FDRL: a data-driven algorithm for forecasting subsidence velocities in Himalayas using conventional and traditional...

Area of Science:

• Geotechnical Engineering
• Data Science
• Environmental Science

Background:

• Landslides pose significant risks in the Himalayas, impacting lives and infrastructure.
• Existing subsidence velocity models often use isolated data sources, limiting accuracy.
• There is a need for integrated approaches combining diverse data for better landslide prediction.

Purpose of the Study:

• To develop an interpretable data-driven model for enhanced landslide subsidence velocity forecasting.
• To integrate traditional soil information with geotechnical features for improved predictive power.
• To bridge local knowledge systems with modern data science for landslide risk management.

Main Methods:

• Developed a stacking ensemble regression model named Forecasting Data-Driven Regression Learning (FDRL).
• Employed machine learning techniques including feature selection (Pearson correlation, mutual information scores).
• Combined quantitative geotechnical variables with qualitative traditional soil indicators.

Main Results:

• The FDRL model achieved a training RMSE of 1.11 mm/year and a test RMSE of 1.32 mm/year.
• Explainability analysis (SHAP) confirmed significant contributions from both geotechnical and traditional soil features.
• Demonstrated the utility of integrating qualitative, locally-sourced soil data into scientific models.

Conclusions:

• The hybrid approach significantly improves landslide subsidence velocity prediction accuracy.
• The FDRL model offers a scalable, interpretable, and locally implementable solution for early warning systems.
• Integrating traditional soil indicators enhances the practical applicability of landslide risk management strategies.