A novel approach for rockfall susceptibility mapping: Transfer learning between boosting models and logistic regression

Abstract

Rockfalls represent one of the most hazardous geomorphological processes in mountainous environments, often causing severe damage to infrastructure and posing a significant threat to human lives. Unlike other types of mass movements, rockfalls are characterized by abrupt initiation, rapid velocity, and a pronounced dependence on local structural and topographic settings. These distinct dynamics necessitate a tailored approach to susceptibility modeling, with careful consideration of contributing geological and geomorphological factors. In recent years, the integration of advanced tools, such as geographic information systems (GIS), remote sensing, and machine learning algorithms, has greatly improved our ability to produce accurate vulnerability maps. However, the success of machine learning based models depends heavily on the selection of parameters and algorithms and the strategy used to combine predictive results, to avoid overfitting and misleading interpretations. This study proposes a new hybrid modeling framework aimed at improving the accuracy and robustness of rockfall susceptibility mapping. Specifically, three hybrid models were developed by integrating logistic regression (LR) with two powerful ensemble learning techniques: Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM). These different algorithms are combined through standardized weight transfer and weighted averages based on the weight/frequency set. The models were applied to a structurally complex Moroccan limestone ridge mountainous area. Performance evaluation was carried out using a comprehensive set of validation measures, including confusion matrices, area under the ROC curve (AUC), F1 score, recall, precision, and G-mean. The validated results of this study highlight significant methodological advances in mapping susceptibility to rockfalls. Firstly, they confirm the effectiveness of innovative combination techniques, which have subsequently enabled the development of new high-performance hybrid models with a bi-algorithmic structure (LR-XGBoost and LR-LightGBM) and a tri-algorithmic structure (LR-XGB-LGBM), the latter of which stands out for its extreme performance, superior to both basic models and bi-algorithmic hybrid models, by enhancing the robustness and reliability of estimates, which represents a leap forward in the field of mapping the susceptibility of ground movements through the use of this type of algorithm structuring.