Deep learning for uncertainty quantification of satellite retrievals of precipitation: Case studies in two complex terrain regions

The performance of various composite satellite precipitation products is severely limited by their individual passive microwave (PMW)-based retrieval uncertainties because the PMW sensors have difficulties in resolving heavy rain and/or shallow orographic precipitation systems, especially during small scale precipitation events. Characterizing the error structure of PMW retrievals is crucial to improving precipitation mapping at different space-time scales. This paper presents an ensemble learning framework to quantify the uncertainties associated with satellite precipitation products with an emphasis on orographic precipitation. A deep convolutional neural network is devised, which utilizes ground-based radar and gauge blended precipitation estimates as target labels to train satellite precipitation products in order to extract the uncertainty features involved in the satellite products. An ensemble strategy is designed to boost the performance of individually trained deep learning models. The ensemble model is then applied to multiple domains with different geophysical characteristics. The precipitation products derived using the NOAA/Climate Prediction Center morphing technique (CMORPH) over Taiwan and the coastal mountain region in the western United States are used to demonstrate the deep learning-based bias correction performance. The impact of topography on satellite-based precipitation retrievals is quantified. The results show that the orographic gradients have a strong influence on precipitation retrievals in complex terrain regions. The accuracy of CMORPH is dramatically enhanced after applying the ensemble learning-based bias correction technique, indicating the great potential of machine learning in improving satellite precipitation retrievals.