A Comparative Assessment of Landslide Prediction Capability of Machine Learning Methods using Frequency Ratio (FR), Shannon Entropy (SE), and Analytical Hierarchy (AHP) Techniques: A Case Study of Uttarakhand, India

Abstract Mapping landslide susceptibility is crucial for defining high-risk zones and preventing property and human casualties. The Uttarakhand provision, which comes under the Himalayan region, has a high potential for landslide occurrence. A landslide susceptibility map was created using satellite imagery, in-depth field research, and aerial photos. The historical landslide inventory of the state's 14698 total landslides was randomly bifurcated into 70% (10289) for training purposes and 30% (4409) for data validation. Eleven landslide-causative factors (Slope, Aspect, Curvature, Topographic Position Index (TPI), Topographic Wetness Index (TWI), Geology, Normalized Difference Vegetation Index (NDVI), Distance to Road, Distance to Stream, Distance to Fault, and Rainfall) were selected for susceptibility assessment. The landslide susceptibility zonation was created using the Shannon Entropy (SE), Frequency Ratio (FR), and Analytical Hierarchy Process (AHP) techniques, along with the causative factors. The AHP method is effectively utilized in LSM to prioritize and weigh the importance of different causative factors contributing to landslide occurrence, while Shannon Entropy uses the method of discrete probability distribution to quantify the uncertainty or variability associated with different causative factors. The FR, AHP, and SE models were validated using the AUC curve, yielding 92%, 89%, and 81% success rates and predictive rates of 90%, 87%, and 77%, respectively. The FR model is most suitable, more efficient, and valuable for future planning in the study area.