A Hybrid Deep Learning Framework to Generate Locally Relevant Streamflow from Large Scale Hydrological Models

The large scale physically based hydrological model outputs can exhibit uncertainties due to various sources such as model structure, parameterization, and inputs. Moreover, these outputs are often available at a coarser resolution due to high computational and data requirements, thus limiting their applicability in operational hydrology. In this study, we present a deep learning framework that reduces the uncertainty in the streamflow generated by a continental scale physical model, using a two-stage process. The hydrological model is setup over the Indian Subcontinent with outputs at 0.1 and daily resolution from 1981-2021. The outputs, including streamflow, are generated by a six-member ensemble consisting of two land surface models (LSM) namely: Noah MP 3.6 and CLSM Fortuna v2.5, and three meteorological forcings namely: MERRA2, CHIRPS and IMD. First, we predict the residual in each of the six-member ensemble streamflow by training a Long Short-Term Memory Network (LSTM) on 220 catchments using dynamic meteorological inputs and static catchment attributes. Next, we pass the six corrected streamflow outputs to another LSTM layer that learns to optimally combine them along with integrating a three-day moving average of observed streamflow, generating an accurate prediction of 1-day ahead streamflow for each catchment. Our results showcase the efficacy of this hybrid framework in significantly enhancing the skill of large-scale hydrological models, imprpoving the national median Kling-Gupta Efficiency (KGE) from 0.01 to 0.628. Moreover, we reproduced extreme event conditions at specific locations with higher accuracy. This study leverages the power of deep learning and extensive observed data to derive locally relevant streamflow predictions from a large-scale hydrological model, offering practical solutions for refining hydrological models and informing water resource management strategies.