UNIVERSITY PARK, Pa. — Floods account for up to 40% of weather-related disasters worldwide, and their frequency has more than doubled since 2000, according to a recent report from the United Nations Office for Disaster Risk Reduction. Global flood losses now average $388 billion a year. Simultaneously, droughts are becoming more widespread and damaging across the world.
To help address these challenges, a team of researchers developed a hydrological model that can forecast flooding impacts and manage water resources on a global scale. The approach combines artificial intelligence (AI) with physics-based modeling to provide communities with reliable, detailed data for managing water, reducing flood risk, planning crops and protecting ecosystems.
The model is currently set with a resolution to simulate areas as small as 36 square kilometers, or 14 square miles, worldwide and zoom in to 6 square kilometers, which is 2.5 square miles, in regions with more detailed data. The team’s findings were published in Nature Communications.
“This model is a game changer for global hydrology,” said Chaopeng Shen, Penn State professor of civil and environmental engineering and corresponding author. “Because of its global coverage, finer resolution and high quality, it becomes plausible for a global-scale model to be genuinely useful for local-scale water management and flood forecasting. It can provide strong prior hydrologic knowledge for global satellite missions. It can also provide practical assistance to underdeveloped regions that have lacked these services.”
According to the team, the model has revealed several important insights. First, it suggests that the balance of water between rivers, groundwater and the landscape is not constant and is strongly shifting year to year and place to place due to changes in climate and precipitation. For example, river flows in Europe have declined, leaving less freshwater for estuaries, raising salinity and altering local ecosystems. Second, the rate of a river or stream rising or falling due to rain and the responses in the environment have also shifted dramatically across the world. The new model accurately captured these hydrologic behavioral changes.
A key strength of the model is that it combines neural networks — AI designed to learn in a way similar to the human brain — with physics-based components that rely on mathematical equations and physical laws, Shen said. The physics-based part represents key parts of the water cycle, including rainfall, soil infiltration, groundwater recharge, streamflow and evapotranspiration, which is the process of water evaporating from soil and transpiring from plants as water vapor. The neural network then learns the parameters controlling these processes and can adjust in real time for any missing or simplified components.
“This end-to-end approach is much more robust, especially for data-scarce regions where the physics-based part guarantees basic behavior,” Shen said. “Neural networks are great at learning from big data and filling in the gaps within data they’ve already seen, but they aren’t as good at predicting beyond that range. That’s why it’s so important to combine neural networks with process-based models that are grounded in the physics of how the system actually works, especially when we’re looking at global patterns.”
This new machine learning approach also greatly reduces the manual effort once required to fine-tune model parameters for different regions, Shen noted.
“Traditional methods were slow, limited in scope and couldn’t directly learn from real-world data,” Shen said. “Parameter calibration was a story of sweat and tears. With differentiable programming, the coupled neural networks can now automatically generate parameters while getting trained using feedback from observations.”
Shen said that AI allows training on trillions of parameters, which is far beyond what was possible before. This brings unprecedented consistency, speed and accuracy compared to past techniques.
Shen said he envisions the model shaping decisions about water use, irrigation, flood management and ecosystem protection worldwide. Future updates could add water quality, nutrient tracking and 3D groundwater mapping.
In addition to Shen, authors include Yalan Song, assistant research professor in the Department of Civil and Environmental Engineering; Kathryn Lawson and Jiangtao Liu, both research associates in the Department of Civil and Environmental Engineering; Haoyu Ji, a doctoral student in the Department of Civil and Environmental Engineering; and Tadd Bindas and Farshid Rahmani, recent graduates from Penn State’s Water Resources Engineering program. A full list of authors is available in the published paper.
This research was funded by the U.S. National Science Foundation, National Oceanic and Atmospheric Administration, U.S. Department of Energy and National Aeronautics and Space Administration.
At Penn State, researchers are solving real problems that impact the health, safety and quality of life of people across the commonwealth, the nation and around the world.
For decades, federal support for research has fueled innovation that makes our country safer, our industries more competitive and our economy stronger. Recent federal funding cuts threaten this progress.
Learn more about the implications of federal funding cuts to our future at Research or Regress.