A new random forest method for landslide susceptibility mapping using hyperparameter optimization and grid search techniques

Landslides are common and frequent in the Indian Himalayan belt. Environmental factors such as precipitation, terrain, geology, seismic activity, stream erosion, and anthropogenic activities significantly trigger landslides. In this paper, three supervised machine learning methods, namely random forest (RF), support vector machine, and logistic regression, were tested for generating landslide susceptibility maps of Himachal Pradesh, India, and RF was proven to be the best performer among the tested methods. This study presents a novel approach to landslide susceptibility mapping by enhancing the RF method with hyperparameter optimization and grid search techniques. This methodology is distinctive as it optimizes RF for improvement. A new landslide inventory for 2023 was developed and digitized using Google Earth, then cross-validated with historical landslide occurrences recorded in the government inventory from 2016 to 2020. The optimized RF model demonstrated strong performance, achieving an accuracy, precision, recall, and F1-measure of 0.92 across all metrics, which is 5.75-55.93% better than the other tested methods. The landslide susceptibility map (LSM) generated from this research provides an insight to the existing government maps by offering a higher resolution and more detailed analysis of landslide-prone areas. Our LSM can contribute to land-use planning by clearly identifying varying risk levels and effectively bridging the gaps found in government maps. This comprehensive map provides valuable information for thorough environmental impact assessments and risk mitigation in response to landslide hazards.