Improving trans-regional hydrological modelling by combining LSTM with big hydrological data

Abstract

Study region

Lancang-Mekong River Basin (LMRB), Brazil.

Study focus

Streamflow prediction in ungauged basins is a significant challenge in hydrology. This study investigates the transferability of deep learning models for hydrological simulations in ungauged basins, focusing on how constraints like catchment attributes, meteorological forcing, and Global Hydrological Models (GHMs) improve model performance when transferring knowledge from gauged to ungauged basins. We applied the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS-BR) dataset alongside GHMs and deep learning techniques to simulate hydrological processes in the LMRB.

New hydrological insights for the region

The results demonstrate that a post-processing scheme combining deep learning, meteorological data, and GHMs significantly improves model accuracy, achieving a median Nash-Sutcliffe Efficiency (NSE) of 0.64, compared to 0.50 for the baseline Long Short-Term Memory (LSTM) model without GHMs. Key factors influencing model performance include catchment attributes, climate variations, and the length of the modelling series. A notable finding is the importance of catchment attributes in defining hydrological similarity, which enhances model migration between regions with differing data availability. Cross-regional migration was particularly successful when hydrological similarities between the Amazon Basin and LMRB were evaluated, achieving an NSE of 0.86 at the Pakse hydrological station. These insights provide a novel modelling framework for hydrological simulations in data-scarce regions, emphasizing the role of physical mechanisms and hydrological similarities in improving model transferability.