Northwestern University and University of California, Los Angeles (UCLA) scientists have developed a new process-based framework that provides a more accurate and dynamic approach to landslide prediction over large areas.
While traditional landslide prediction methods often rely solely on rainfall intensity, the new approach integrates various water-related processes with a machine-learning model. By accounting for diverse and sometimes compounding factors, the framework offers a more robust understanding of what drives these destructive events.
With further development, the new framework could help improve early warning systems, inform hazard planning and enhance strategies for climate resilience in regions vulnerable to landslides. Ultimately, these approaches could help save lives and prevent damage.
The study was published today (July 25) in Geophysical Research Letters.
"Current early warning systems tend to derive information from historical precipitation events and landslides," said Chuxuan Li, the study's first author. "Because it's based on historical data, it doesn't consider the changing climate. In the future, we expect more intense precipitation and higher numbers of heavy precipitation events. These systems often don't consider snow melt or other ground conditions. Our model considers a wider range of factors, so we can identify more diverse pathways leading to landslides over a large spatial scale."
"Different landslides can be caused by different hydrological processes," said Northwestern's Daniel E. Horton, the study's senior author. "We're trying to identify which landslides are caused by which processes. But we're also trying to think about it from a much broader scale; a scale that is consistent with the storms that cause these events. Our ideal is to develop tools that could be useful across a broad region, such as the state of California."
Horton is an associate professor of Earth, environmental and planetary sciences at Northwestern's Weinberg College of Arts and Sciences, where he leads the Climate Change Research Group. Li is a Ph.D. graduate from Horton's laboratory at Northwestern and current postdoctoral researcher at UCLA.
Simulating a 'parade' of storms
Dangerous flows of water, mud and rocks, landslides can be difficult to predict - especially across large areas with varied landscapes and different climates. To better understand how and why widespread landslides occur, the Northwestern and UCLA team looked to one month of extreme weather in California.
During the winter of 2022-23, California experienced an unprecedented "parade" of nine consecutive atmospheric rivers, which caused catastrophic flooding and more than 600 landslides. To understand the pathways that caused these landslides, the scientists adopted a community-developed computer model that simulates how water moves through the environment, including rain infiltrating into the ground, running off on the surface, evaporating, and freezing or melting of snow and ice.
To drive the model, the team used a diverse array of meteorological, geographical and historical data. This included information about terrain, soil depth, past wildfires, precipitation, and meteorological and climatic conditions.
Using model outputs, the team developed a metric, called "water balance status" (WBS), to assess when there is too much water in a particular area. A positive WBS means there's more water than the ground can handle through absorption, storage, evaporation or drainage. This also means there's higher potential for landslides.
Identifying main pathways
Finally, the Northwestern and UCLA team applied a machine-learning technique to group together similar landslides based on their sites' specific conditions. Through this technique, they identified three main pathways that led to the California landslides: intense rainfall, rain on already saturated soils and melting snow or ice.
The team predicts that heavy, rapid downpours caused about 32% of the landslides. Roughly 53% of the landslides occurred after moderate rain fell on soils already saturated from previous storms. And about 15% of the landslides were linked to snow or ice, with rain accelerating the snowmelt or ice thaw.
"We found most of the landslides were triggered under excessively wet conditions," Li said. "By excessively wet, we mean the precipitation exceeds the soil's capacity to hold or drain water. This can be especially dangerous on steep slopes."
When the scientists compared these events to their model, they found a significant majority (89%) of California's landslides occurred in areas where the WBS was positive. This finding validated that the metric can accurately identify conditions ripe for landslides.
"While this study looks backwards to understand a past event, our ultimate goal is for the method to look forward to make predictions," Horton said. "We plan to take this modeling framework that we developed and use it in concert with weather forecasting models."
Better models for an uncertain future
As the global climate continues to change, prediction systems are more important than ever. Because warmer air can hold more water vapor, storms can dump more water. And more water often indicates more dangerous flooding and landslides.
In a recent review published in the journal Science, Horton and his collaborators examined how natural hazards, such as atmospheric rivers, often trigger other disasters to create a chain reaction. In the piece, the authors emphasize the critical need for integrating diverse datasets and building advanced models to improve the ability to predict and prepare for natural disasters.
"Atmospheric rivers are not necessarily becoming more common," Horton said. "But, when they do make landfall, their impact is becoming more severe. Lately, we have seen an increase in the intensity of their precipitation. This is consistent with the global trend of experiencing more intense precipitation events due to human-caused climate change."
The study, "Mixed hydrometeorological processes explain regional landslide potential," was supported by the National Science Foundation (PREEVENTS grant numbers 1854951 and 2023112).
