A new strategy to map landslides with a generalized convolutional neural network.

Area of Science:

• Geosciences
• Remote Sensing
• Artificial Intelligence

Background:

• Rapid landslide mapping is critical for disaster response and damage assessment.
• Traditional methods involve visual interpretation or machine learning algorithms trained on limited local data.
• Convolutional Neural Networks (CNNs) show promise for landslide detection from remote sensing data.

Purpose of the Study:

• To develop a generalized CNN model for direct application to unexplored areas for event landslide mapping.
• To overcome the limitation of requiring local landslide inventories for training deep learning models.
• To improve the efficiency and speed of post-disaster landslide mapping.

Main Methods:

• Progressive CNN training using combined landslide inventories from multiple event datasets.
• Validation of CNN effectiveness on event landslide inventories from earthquake and extreme weather events.
• Application of trained CNNs to map landslides in new, geographically diverse regions using optical sensor imagery at various resolutions (6m, 10m, 30m).

Main Results:

• CNNs trained on combined inventories demonstrate improved generalization performance in new regions.
• The combined training model achieved a high Matthews correlation coefficient (MCC) of 0.69.
• The method achieves high precision and low recall, with slightly reduced performance but overcomes the need for local training data.

Conclusions:

• Combined CNN training provides a generalized model for rapid, automated landslide mapping in post-disaster scenarios.
• This strategy significantly reduces the dependency on localized training data, facilitating faster response.
• Future research should extend this approach to less vegetated regions.