Locally Relevant Streamflow by Integrating a Land Surface Model Ensemble With a Two‐Stage LSTM Post‐Processor

Abstract Process‐based land surface models (LSMs) are widely used for global water cycle and runoff assessments, but when integrated with hydrodynamic models, the streamflow simulations exhibit significant uncertainties in uncalibrated mode, limiting their effectiveness in local hydrology applications. The calibration of LSMs against observed streamflow across large basins and regions is computationally prohibitive and sometimes degrades performance of other variables. In contrast, deep learning models, particularly Long‐Short Term Memory (LSTM) networks, have shown promising results in streamflow simulations, but they are often limited by poor reproducibility of other water cycle variables. This study presents a hybrid modeling framework that integrates process‐based models with deep learning to improve daily streamflow simulations without requiring basin‐specific calibration. The framework is showcased on a national scale using a multi‐model hydrologic ensemble from the Indian Land Data Assimilation System (ILDAS). It is integrated with a proposed two‐stage post‐processor, which pairs a residual error prediction LSTM with an auto‐regressive meta‐learning LSTM to predict 1‐day ahead streamflow. Trained on multi‐decadal data from 220 catchments across India, the framework improves Kling‐Gupta Efficiency in 208 catchments, raising the national median from 0.18 (uncalibrated) to 0.60. It also reduced peak flow timing error and peak mean absolute percentage error by 25% in 135 catchments. During monsoon and post‐monsoon periods, residual error interquartile range (IQR) decreased by 66.3% and 81.7%, respectively. This approach has the potential to integrate LSMs with deep learning for more accurate and locally relevant streamflow predictions, while enhancing other water cycle variables through methods like data assimilation.