Using deep learning to understand flood variability across the last millennium from GCM atmospheric variables in two contrasting catchments

Abstract

Understanding of historical flooding characteristics is conducive for predicting future floods and their characteristics. This study applies deep learning techniques to explore nonlinear long-term relationships between atmospheric variables simulated by the NorESM1-F model and river flow within two selected catchments, the Wujiang basin in Southern China and the Bulken basin in Western Norway. We investigate the feasibility of using atmospheric variables for long-term daily discharge simulations, especially in the context of cold-warm and dry-wet fluctuations over the past 1000 years. Our analysis delves into the changing patterns of atmospheric variables and their impact on discharge and flood patterns. The results indicate that (1) The deep state-space model could effectively simulate daily discharge at the catchment scale by incorporating relevant atmospheric variables of reanalysis data; (2) In our paleoclimate simulations, there is a noteworthy correlation between temperature and precipitation data from the NorESM1-F model over the past millennium with the reconstructed temperature and a proxy indicator for dry-wet conditions in the study basins; (3) Our investigation highlights differences in the simulation of solar1 and solar2, particularly in relation to climate variability associated with the Medieval Warm Period (MWP) and the Little Ice Age (LIA). We observe that, during the periods characterized by larger oscillations in precipitation and temperature, the frequency of floods tends to increase.