Advanced graph-based and deep learning frameworks for landslide susceptibility mapping in mountain transportation...

Area of Science:

• Geosciences and Artificial Intelligence
• Geotechnical Engineering
• Machine Learning for Geospatial Analysis

Background:

• Landslide susceptibility mapping in mountainous transportation corridors is crucial for infrastructure safety.
• Existing methods struggle to capture complex nonlinear relationships and spatial dependencies among geo-environmental factors.
• There is a need for advanced modeling approaches that explicitly account for topological dependencies in terrain units.

Purpose of the Study:

• To propose and validate a hybrid graph-based artificial intelligence framework for landslide susceptibility assessment.
• To model terrain units as nodes in a spatially structured graph, capturing topological dependencies.
• To compare the proposed framework against state-of-the-art deep learning and hybrid machine learning models.

Main Methods:

• Development of a hybrid GraphSAGE-CatBoost framework optimized by the Reptile Search Algorithm (RSA).
• Compilation of a comprehensive database including 409 landslides, 409 non-landslides, and ten conditioning factors.
• Benchmarking against GoogleNet-CNN (Harris Hawks Optimization) and Autoencoder-XGBoost using random split and spatial cross-validation.

Main Results:

• The GraphSAGE-CatBoost-RSA model achieved superior performance (AUC-ROC=0.972, Accuracy=0.932) over benchmarks.
• The model demonstrated strong generalization capabilities with spatial cross-validation (mean AUC-ROC=0.924).
• High-susceptibility zones are linked to steep slopes, weak lithologies, proximity to roads/rivers, and north-facing aspects.

Conclusions:

• The proposed graph-based AI framework offers a robust and spatially explicit approach for landslide susceptibility mapping.
• This method effectively captures complex spatial dependencies, outperforming conventional and advanced alternatives.
• The resulting susceptibility map aids in identifying critical areas for targeted mitigation along transportation corridors.