Evaluation and machine learning improvement of global hydrological model-based flood simulations

Abstract A warmer climate is expected to accelerate global hydrological cycle, causing more intense precipitation and floods. Despite recent progress in global flood risk assessment, the accuracy and improvement of global hydrological models (GHMs)-based flood simulation is insufficient for most applications. Here we compared flood simulations from five GHMs under the Inter-Sectoral Impact Model Intercomparison Project 2a (ISIMIP2a) protocol, against those calculated from 1032 gauging stations in the Global Streamflow Indices and Metadata Archive for the historical period 1971–2010. A machine learning approach, namely the long short-term memory units (LSTM) was adopted to improve the GHMs-based flood simulations within a hybrid physics- machine learning approach (using basin-averaged daily mean air temperature, precipitation, wind speed and the simulated daily discharge from GHMs-CaMa-Flood model chain as the inputs of LSTM, and observed daily discharge as the output value). We found that the GHMs perform reasonably well in terms of amplitude of peak discharge but are relatively poor in terms of their timing. The performance indicated great discrepancy under different climate zones. The large difference in performance between GHMs and observations reflected that those simulations require improvements. The LSTM used in combination with those GHMs was then shown to drastically improve the performance of global flood simulations (especially in terms of amplitude of peak discharge), suggesting that the combination of classical flood simulation and machine learning techniques might be a way forward for more robust and confident flood risk assessment.