Deep learning for multi-source precipitation fusion on the Qinghai–Tibet Plateau

Accurate precipitation estimation remains a considerable challenge across the Qinghai–Tibet Plateau (QTP). By integrating the strengths of in situ gauge accuracy and the spatiotemporal continuity of gridded products, this study developed a highly accurate spatiotemporal deep fusion model using Convolutional Neural Networks (CNN) and long short-term memory (LSTM) networks (CNN-LSTM) to combine remote sensing estimates (TRMM-3B42V7, IMERG-FinalV06), reanalysis products (ERA5-Land), and in situ rain gauges. The CNN handles spatial feature extraction, whereas the LSTM captures temporal dependencies. Evaluations at 104 rain gauge stations across the QTP demonstrated that CNN-LSTM model significantly outperforms both traditional ANN/CNN fusion models and individual gridded precipitation products. It achieved an overall correlation coefficient (CC) of 0.61 and a root mean square error (RMSE) of 2.73 mm/d at the daily scale. Furthermore, its advantages are manifested in three key aspects: temporally, it consistently performs best across multiple time scales (daily, monthly, seasonal, and annual); spatially, it performs better at the vast majority of stations, particularly in the southeastern and southern regions with complex precipitation gradients; additionally, it demonstrates superior detection performance across all precipitation intensities, particularly for heavy and torrential rainfall events. This result enhance the understanding of the water cycle in the QTP.