Comparison between U-shaped structural deep learning models to detect landslide traces.

Area of Science:

• Geosciences
• Remote Sensing
• Artificial Intelligence

Background:

• Landslides pose significant risks to lives and infrastructure in mountainous regions.
• Real-time landslide monitoring is challenging due to terrain and safety concerns.
• Modern technology offers potential for early landslide detection and warnings.

Purpose of the Study:

• To propose indicators for detecting landslide traces in situ and via remote sensing.
• To develop deep learning (DL) models for automated landslide identification from Sentinel-2 imagery.
• To apply trained DL models for landslide detection in specific regions of Vietnam.

Main Methods:

• Trained nine DL models using U-Net, U2-Net, and U-Net3+ architectures with varying input sizes.
• Utilized multi-temporal Sentinel-2 images as input data for model training.
• Evaluated model performance based on detection accuracy and loss functions.

Main Results:

• The U-Net model (32x32 input) achieved 97% performance with a 0.01 loss function for detecting typical landslide traces.
• The U-Net model (64x64 input) demonstrated capability in detecting larger landslide features.
• A case study in Vietnam successfully utilized the trained U-Net (32x32) model to track landslide occurrences over time.

Conclusions:

• Deep learning models, particularly U-Net, can effectively automate landslide detection from satellite imagery.
• The developed models provide a valuable tool for real-time landslide monitoring and early warning systems.
• This technology can help mitigate risks associated with landslides and associated flash floods in vulnerable areas.