Improving Subseasonal Indian Summer Monsoon Rainfall Forecasts with U-Net Calibration

Abstract Accurate subseasonal to seasonal (S2S) forecasts of Indian Summer Monsoon Rainfall (ISMR) are vital for agricultural planning, water resource management, and disaster risk reduction. Conventional post-processing techniques for S2S General Circulation Model (GCM) forecasts predominantly rely on linear methods, which often exhibit limited predictive skill. In this study, we investigate the use of deep learning—specifically, U-Net convolutional neural networks (CNNs)—for improving probabilistic ISMR forecasts. We apply U-Net-based bias correction to three state-of-the-art S2S GCMs: GEFSv12, ECMWF, and IITM ERPv2, training the U-Net models to calibrate tercile probabilities of weekly accumulated rainfall at lead times of 1 to 4 weeks for the monsoon season (June-September) over India. Our results show that the U-Net enhances probabilistic forecast skill, consistently outperforming the baseline Extended Logistic Regression (ELR) approach. Using hindcast data from 1989 to 2022, we observe a substantial improvement in weeks 3--4 probabilistic forecast skill, measured by the Ranked Probability Skill Score (RPSS). The extent of improvement, however, varies with the amount of training data available for each GCM, and can reach as high as a twofold increase, from 2% in the baseline to 4% with our U-Net model for Weeks 3--4. We also construct a multi-model ensemble (MME), which improves skill across all lead times. Additionally, the U-Net demonstrates superior performance to the linear baseline in a real-time forecasting context, for the summer of 2023. These findings underscore the promise of deep learning approaches to improve Indian Summer Monsoon Rainfall forecasts at the S2S timescale.