Landslide Susceptibility Mapping in Chittagong District of Bangladesh using Support Vector Machine integrated with GIS

The key objective of this research was to identify landslide susceptible areas in Chittagong District of Bangladesh using Support Vector Machine and Geographic Information System. A landslide inventory database consisting of 105 historical landslide locations was used to make partition randomly into two parts such as the training dataset containing 70% (73 landslide locations) and checking dataset 30% (32 landslide locations) for the SVM models. In this study ten major landslide conditioning parameters such as topography, slope-angle, slope-curvature, slope-aspect, distance to-stream, road and fault line, precipitation, Normalize Difference Vegetation Index (NDVI) and soil type data were used as the triggering factors for the landslide susceptibility mapping in the corresponding area. To compare the prediction performance of the landslide models six different types of kernel functions namely linear, cubic, quadratic, fine-Gaussian, coarse-Gaussian and medium-Gaussian were applied. The obtained results showed that the cubic kernel demonstrated the highest average training accuracy (98.7%) among the six different kernel functions. In the next step, an SVM model was constructed based on the model with maximum accuracy. The experimental results of the SVM models were also validated using Receiver Operating Characteristic (ROC) curve method. The validation results also showed that the Area Under the Curve (AUC) for cubic kernel was highest and it was98.3% which demonstrated very excellent prediction performance for regional scale landslide mapping. Finally a landslide susceptible map of Chittagong district was produced to visualize the landslide prone areas.