Mapping Fractional Land Use and Land Cover in a Monsoon Region: The Effects of Data Processing Options
Existing global land use/land cover (LULC) raster maps have limited spatial and thematic resolution relative to the strong heterogeneity of agricultural landscapes. One promising approach to derive more informative maps is using fractional cover instead of hard classification. Here, we evaluate the effect of three key data processing options on the performance of random forest (RF) fractional cover models for moderate resolution imaging spectroradiometer (MODIS) data in a heterogeneous agricultural landscape in a monsoon region: 1) selection of spectral predictor sets [normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), surface reflectance (SR), and all combined (Full)]; 2) time interval (8-day vs. 16-day); and 3) smoothing (no smoothing versus Savitzky–Golay (SG) filter). Model performance was assessed with spatially stratified root-mean-square error (RMSE), Spearman’s rank correlation, and , per LULC type and averaged over all types. We found adequate performance of the best model (avg. ) that used all predictors, 8-day interval and no smoothing. Among the different alternatives, the choice of predictors accounted for 36.3% of the variation, smoothing for 19.0%, and time interval for 17.9%. The intrinsic dimensionalities of the spectral predictors were investigated to complement the variable importance analyses. Although predicting LULC fractions for minor types remained difficult, our results suggest that existing satellite products can be a useful source of information about LULC at subpixel level provided the data-processing options are properly chosen.