DOWNSCALING CHIRTS GRIDDED AIR TEMPERATURE DATASET INTEGRATING MODIS LAND SURFACE TEMPERATURE ESTIMATES IN MACHINE LEARNING MODELS

This study assesses the potential of integrating various satellite-based products in a Random-Forest (RF) machine-learning model to enhance CHIRTS (Climate Hazards center InfraRed Temperature with stations) gridded estimates across Madagascar. The considered satellite-based products are based upon their close relationships with temperature pattern in space and time. It includes, SRTM digital elevation model (elevation, slope, aspect) along with MODIS NDVI and LST estimates. Julian day is also considered in order to integrate the temporal dynamic. To assess the reliability of the method, the model outcomes were compared with observed temperature records not considered in the RF model training. Results show that the integration of such datasets in a Random-Forest model considerably improve CHIRTS estimates with an overall decrease of RMSE and MAE of 59% and 53%, respectively. Besides, the method also permit to downscale CHIRTS estimates from its native spatial resolution of 5km to 1km.