1:45 PM - 3:15 PM
[HTT14-P03] GIS-based machine learning models for assessing landslide impact:
A case study in King County, State of Washington, USA
Keywords:GIS, Landslide susceptibility, Landslide runout distance, Artificial Neural Networks, Random walk
Under the impact of global climate change and population increase, landslides affecting people’s lives have drastically increased in recent years. The impact caused by landslides can be determined by both landslide occurrence probability and the runout distance of landslides. This study aims to evaluate the potential damage caused by landslides in urbanized areas using two different methods: the Artificial Neural Networks (ANNs) method produces a landslide susceptibility map, and the constrained random walk method deals with probable landslide runout distance. These methods are applied to the study area in King County, Washington, USA. The landslide inventory data used included 2331 historical landslides. We divided each landslide area into the source and deposition areas. We also selected 13 conditioning factors for landslide occurrence in the source areas: elevation, slope gradient, slope aspect, plan curvature, profile curvature, lithology, Stream Power Index (SPI), Topographic Wetness Index (TWI), Sediment Transport Index (STI), land cover, distance to roads, distance to railways, and population density. The overall accuracy and the value of the Area Under the Curve (AUC) of landslide susceptibility assessment using these factors are 0.86 and 0.92, respectively, showing the high performance of the applied model. The result of the random walk method calculated the impact probability, which represents the probability for a raster cell to be affected by the movement of a landslide. The overall accuracy and the AUC value of landslide impact probability analysis are 0.77 and 0.87, respectively, proving that the random walk method can predict the impact probability. We also introduced a concept of vulnerability to evaluate the risk in each specific area.