Landslide Susceptibility Prediction Model using Random Forest for Kalutara District, Sri Lanka

Abstract

Abstract Landslides are one of the most recurrent and prominent natural disasters in Sri Lanka. An area of nearly 20,000 sq. km encompassing 10 districts is prone to landslides. According to statistics provided by the National Building Research Organization landslides have destroyed over 800 lives in Sri Lanka over the last decade. In 2017 Kalutara district reported the maximum number of deaths of 101 due to landslides. Owing to haphazard, unplanned land use, inappropriate construction methods, wanton human intervention and other other geological and morphological causes, the trend of landslide occurrence will continue in the next eras. Therefore prediction of landslide susceptibility is indispensable for disaster management and ensure sustainability of developments. The main focus of this study was to investigate the applicability of 12 landslide conditioning factors including slope, aspect, hydrology, Stream Power Index(SPI), Topographic Wetness Index(TWI), Sediment Transport Index(STI), geology, land form, land use, soil type, soil thickness and rainfall in the prediction of landslide susceptibility in Kalutara district using Random Forest machine learning algorithm. In order to achieve this a Geographical Information System(GIS) was used to manipulate and analyze the spatial data while the implementation of the prediction model was carried out using python. A pilot study was carried out to analyze the correlation between the landslide conditioning factors and landslide occurrence and to select the most appropriate set of conditioning factors for the prediction. A landslide inventory of 84 landslides occurrences in Kalutara district, was utilized along with randomly generated 84 non-landslide locations from the landslide-free area of Kalutara district. Random Forest (RF), a non-parametric supervised classification algorithm was employed to construct the prediction model. The efficiency of the Random Forest model was evaluated using Receiver Operating Characteristic(ROC), accuracy, sensitivity and specificity. The results indicated 76.92% specificity, 84.00% specificity, and accuracy of 80.39%. The area under the ROC curve demonstrated 79.46% of predictive capability for the model. Keywords: Landslide Susceptibility, Machine Learning, GIS, Random Forest