Reconstructing daily streamflow and floods from large-scale atmospheric variables with feed-forward and recurrent neural networks in high latitude climates

Abstract

ABSTRACT With climate change, decreases in winter snow storage and increases in precipitation duration and intensities will alter the occurrence of floods in high-latitude countries. The state-of-the-art hydrological climate-impact model chain consists of one or more global climate models, downscaling and bias-correction techniques, and one or more hydrological models. Machine learning offers a complementary approach to hydrological climate-impact modelling by facilitating direct downscaling from large-scale atmospheric variables to streamflow. This paper presents the development of multilayer perceptron (MLP) and long short-term memory (LSTM) neural networks benchmarked against regression tree models for reconstruction of daily streamflow and floods from atmospheric reanalysis data with comparable resolution to global climate model outputs. Catchment-specific, physically-based input features representing the dominant flood drivers were identified for 27 catchments in Norway. Overall, the LSTM obtained the highest accuracy. This article provides a springboard for future research on hydrological climate-impact modelling with neural networks in high-latitude countries.