FEATURED NEWS
UCLA team uses artificial intelligence to predict landslides
This article was originally published by UCLA Chemistry & Biochemistry
Professor Louis Bouchard’s group and UCLA collaborators have developed a new approach for predicting landslides that decouples the analytic power of DNNs from their complex adaptive nature in order to deliver more useful results.
A professor of chemistry and bioengineering, Bouchard and his former student Khalid Youssef (Ph.D. ’15, Biomedical Engineering), now a Senior Scientist in Medicine at Indiana University, developed an explainable AI algorithm for geospatial and climatic data.
Seulgi Moon, an associate professor of earth, planetary, and space sciences (EPSS) at UCLA is co-senior author on the team’s paper, along with Bouchard. Youssef is co-first author along with Moon group graduate student Kevin Shao.
By Holly Ober (UCLA Newsroom):
Artificial intelligence can predict landslides
New method improves on older machine learning techniques by revealing which variables were most important in causing the earth to slip
UCLA geologists are using artificial intelligence to predict landslides. The new method improves the accuracy and interpretability of older machine learning techniques and is more broadly applicable than the more common physical and statistical analysis. The new method is described in a paper published in Communications Earth & Environment.
Many factors influence where a landslide will occur, including shapes of the terrain, such as slope, relief, and drainage areas, material properties of the soil and bedrock, and environmental conditions such as climate, rainfall, hydrology, and ground motion due to earthquakes. With so many variables to consider, predicting when and where a chunk of earth is likely to lose its grip and slip down a slope is as much an art as a science.
In places like California, where drought, wildfires and earthquakes create the perfect recipe for landslide disasters, the situation is likely to get worse as climate change brings stronger and wetter storms. Better prediction methods could save lives, property, and money.
Geologists incorporate these controlling factors into physical and statistical models to estimate the susceptibility of an area to landslides. With enough data, physical models can be developed that achieve reasonable predictions. However, they take a lot of time and resources to develop accurately and can’t be applied over broad areas. Statistical models, on the other hand, can reasonably estimate landslide probability but give little insight into their decision-making process.
In recent years many research groups have trained machine learning models such as deep neural networks, or DNNs to predict landslides. Deep neural networks are very large, interconnected networks of algorithms that “learn” from data while achieving excellent accuracy and predictive power. These artificial intelligence programs can process large amounts of information from many variables quickly and accurately but produce results that are hard for humans to interpret. In fact, they provide little to no insight into the results they produce.
“DNNs will deliver a percentage likelihood of a landslide that may be accurate but we are unable to figure out why and which specific variables were most important in causing the landslide,” said co-first author Kevin Shao, a doctoral student in earth, planetary, and space sciences.
This is because in a DNN, the different layers feed into and “learn” from each other to quickly produce an accurate result, a useful quality when an answer is all that’s needed but less useful when there is a need for the computer program to “show its work.”
“The predictions obtained from DNNs do not enable clear separation of the results from different data inputs and the results of this analysis are not useful when the goal is to determine which factors are the most important contributors to natural disasters,” said co-first author Khalid Youssef, a former student of biomedical engineering and postdoctoral researcher at UCLA.
Youssef and Shao, co-corresponding authors Seulgi Moon, an associate professor of earth, planetary, and space sciences and Louis Bouchard, a professor of chemistry and bioengineering, teamed up and developed a new approach for predicting landslides that decouples the analytic power of DNNs from their complex adaptive nature in order to deliver more useful results.
Their method uses a type of artificial intelligence called a superposable neural network, or SNN, in which the different layers of the network run alongside each other, retaining the ability to learn complex relationships between data inputs and output results, but only converging at the very end to yield the prediction. They hypothesized that this approach would produce an accurate prediction and reveal how it arrived at the prediction. The new artificial intelligence program runs effectively with relatively little computing power.
“The SNN is so small it can run on an Apple Watch, as opposed to DNNs, which require powerful computer servers to train,” said Bouchard.
The researchers fed the SNN data from 15 geospatial and climatic variables pertaining to the eastern Himalaya mountains. This particular region was selected because the majority of human losses due to landslides occur in Asia, with a substantial portion of them occurring in the Himalayas.
The SNN model predicted landslide susceptibility for these areas with accuracy rivaling that of DNNs, but using far fewer computer resources. More importantly, the researchers could tease apart the variables to see which ones played a bigger role in producing the results.
“Similar to how autopsies are required to determine the cause of death, identifying the exact trigger for a landslide will always require field measurements and historical records of soil, hydrologic, and climate conditions, such as rainfall amount and intensity, which can be hard to obtain in remote places like the Himalayas,” said Moon. “Nonetheless, our new AI prediction model can identify key variables and quantify their contributions to landslide susceptibility.”
The team looks forward to extending their work to other regions of the world that are landslide prone. In California, for example, the problem is exacerbated by frequent wildfires and earthquakes. The new system may help develop early warning systems that account for a multitude of available signals and predict a range of other surface hazards, including floods.