Harnessing Deep Learning Methods for Rainfall Downscaling and Projections Over India: A CMIP6 Based Assessment of Future Shifts
Abstract
The intensification of hydroclimatic extremes under global warming poses serious risks to water security, agriculture and disaster resilience in monsoon‐dependent regions. However, reliable regional‐scale rainfall projections remain challenging due to scale mismatches between global climate models and localised hydroclimatic variability. To address this limitation, we apply two sequential deep learning (DL) methods, bidirectional long short‐term memory (BiLSTM) and gated recurrent unit (GRU), to downscale daily precipitation from 12 CMIP6 models over India for the historical period (1951–2010) and future projections (2040–2069; 2070–2100) under SSP2‐4.5 and SSP5‐8.5 scenarios. BiLSTM consistently outperforms GRU in reproducing observed rainfall variability, achieving lower RMSE (≤ 4.5), reduced MAE (1.16–1.63) and stronger correlation ( r = 0.85–0.94). The model effectively captures observed annual variability (~14 mm) and monsoon (JJAS) intensity (~24 mm), while substantially reducing systematic overestimation in raw CMIP6 outputs. Spatial gradients across the Western Ghats, Northeast India and core monsoon regions are more realistically represented, indicating improved preservation of regional rainfall dynamics. Future projections reveal a coherent redistribution of monsoonal rainfall. Under SSP5‐8.5 (2070–2100), precipitation increases by up to ~20% over western India and ~40% over parts of southern India, the west coast and the western Himalayas, while central India experiences reductions approaching ~50%. Similar but less pronounced patterns emerge under SSP2‐4.5. The majority of the models reveal notable monsoon shift towards south India in the future projection, under both scenarios. These results suggest intensifying regional hydroclimatic contrasts and enhanced spatial polarisation of monsoon rainfall under continued warming. Overall, DL‐based downscaling substantially improves the spatial fidelity and usability of climate projections for regional adaptation planning.