A multi-task deep learning approach for landslide displacement prediction with applications in early warning systems

Area of Science:

• Geosciences and Environmental Science
• Artificial Intelligence in Earth Sciences
• Geotechnical Engineering and Landslide Monitoring

Background:

• Accurate landslide displacement prediction is vital for developing reliable landslide early warning systems (LEWS).
• Deep neural networks are commonly used for landslide displacement modeling, but may not be optimal for LEWS goals.
• Existing models focusing on low prediction residuals can lead to inefficient threshold-based warning predictions.

Purpose of the Study:

• To propose and validate a multi-task deep learning approach for landslide displacement prediction tailored for LEWS.
• To enhance the efficiency of threshold-based warning predictions in landslide monitoring.
• To develop a displacement prediction model for the Urbas landslide as a step towards improving existing alarm systems.

Main Methods:

• A multi-task learning approach was developed to optimize models for LEWS-relevant performance using auxiliary targets.
• A convolutional neural network was employed for day-ahead displacement prediction.
• The model utilized landslide activity, hydrometeorological, and seismological data, validated on the Urbas landslide (Slovenia).

Main Results:

• The proposed multi-task model achieved a competitive R² score for warning prediction.
• The model demonstrated a significantly lower mean absolute error compared to reference models.
• The methodology proved effective in improving landslide displacement prediction for LEWS.

Conclusions:

• The multi-task approach optimizes landslide displacement models for LEWS, improving warning efficiency.
• The developed model shows strong performance for the Urbas landslide, contributing to enhanced monitoring.
• This methodology is broadly applicable for improving landslide modeling and early warning systems globally.