Predictive Analysis of Landslide Activity Using Remote Sensing Data

Landslides are historically one of the most damaging geohazard phenomena in terms of death tolls and socio-economic losses. Therefore, understanding the underlying causes of landslides and how environmental phenomena affect their frequency and severity is of critical importance. Of specific importance for mitigating future damage is increasing our understanding of how climate change will affect landslide severity, occurrence rates, and damage. We are developing data driven models aimed at predicting landslide activity. The models learn multi-dimensional weather and geophysical patterns associated with historical landslides and estimate location-dependent probabilities for landslides under current or future weather and geophysical conditions. Our approach uses machine learning algorithms capable of determining non-linear associations between dependent variables and landslide occurrence without requiring detailed knowledge of geomorphology. Our primary goal in year one of the project is to evaluate the predictive capabilities of data mining models in application to landslide activity, and to analyze if the approach will discover previously unknown variables and/or relationships important to landslide occurrence, frequency or severity. The models include remote sensing and ground-based data, including weather, landcover, slope, elevation and drainage information as well as urbanization data. The historical landslide dataset we used to build our preliminary models was compiled from City of Seattle landslide files, United States Geological Survey reports, newspaper articles, and a verified subset of the Seattle Landslide Database that consists of all reported landslides within Seattle, WA, between 1948 and 1999. Most of the landslides analyzed to-date are shallow. Using statistical analysis and unsupervised clustering methods we have thus far identified subsets of weather conditions that lead to a significantly higher landslide probability, and have developed statistically predictive models for individual storms. We have also developed a topographic probabilistic map indicating current hotspots - areas where the probability of a landslide is heightened - and tested it on an unseen set of data.