Ikonos Quickbird and other data as available In addition comparisons were made to Google Earth where tree crowns can be seen distinctly Adjustments were made to the discrete values as necessary to match observed conditions from the fine resolution data The 30m data were then averaged to 250m spatial resolution yielding a continuous representation of the surface for that scene from 0 100 percent Once created the training data is a static data set However if specific problem areas emerge the training data will be augme
12. oudy data 4 Standard Deviation of models The first layer is the primary data layer with the ensuing 3 layers providing the user with indications of the overall quality of the data for any given pixel 4 1 Percent_tree_cover The percent tree cover layer is a primary data layer which describes the percent of a pixel which is covered by tree canopy This is defined as light penetration to the ground as compared to crown cover which describes the amount of the ground which is encompassed by the tree s crown regardless of whether light penetrates The information in this layer can be used to identify forested areas for a variety of applications from resource management to the creation of plant functional types for climate modeling Valid values in the VCF percent tree cover layer are 0 100 percent tree cover 200 water 253 fill outside of projection 4 2 Quality The Quality sds is an 8 bit packed bit layer which describes per pixel those inputs that had poor quality defined by the MODIS surface reflectance quality assurance In this case we define poor quality as those pixels which are cloudy high aerosol cloud shadow or view zenith gt 45 The bit field is described in table 1 below Bit Input layers State O DOY 065 097 O clear 1 bad 1 DOY 113 145 O clear 1 bad 2 DOY 161 193 O clear 1 bad 3 DOY 209 241 O clear 1 bad 4 DOY 257 289 O clear 1 bad 5 DOY
13. per on Land Cover Land Cover Change NASA Greenbelt MD pp 9 ftp ftp iluci org Land ESDR Landcover change Masek_whitepaper pdf Olson D M Dinerstein E Wikramanayake E D Burgess N D Powell G V N Underwood E C D Amico J A Itoua I Strand H E Morrison J C Loucks C J Allnutt T F Ricketts T H Kura Y Lamoreux J F Wettengel W W Hedao P and Kassem K R 2001 Terrestrial Ecoregions of the World A New Map of Life on Earth BioScience 51 11 933 938 Townshend J R G and Justice C O 1990 The spatial variation of vegetation at very large scales International Journal of Remote Sensing 11 149 157
14. r Figure 1 Global Vegetation Continuous Fields percent tree cover for 2001 Darker greens indicate denser tree cover pale colors indicate light tree cover and gray indicates completely bare The three components are each stored in separate layers so they can be used independently to look at a particular type of ground cover or collectively to look at the entire surface When originally proposed the VCF product represented a revolutionary new approach to the characterization of vegetative land cover DeFries and Townshend 1994 DeFries et al 1997 Conventional land cover classifications suffer from the imposition of arbitrary thresholds between classes and the characterization of the land surface is highly dependent on the a priori class boundaries which are chosen Hansen et al 2002 Moreover when land cover products are used in models parameterization has to be carried out which is often crude and inaccurate By depicting each pixel as a proportion of characteristics such as percentage tree cover non tree vegetation cover and bare ground a genuinely quantitative depiction of land cover became possible The advantages of this approach have been recognized by the widespread adoption of the VCF product by many users in the modeling and monitoring communities The VCF product has also been identified as an Earth System Data Record ESDR by the science community Masek et al 2006 Historically the creation of the algorithms for the prod
15. uction of global land cover maps has been largely hand crafted as human intervention was required to help the algorithm make distinctions between certain land cover types such as wetlands distinction from forest The current version of the VCF algorithm endeavors to minimize the human element and allow the algorithm to make final decisions The early MODIS VCF algorithms were developed using a semi automated process where the regression trees were created using machine learning software These trees were then evaluated by an operator who could then add training at certain branches of the tree or simply eliminate parts of that tree This human interaction was necessary because the training data though very good at the time it was created had some inconsistencies In the new approach the training data has been completely updated using Landsat Geocover data and has been revised and refined using the plethora of fine and ultra fine resolution data available through the NASA science data purchase Google Earth among many others The improved training data and the implementation of new and improved data mining software have resulted in much greater accuracy in the final product without human intervention The final algorithm for VCF operates in a completely automated fashion with the results published upon completion The following pages describe in detail how this algorithm came to be and basically how it works 3 Algorithm The first step in the pr
Download Pdf Manuals
Related Search