Spatio-statistical comparative approaches for landslide susceptibility modeling: case of Mae Phun, Uttaradit Province, Thailand

Abstract

In this study for landslide susceptibility modeling, three quantitative techniques, i.e., frequency ratio (FR), information value (IFV), and weight of evidence (WOE), were evaluated and compared. For this purpose, landslide inventory map was prepared using visual interpretation on SPOT-5 image and field survey was carried out for ground truthing of landslide sites and total 677 landslides were identified. The inventory map was divided into training and validation datasets, and from total, 473 landslides (70%) were for training to run the model and 30% (204 landslides) for validation purpose. Total 11 landslide conditioning factors were used in this study that are: elevation, slope, aspect, curvature, plan curvature, profile curvature, land use/land cover (LULC), topographic wetness index (TWI), stream power index (SPI), proximity to road, and proximity to stream. Three different landslide susceptibility maps were produced based on analyzing the relationship of landslides with conditioning factors using FR, IFV, and WOE in GIS environment. The results of FR model indicated that almost 40% of the total study area fall in high to very high landslide susceptibility zones, while in WOE and IFV models, it was found almost 50% of the total area. The landslide susceptibility maps were validated using prediction and success rate curve techniques. The prediction rate curve gives us a glimpse of future landslides based on present landslide susceptibility maps. The results obtained from validation showed that the area under curve (AUC) based on prediction rate curve for FR, IFV, and WOE was 80.78%, 72.88%, and 72.33%, respectively. However, the AUC obtained through success rate curve for the models in this study was 74.60%, 75.04%, and 72.54% for FR, IFV, and WOE, respectively. Moreover, the evaluation of landslide density test and seed cell index area (SCAI) indicated that calculated and classified landslide susceptibility maps are in a good agreement with the field conditions. Thus, it was observed from this study that the frequency ratio has better accuracy as compared to information value and weight of evidence, but in success rate curve, almost all the models showed the same results. Consequently, it can be concluded that the susceptibility maps produced from FR, IFV, and WOE are in good agreement because more than two-third of landslides falls in high and very high susceptibility zones of each model. From this study, it was found that slope angle, elevation, land use/land cover, and roads play a major influencing role in the occurrence of landslide in the study area. The maps produced based on these landslide susceptibility models provide a base for engineers and land use planner to develop landslide management strategies before any development on slopes.