Deriving Water Fraction and Flood Maps From MODIS Images Using a Decision Tree Approach

This study investigates how to derive water fraction and flood mapping from the Moderate-Resolution Imaging Spectroradiometer (MODIS) onboard the Earth Observing System (EOS) satellites using the linear mixture model and decision-tree approach. The recent floods in the Midwestern United States in June 2008 and in the New Orleans area in August 2005 were selected for this study. MODIS surface reflectance with the matched land cover data in the Midwest prior to the flooding events were used for the training dataset, with the split test mode of 50% for training and the remaining 50% for testing. The precision, or accuracy rate, of the water classification reaches over 90% from the test. Our results demonstrate that the reflectance difference (CH2-CH1) between the MODIS channel 2 (CH2) and channel 1 (CH1) is the most useful parameter to derive water fraction from the linear mixture model. Rules and threshold values from the decision tree training were applied to real applications on different dates (June 1, 17, and 19, 2008 for the Midwestern region of the U.S.) and at different locations (New Orleans in 2005) to identify standing water and to calculate water fraction. The derived water fraction maps were evaluated using higher resolution Thematic Mapper (TM) data from Landsat observations. It shows that the correlation between water fractions derived from the MODIS and TM data is 0.97, with difference or “bias” of 4.47%, standard deviation of 4.40%, and root mean square error (rmse) of 6.28%. Flood distributions in both space and time domains were generated using the differences in water fraction values before and after the flooding.