Forest Cover Change: Using Rapid Eye Data to Validate the Accuracy of Landsat Data in Brazil Alexandra C. Rodriguez, Jacob A. Noel, Matthew C. Hansen, PhD BSOS SRI 2014 – University of Maryland, College Park Using remote sensing for forest monitoring is important for quantifying changes to this key natural resource. Brazil, for example, has been able to reduce deforestation by using remote sensing technology to track deforestation in near-real time. Methods have also been developed to map global forest cover loss and gain (Hansen et al., 2013). Such products allow for the estimation of habitat loss, greenhouse gas emissions due to deforestation, and other important measures of environmental change. Landsat data, with a spatial resolution of 30 meters, have been the main satellite data source for forest monitoring and were used to create the aforementioned global product. More recently, Rapid Eye data, with a 5 meter resolution, have become available to researchers. Rapid-Eye derived forest cover loss, quantified at 5m, can be held as ‘truth’ in assessing the accuracy of the global Landsat 30m product in detecting and estimating area of forest cover loss. The study presented here is a preliminary comparison of Landsat- and Rapid Eye- derived forest cover loss for a number of sites over Brazil. Abstract The overall results of the 9 sample blocks: • Rapid Eye (5m) Image 1022373: Forest cover change is. 691,702.7 acres (2799.2 km²) (Figure 4a) • Landsat Data (30m) Image 1022373: Forest cover change is 406,739.8 acres (1646 km²). (Figure 4b) • Difference between Rapid Eye and Landsat estimated area is 284,962.9 acres (1153.2 km²). (Figure 4c) There is more forest cover change being detected through Rapid Eye data than Landsat data. Data A total of nine sample blocks were used for this analysis. For each sample block, Rapid Eye data were collected for two dates between 2010 and 2013. Mapping and change detection was implemented using a classification tree algorithm based on user applied training data resulting in a change/no-change map (Noel et al., 2013). Decision tree (“Classification and regression trees” – CART; Breiman et al., 1984) is a hierarchical classifier that predicts class membership by recursively partitioning a data set into more homogeneous subsets. This splitting procedure is followed until a perfect tree (one in which every pixel is discriminated from pixels of other classes, if possible) is created with all pure terminal nodes or until preset conditions are met for terminating the tree’s growth. The result from the tree algorithm will then be compared to the 2013 Hansen et al. Global Forest Change product. EXAMPLE BLOCKS Figure 4 depicts Date 1 and Date 2 of the sample block. This block is located in the third largest, area wise, Brazilian state of Mato Grosso. It is located in the western part of Brazil. The sample falls under the Tropical Dry stratum that is experiencing > 2.10% forest cover change over 3 years. Date 1 (Figure 4a) is dated September 2010 and Date 2 (Figure 4b) is dated May 2013. Methods As a result of comparing both satellite data products, Rapid Eye data detects forest cover change where Landsat data was not able to. Due to it’s coarser resolution, Landsat data underestimates the total area of forest cover change. The goal over time is to improve analysis by reviewing the error found in the Confusion (Error) Matrix (Table 3) and finding a solution to diminish the error as much as possible. This error could be as a result of a geographic shift between both satellite products. Also, the difficulty of knowing exact clearing dates plays a factor. Lastly, improving the user applied training data would lessen the error. This analysis is only the first step of providing more accurate and precise data regarding forest cover loss and gain. A finer resolution will be able to provide countries like Brazil the opportunity to improve their existing data in order to implement better policies and regulations Discussion Landsat Wavelength (mm) Resolution (m) Band 1 0.45 - 0.52 (blue) 30 Band 2 0.52 – 0.60 (green) 30 Band 3 0.63 – 0.69 (red) 30 Band 4 0.76 – 0.90 (NIR) 30 Band 5 1.55 – 1.75 (SWIR) 30 Band 6 10.40 – 12.50 (SWIR) 120* 30 Band 7 2.08 – 2.35 (TIR) 30 The processed data is a composite of images taken over a period of time. The Landsat data being validated is taken from 2000 to 2012. The Rapid Eye data used as reference is taken from 2010 to 2013. A 400km² sample block was then created for various locations in Brazil. A few images have cloud interference but that it is removed through cloud masking. Landsat data contains 7 Spectral Bands but for this study only Bands 3 to 7 were used. Rapid Eye data contains 5 Spectral Bands and all of them were used. Results Figure 4b: Forest Cover Change Using 30m Landsat Data Table 1. Landsat Thematic Mapper (TM), USGS. Rapid Eye Wavelength (nm) Resolution (m) Band 1 440 – 510 (blue) 5 Band 2 520 – 590 (green) 5 Band 3 630 – 685 (red) 5 Band 4 690 – 730 (red edge) 5 Band 5 760 – 850 (NIR) 5 Table 2. Rapid Eye Satellite Sensor Specifications Figure 1: Study area of the 9 sample blocks located in Brazil Figure 4a: Forest Cover Change Using 5m Rapid Eye Data Figure 4c: Spatial Comparison of Rapid Eye and Landsat Change Products Figure 4c Figure 4b Figure 4a TABLE 3: CONFUSION (ERROR) MATRIX (1022373) LANDSAT NO LOSS LOSS TOTAL USER'S ACCURACY PRODUCER'S ACCURACY RAPID EYE NO LOSS 12586603 290583 12877186 61.70% 55.60% LOSS 225288 363631 588919 TOTAL 12811891 654214 13466105 OVERALL ACCURACY 96.20% RAPID EYE LANDSAT FIGURE 4: FOREST LOSS AREA, ACRES RAPID EYE LANDSAT 406,739.8 691,702.7 Figure 3: Stratification and sample block locations ³ Stratified random sampling was used in this analysis. Stratification of these Biomes were based on forest loss data from 2007-2009 (Figure 2) which were applied to estimate 2010-2012 change. Criteria for each strata within each respective biome: - *minimal % change over 3 years time - *moderate % change over 3 years time - *major % change over 3 years time - >= 3 % minimum of loss/gain (rotation) over 12 years (industrial forest) *(The cutoff points for percent change are region specific) Equal allocation method was used for selecting random samples within each biome using this criteria resulting in 25 400 km 2 sample blocks for each of the 5 biomes, resulting in 500 sample blocks. The stratification and block locations are depicted in Figure 3. Stratified Random Sampling 1. Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A.Kommareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Townshend. "High-Resolution Global Maps of 21st-Century Forest Cover Change."Science 342.6160 (2013): 850-53. Web. 2. Breiman L, Friedman JH, Olshen RA, Stone CJ. “Classification and Regression Trees.” (1984) Chapman & Hall, New York. 3. Noel, Jacob A., Alexander Krylov, Peter Potapov, Steve Stehman, and Matthew C. Hansen. "Very High Spatial Resolution Rapid Eye Data Used for Validating Large Area Landsat Based Forest Disturbance Map."Department of Geographical Sciences – University of Maryland, College Park (2013): n. pag. Web. References Figure 2: FAO Biomes