Landslide susceptibility prediction based on GIS and Artificial Neural Network

Abstract

Landslide is one of the hazardous events which cause lives and property damages. Many factors, such as soil type, slope of the terrain, precipitation and manmade activities incorporate for landslides. Frequent of occurring landslide is increasing in Sri Lanka due to the erratic rainfall pattern with climate change especially in the monsoon seasons. Therefore, to prevent loss of lives and damage to property, proper observation and analysis of unstable slope behaviour is crucial. This research inspects the success rate and the effectiveness of using the Artificial Neural Network (ANN) which is a widely used machine learning technique for pattern recognition and prediction purposes and using modern GIS tools and technologies. Badulla district was selected as the study area of the study which is a protruding geographical area of the country when considering landslide susceptibility due to uneven sloppy shapes of the terrain. Seven landslide causing factors are selected after performing statistical and GIS analysis. The factors consist precipitation, surface slope, aspect, soil density, land use and the distance to water bodies. The extracted data of these factors are then used for the ANN network training process. The factors were identified by interpreting satellite images, topographical maps and factor layer maps obtained from National Building Research Organisation and Survey Department. ArcMap for desktop and several other GIS tools were used for GIS analysis and factor data extraction. A comprehensive geospatial database was created using the above collected factor data. This geospatial database contains data related to the terrain, surface as the varying factors. Precipitation was identified as the most prominent and triggering factor for landslides. The above geospatial database was then used for landslide susceptibility prediction using Artificial Neural Network. MATLAB scripting language was used for ANN network creation and for manage training process. After training and several neuron weight adjustments, successful landslide prediction results could be obtained by the network. These results were further processed and then used for the evaluation processes. The Evaluation was done comparing several previously occurred known landslide events and the prediction results returned by the above trained ANN network. Several geoprocessing tools of ArcMap were used for the comparing process. Further, the performance measures of ANN was determined by calculating accuracy, precision, recall and confusion matrix. The network returned 71.3% prediction accuracy according to the performance calculations.