A Hybrid Ensemble Framework for Landslide Susceptibility Mapping: Integrating Spatial Clustering and Multi-Scale Terrain Features

This study presented a hybrid machine learning framework for high-resolution landslide susceptibility mapping in the Svaneti region of the Georgian Greater Caucasus, a steep alpine environment prone to frequent mass movements. The objective was to enhance predictive accuracy and spatial generalization of susceptibility maps by integrating geomorphological knowledge with data-driven modeling.

We combined geostatistical, geomorphometric, and spectral terrain features, including slope derivatives and multi-scale gray-level co-occurrence matrix texture extracted from diverse raster sources. These were systematically preprocessed and complemented by spatial clustering, which segmented the landscape into zones with internally coherent terrain characteristics using tree-based projections, dimensionality reduction, and spatially-aware k-means.

Supervised classification leveraged a binary landslide inventory aligned to the feature grid. Class imbalance was addressed through stratified undersampling. Base classifiers, including extreme gradient boosting, random forests, k-nearest neighbors, and a multi-layer perceptron neural network, were tuned using Bayesian optimization and combined into a stacked ensemble with extreme gradient boosting as the meta-learner. The final model yielded high cross-validated performance (AUC = 0.98, balanced accuracy = 0.95).

The resulting susceptibility map showed good spatial agreement with known events and provided a detailed zonation of hazard potential. Despite a high recall (0.98) for landslide prediction, the low precision (0.13) and moderate Cohen’s Kappa (0.21) reflected challenges posed by extreme class imbalance and spatial noise in event inventories. Overall, the proposed framework demonstrated strong potential for integrating spatially-informed clustering with ensemble learning to enhance landslide susceptibility mapping in complex mountainous regions and supported scalable, region-specific hazard assessment.