Hybrid Integration of Land Surface Models and Deep Learning for Enhanced Streamflow Prediction Across India

Process-based land surface models (LSMs) are essential tools for global water cycle and runoff assessments. However, when coupled with hydrodynamic models, their streamflow simulations often exhibit considerable uncertainties in uncalibrated settings, making them less effective for local hydrology applications. Calibrating LSMs using observed streamflow data across large basins and regions is computationally expensive and can even impair the performance of other variables. On the other hand, deep learning models, particularly Long-Short Term Memory (LSTM) networks, have shown promise in streamflow simulations, but often struggle to reproduce other water cycle variables reliably. In this study, we propose a hybrid modeling framework that combines process-based models with deep learning to enhance daily streamflow simulations without requiring basin-specific calibration. Applied at a national scale, the framework integrates a multi-model hydrologic ensemble from the Indian Land Data Assimilation System (ILDAS) with a novel two-stage post-processor. This post-processor uses a residual error prediction LSTM alongside an auto-regressive meta-learning LSTM. Trained on 220 catchments across India, the framework significantly improves streamflow predictions, raising the Kling-Gupta Efficiency in 208 catchments, with the national median improving from 0.18 (uncalibrated) to 0.60. Additionally, peak flow timing error and peak mean absolute percentage error were reduced by 25% in 135 catchments. This approach demonstrates the potential to integrate LSMs with deep learning to provide more accurate and locally relevant streamflow predictions.