91ÃÛÌÒ¸ó

Raw precipitation forecasts from numerical weather prediction 91ÃÛÌÒ¸ó often exhibit systematic biases and spatial inconsistencies that limit their direct use in hydrological forecasting applications. We develop a Bayesian hierarchical post-processing framework that integrates spatial clustering with probabilistic modeling to generate calibrated precipitation forecasts at both basin-average and gridded scales. Spatially homogeneous rainfall regions are first identified using a modified partitioning around medoids (PAM) clustering algorithm that combines an extreme-value distance metric (F - madogram) with a spatial penalty to ensure geographically contiguous precipitation clusters. Cluster-average precipitation is then post-processed using a hierarchical Bayesian model that accounts for temporal persistence and spatial dependencies among neighboring clusters via regime-dependent predictors. The same hierarchical framework is subsequently applied at individual grid cells within each cluster, using the covariate structure identified at the cluster level to generate spatially coherent gridded precipitation forecasts across the basin. The framework is demonstrated on the Narmada River Basin in central India using IMD forecasts (2000–2014 July–August) and the larger Brahmaputra River Basin in northeast India using NCMRWF forecasts (2020–2022, JJAS).Ìý

Forecast skill was evaluated at 1–3 day lead times using probabilistic metrics including Ranked Probability Skill Score (RPSS), Brier Skill Score (BSS), and Continuous Ranked Probability Skill Score (CRPSS). Results show substantial improvements over raw forecasts across both basins. In the Narmada Basin, 0.36 ≤ RPSS ≤ 0.46 and 0.32 ≤ BSS ≤ 0.41 across sub-basins for lead times up to three days while in the Brahmaputra Basin, cluster level forecasts show even stronger skill, reaching RPSS ≈ 0.73 and BSS > 0.84 at 1-day lead time in the western cluster, with consistently positive skill across locations and lead times. CRPSS also shows strong improvement, 1-day ≈ 0.55, 2-day ≈ 0.50, and 3-day ≈ 0.45, demonstrating substantial gains in overall probabilistic forecast accuracy. Spatial maps of cluster-wise extreme precipitation events from the gridded post-processed forecasts are generated and compared with observations and raw forecasts after model cross-validation, show promising agreement with the observed spatial precipitation patterns and improved representation of high-rainfall events, demonstrating its potential to provide improved precipitation inputs for hydrological forecasting systems.