Observation‐Constrained Projection of Global Flood Magnitudes With Anthropogenic Warming

Abstract

River flooding is among the costliest natural disasters with severe economic, societal, and environmental consequences. However, substantial uncertainties remain in global and regional projections of future flood conditions simulated by global climate models (GCMs) and/or global hydrological models (GHMs). Using physical models coupled with machine learning (ML), for the first time, we project changes in flood magnitudes of 2062 global river basins by constraining physical‐based streamflow simulations with observations under 1.5°C and 2°C warming scenarios identified for the Representative Concentration Pathway 8.5. We found that, during the validation period, the GHMs‐simulated flood magnitudes would improve with reduced uncertainty over the selected river basins after ML with a Long Short‐Term Memory network. Our estimation suggested that flood magnitudes would increase in many Northern Hemisphere mid‐ and high‐latitude rivers (e.g., Lena River, Amur River and Volga River) but decrease in some river basins in southern Finland and Eastern Europe in future periods (i.e., 1.5°C and 2°C warming levels). In 1.5°C and 2°C warmer worlds, the decreasing flood magnitudes in most South American rivers are associated with decreased soil moisture and increased evapotranspiration induced by warmer temperatures. Although the geographical pattern of changes in flood magnitudes for the +2°C experiment is close to that of the +1.5°C experiment, a 1.5°C warming target is more likely to reduce flood magnitudes of many river basins worldwide (e.g., in central and eastern Siberia, Alaska/Northwest Canada and South America).

Hybrid forecasting: using statistics and machine learning to integrate predictions from dynamical models

“…LSTMs can also be used to post-process outputs from physics-based models, such as long-term streamflow projections (Liu et al, 2021) and streamflow simulations (Frame et al, 2021) to make them more realistic. Liu et al (2021) implemented a physics-informed approach to post-process the streamflow projections from GCMs, GHMs and the Catchment-based Macroscale Floodplain model (CaMa-Flood).…”

“…LSTMs can also be used to post-process outputs from physics-based models, such as long-term streamflow projections (Liu et al, 2021) and streamflow simulations (Frame et al, 2021) to make them more realistic. Liu et al (2021) implemented a physics-informed approach to post-process the streamflow projections from GCMs, GHMs and the Catchment-based Macroscale Floodplain model (CaMa-Flood). The LSTMs were trained to learn a relationship between simulated streamflow (from the physics-based model GHMs-CaMa-Flood), basin averaged daily precipitation, temperature, windspeed and observed streamflow.…”

Future Changes in High and Low Flows under the Impacts of Climate and Land Use Changes in the Jiulong River Basin of Southeast China

“…Numerous studies have detected possible future changes in runoff under the impacts of climate and land use changes, and then analyzed with great detail the potential changes in different hydrological regimes including high and low flows. Most of them adopted hydrological models with projected future climate forcing data from general circulation models (GCMs) under the Coupled Model Intercomparison Project (CMIP) [12][13][14][15][16]. Based on a global river routing model with outputs of 11 GCMs from CMIP5, Hirabayashi et al [14] projected future changes in runoff on a global scale and concluded that a warmer climate would significantly increase future high flows in Southeast Asia, peninsular India, eastern Africa and the northern half of the Andes.…”

Machine Learning Reveals a Significant Shift in Water Regime Types Due to Projected Climate Change

“…The presented modeling chain that connects GCM outputs with hydrological models has been established in numerous research projects to estimate the impact of projected climate change on river runoff [18][19][20]. Depending on the underlying hydrological model used, runoff estimates could be represented at different spatial (global, regional, basinscale) and temporal (daily, monthly, annual) resolutions.…”

“…Depending on the underlying hydrological model used, runoff estimates could be represented at different spatial (global, regional, basinscale) and temporal (daily, monthly, annual) resolutions. Then these estimates serve as a basis of impact assessment-e.g., depicting trends [21] or analyzing differences between runoff characteristics such as flood magnitude and timing on historical and projected periods [13,[16][17][18][19][20]22]. Thus, individual runoff characteristics are in the strong focus of studies devoted to climate change impact assessment.…”