Machine Learning Based Bias Correction for Subseasonal Indian Monsoon Precipitation Forecasts

Accurate subseasonal precipitation forecasts are crucial for water resource management, agricultural planning, and climate adaptation strategies, yet significant biases persist in numerical weather prediction at these timescales. This study presents an analysis of a hierarchy of bias-correction models for subseasonal Indian monsoon precipitation deterministic forecasts from the Global Ensemble Forecast System (GEFS), comparing linear regression (LR), multilayer perceptron (MLP), convolutional neural networks (U-Net), and an ensemble approach—Bayesian-optimized stacking (BOS)—which adaptively stacks predictions from these three models via Bayesian optimization. The value added by these machine learning models is found to be small compared to LR, but that BOS consistently outperforms the individual models. No individual model significantly outperformed the others, when evaluated over the Indian region at lead times of 1--7, 8--14, and 15--28 days. However, the BOS approach produced statistically significantly better results than LR and MLP (p < 0.004). Specifically, BOS reduced spatially averaged mean squared error (MSE) by 6.2% for 1--7 days, 4.0% for 8--14 days, and 2.2% for 15--28 days relative to LR—surpassing the corresponding gains from U-Net (4.6%, 1.6%, and 0.7%, respectively). Further analyses across seven homogeneous climate zones in India showed BOS consistently improved upon its constituent models for most regions and lead times. Overall, BOS provides robust improvements over existing approaches while remaining computationally efficient and operationally feasible for subseasonal forecasting across South Asia, effectively addressing the limitations of individual machine learning models.