This research estimated the surface water fraction from MODIS 16-day 500 m Bidirectional Reflectance Distribution Function (BRDF) corrected surface reflectance image composites. The spatial extent of two Landsat tiles over Spain were selected as test areas. We obtained a 500 m reference dataset on surface water fraction by spatially aggregating 30 m binary water masks derived from the C-version of Function of Mask (CFmask), which themselves were evaluated against high-resolution Google Earth imagery. Twelve regression tree models were developed with two approaches, Random Forest and Cubist, using spectral metrics derived from MODIS data and topographic parameters generated from a 30 m spatial resolution digital elevation model. Results showed that accuracies were higher when we included annual summary statistics of the spectral metrics as predictor variables. Models trained on a single Landsat tile were ineffective in mapping surface water in the other tile, but global models trained with environmental conditions from both tiles can provide accurate results for both study areas. We achieved the highest accuracy with Cubist global model (R2 = 0.91, RMSE = 11.05%, MAE = 7.67%). Our approach is promising for monitoring surface water fraction at high frequency time intervals over much larger regions provided that training data are collected across the spatial domain for which the model will be applied.
Remote sensing, Hydrology, Machine learning, Time series
|Date made available||2 Oct 2018|
|Temporal coverage||2000 - 2015|
|Date of data production||2018|
|Geospatial point||39.388579, -3.720733|