Explainable AI Integrated Feature Selection for Landslide Susceptibility Mapping using TreeSHAP

Explainability improved model efficiency: By using TreeSHAP to identify which features actually mattered for predictions, researchers reduced the number of factors from 15 to 9 (a 40% reduction) while maintaining superior performance, demonstrating that understanding why a model makes decisions can lead to simpler, more practical systems.

Transparency revealed the most important warning signs: The analysis identified that slope, elevation, and terrain wetness index (TWI) were the strongest predictors of landslide risk, while land use and vegetation indices had minimal impact—information that could help communities and policymakers focus resources on monitoring the most critical factors.

Game-theory based explanations enabled trustworthy predictions: TreeSHAP's statistical foundation provided a principled way to explain each prediction, moving beyond "black box" machine learning to a system where stakeholders can understand and verify how landslide risk assessments are being made.

High accuracy with interpretability addresses real-world needs: The optimized model achieved 94.62% accuracy while remaining explainable, suggesting that responsible AI doesn't require choosing between performance and transparency—both can be achieved simultaneously for critical applications like disaster prediction.

Responsible feature selection reduces bias and costs: By systematically identifying which environmental factors truly drive landslide susceptibility, the approach helps prevent models from relying on spurious correlations or irrelevant data, making predictions more reliable and deployment more cost-effective.