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.

Observation‐Constrained Projection of Global Flood Magnitudes With Anthropogenic Warming

“…Krinner and Flanner (2018) found that the bias patterns of key climate variables did not change substantially under strong climate change, which enables us to apply the fine-tuning physical-guide machine learning model to future periods with higher confidence. The LSTM-based post-processing bridges the gap between process-based and data driven models and can provide us with an operational way to constrain the GCMs-GHMs-simulated historical and future streamflow through observations (e.g., Sungmin et al, 2020;Yang et al, 2019). It can thus improve the reliability of GCMs-GHMs-based flood projections in the case the physical models cannot adequately represent/simulate the real-world processes of floods.…”

“…Moreover, GHMs usually performed worse in simulating hydrological extremes than normal values in some river basins. Possible reasons include the inadequate mathematical representation of hydrological systems (despite including some human impact parameterizations), limited availability and coarse spatial‐temporal resolution of global forcing data, and insufficient calibrations conducted in a few basins (Bierkens, 2015; W. B. Liu, Lim et al., 2018; Mateo et al., 2017; Veldkamp et al., 2018; Yang et al., 2019; Zaherpour et al., 2018).…”

Design flood estimation for global river networks based on machine learning models

“…After quality control of the daily streamflow data, 15 types of time-series indices are provided at yearly, seasonal, and monthly resolutions in the GSIM archive. To date, this archive has been successfully used in flood classification (Stein et al, 2019), streamflow trend analysis (Do et al, 2019) and hydrological model evaluation (Yang et al, 2019a) at the global scale. To make reliable estimates for at-site DFs, a station selection is needed based on some quality control criteria.…”

“…Return period flows along the river network in the cascade model type are derived by an at-site flood frequency analysis of the resulting land surface model streamflow. However, due to the coarse resolution (usually 0.5 degrees) of global land surface models (Yang et al, 2019b;Liu et al, 2019), some downscaling and bias correction methods usually need to be adopted for high-resolution flood hazard mapping (Mueller Schmied et al, 2016;Frieler et al, 2017;Schumann et al, 2014a). Unlike the cascade model type, the gauged flow data model type uses observed gauged discharge and regional flood frequency analysis (RFFA) approaches to produce different return period flows along the global river network (Trigg et al, 2016).…”

PrecipGAN: Merging Microwave and Infrared Data for Satellite Precipitation Estimation Using Generative Adversarial Network

“…DL has shown great capability in approximating nonlinear deterministic functions (LeCun et al., 2015; Schmidhuber, 2015), and is successfully applied in fields including computer vision, speech recognition, natural language processing, robotics, self‐driving cars and medical image processing (Amodei et al., 2016; He et al., 2016; Ramos et al., 2017; Ronneberger et al., 2015; Sünderhauf et al., 2018; Young et al., 2018). There are also many applications of DL in geoscience fields, such as weather prediction (Pan et al., 2019; Shi et al., 2015), soil moisture (Fang et al., 2017), hydrological model (Tan et al., 2018; Yang et al., 2019; Zhao et al., 2019), climate modeling (Gentine et al., 2018; Rasp et al., 2018) and land cover classification (Fang et al., 2019; Zhang et al., 2019; Zhong et al., 2017).…”