Reconstructing floods from large-scale atmospheric variables with neural networks in high latitude climates

<p>Climate change is expected to alter the occurrence of floods in high latitude countries; evidence of earlier spring floods and more frequent rainfall-driven floods has already been detected in Norway. While the state-of-the-art hydrological climate-impact model chain embeds explicit assumptions about stationarity, machine learning offers a complementary approach to hydrological climate-impact modelling by facilitating direct downscaling from large-scale atmospheric variables to streamflow, thus making downscaling and bias-correction implicit. While applications of machine learning algorithms for streamflow and flood modelling are well documented in the scientific literature, few studies have linked large-scale atmospheric variables directly to streamflow without including observed streamflow as part of the input variable selection. Such autoregressive models have limited application for climate-impact studies, as future streamflow is yet to observe. Furthermore, most studies linking large-scale atmospheric forcing to catchment response have focused on monthly, seasonal, or annual streamflow. This study presents the application of feed-forward and recurrent neural networks for daily streamflow and flood reconstruction from atmospheric reanalysis data with comparable spatiotemporal resolution to global climate model outputs. Two widely applied neural network types, namely multilayer perceptron (MLP) and long short-term memory (LSTM), were benchmarked against gradient boost regression tree models. Catchment-specific, physically-based input variable selections representing the dominant flood-drivers were identified for 27 catchments in Norway. The selected catchments have low degrees of basin development and anthropogenic influence so that the established statistical links only reflect the forcing-response relationship between the atmosphere and the catchments. Overall, the LSTM obtained the highest accuracy, with a median Nash Sutcliffe Efficiency (NSE) of 0.88 on the training set (1950-2000) and 0.76 on the testing set (2006-2010). However, the MLP proved more robust, with a smaller drop in NSE from training (0.76) to testing (0.72), indicating that further restricting the input variables based on hydrological theory and physical interpretability may increase the robustness of neural networks in the context of daily streamflow modelling. The median NSE of the regression tree models was lower on both the training set (0.73) and the testing set (0.66). The results point to the potential of neural networks for hydrological climate-impact modelling in catchments where both snowmelt and rainfall constitute flood-drivers in the present climate. This research provides a springboard for future studies employing neural networks for hydrological climate-impact modelling in high latitude countries. Future research should assess the potential for regionalization by including catchment characteristics through clustering techniques like Kohonen Self-Organizing Maps.</p>