Development and verification of a non-linear disaggregation method (NL-DisTrad) to downscale MODIS land surface temperature to the spatial scale of Landsat thermal data to estimate evapotranspiration

Abstract A nonlinear method (NL-DisTrad) was developed and tested to disaggregate satellite-derived estimates of land surface temperature of MODIS (Moderate Resolution Imaging Spectrometer) with a resolution of 960 m to the scale of Landsat 7 ETM + (Enhanced Thematic Mapper Plus) at 60 m. This method uses the relationship that is captured at the hot edge pixels in the feature space between the Normalised Difference Vegetation Index (NDVI) and the land surface temperature (LST) at a coarse resolution to disaggregate the LST to a finer resolution. The residuals that are generated at the coarse resolution are modelled using an Artificial Neural Network model (ANN), and the resulting residuals are added to the disaggregated LST at a fine resolution. The ANN model was built using the NDVI from the neighbourhood pixels. The hypothesis is that the LST of a pixel will not only be affected by the vegetation within the pixel but also by the vegetation of surrounding pixels. The performance of this hybrid model NL-DisTrad (Hot edge model + ANN model) is assessed by comparing the results to the existing disaggregation method, TsHARP, and the observed Landsat LST. The NL-DisTrad disaggregation results were comparable to the observed Landsat LST even for pixels with non-uniform vegetation. The statistical analysis indicated that the proposed model disaggregates the LST better than TsHARP, based on the high Nash Sutcliffe Efficiency (NSE > 0.80) and low root mean square error value (RMSE