Improving generalization performance of landslide susceptibility model considering spatial heterogeneity by using the geomorphic label-based LightGBM

Previous landslide susceptibility assessments often overlook the heterogeneity of landslides across different regions and their generalization ability. The objective of this study was to explore the landslide susceptibility model from geomorphic zones-based K-means constructed sub-regions to geomorphic label-based matching whole regions. To achieve this, we divided the study area into slope units based on curvature calculation and utilized a Light Gradient Boosting Machine (LightGBM) to construct the susceptibility model. For model training, we randomly selected fourteen townships in Fengjie County, Chongqing, China. By employing K-means clustering, we established geomorphic zones for the landslide susceptibility model, which were assigned a corresponding geomorphic label. Notably, the model divided the entire study area into four geomorphic zones, with zone 4 demonstrating the highest accuracy of 0.85. Additionally, we developed two comparative models to assess the effectiveness of our proposed approach. The first comparative model considered only ten conditioning factors, resulting in a model accuracy of 0.73. The second comparative model incorporated the classification results of geomorphic labels as explanatory variables for landslides and achieved a model accuracy of 0.75. In this study, we successfully established a model with an accuracy of 0.87 by automatically matching and generalizing the remaining townships with the existing geomorphic labels. Our findings suggest that a geomorphic label-based model, adopting a framework from sub-regions to the whole region, offers better quantification of the landslide occurrence mechanism in the region and demonstrates stronger generalization ability compared to the overall-based model.