Differences in extremes and uncertainties in future runoff simulations using SWAT and LSTM for SSP scenarios

Abstract

This study compared the performance of Long Short-Term Memory networks (LSTM) and Soil Water Assessment Tool (SWAT) in simulating observed runoff and projecting future runoff using 11 CMIP6 GCMs. The projected runoff was estimated for two Shared Socioeconomic Pathways (SSPs), 2-4.5 and 5-8.5 for near (2021-2060) and far (2061-2100) futures, respectively. The biases in GCM simulated climate variables were corrected using quantile mapping considering observations at 6 weather stations as reference data over the historical period (1985-2014). Five evaluation metrics were used to quantify the GCM's and hydrological models' capability to reconstruct climate variables and runoff in the Yeongsan Basin of South Korea. Uncertainties in LSTM and SWAT simulated runoff for the historical and projected periods were quantified using Bayesian Model Averaging (BMA) and reliability ensemble averaging (REA), respectively. The results showed significant improvement in bias-corrected GCMs in replicating observations in terms of all evaluation metrics. The extreme runoff estimated using general extreme value (GEV) distribution revealed the better capability of LSTM than SWAT in reproducing observed runoff at all gauging locations. The SWAT projected an increase (17.7%) while LSTM projected a decrease (-13.6%) in the future runoff for both SSPs at most locations. The uncertainty in LSTM simulated runoff was lower than in SWAT runoff at all stations for the historical period. However, the uncertainty in SWAT projected runoff was lower than LSTM projected runoff for both SSPs. This study helps assessing the ability of deep-learning versus physically-based models in hydrological modeling and therefore opens new perspectives for hydrological modeling applications.

Keywords: Bayesian model averaging; CMIP6; LSTM; Reliability ensemble averaging; SWAT; Shared socioeconomic pathways; Uncertainty.