A New Satellite-Based Precipitation Downscaling Scheme for Data-sparse Areas Using Deep Learning and Transfer Learning

Numerous statistical downscaling techniques were proposed to improve the spatial resolution of satellite-based precipitation data. However, these downscaling methods generally required observation data, making it difficult to apply in data-scarce areas. To address this issue, this study presented a framework based on transfer learning (TL), in which the pretrained convolutional neural network model in source domain was fine-tuned and implemented in target domains. Results showed that the pretrained model cannot be directly applied in transferred regions due to its poor performance, but the TL model with at least one fine-tunable layers achieved significant improvements and can be employed successfully. It was notable that the fine-tuning model obtained even higher accuracy than the model trained independently with data of target domains. Results also showed that TL model attained a higher performance with more fine-tunable layers, and different fine-tunable layers would impact the downscaling results and should be selected during TL.