Deep learning-based debris flow hazard detection and recognition system: a case study.

Area of Science:

• Geological Hazard Monitoring
• Computer Vision
• Deep Learning

Background:

• Debris flows pose significant threats in mountainous regions due to their destructive nature.
• Current surveillance camera use for debris flows is primarily for post-event analysis, lacking active monitoring capabilities.
• Anomaly detection in computer vision offers potential for real-time hazard identification.

Purpose of the Study:

• To develop an automatic debris flow detection and recognition system using deep learning.
• To enhance geological hazard monitoring with active, real-time warning capabilities.
• To leverage computer vision for improved debris flow early warning systems.

Main Methods:

• A deep learning system comprising a 3D CNN for feature extraction, an MLP for detection, and another CNN for recognition.
• Utilizing video sequences from surveillance cameras as input data.
• Training and evaluation on the newly annotated Debrisflow23 image dataset.

Main Results:

• The system achieved a detection accuracy of 86.3% AUC and recognition accuracy of 83.7% AUC.
• Overall debris flow identification accuracy reached 88.1% AUC on the test dataset.
• Demonstrated accurate and reliable debris flow warning capabilities.

Conclusions:

• The proposed deep learning system offers a reliable method for automatic debris flow detection and recognition.
• This technology can significantly improve early warning systems for geological hazards.
• Advance warnings can mitigate damage to infrastructure and protect populations in vulnerable areas.