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

Debris flows are characterized by their suddenness, rapidity, large scale and destructive power, causing serious threat to the population in mountainous areas. Surveillance cameras are widely used in geological hazard monitoring and early warning projects. So far, video cameras are used as a passive tool for post inspection and not as an active role for debris flow monitoring and early warning. Inspired by recent developments of anomaly detection in the field of computer vision, in this paper, we propose a novel automatic debris flow detection and recognition system based on deep learning. It consists of a video feature extraction network using a 3D convolutional neural network (CNN), a debris flow hazard detection network using a multi-layer perceptron (MLP), and a debris flow hazard recognition network for verification employing another CNN. The proposed system takes the video sequences captured by the cameras as inputs and enables the detection and recognition of debris flow hazards. All the networks are optimized and evaluated on a newly annotated image dataset called Debrisflow23. Extensive experimental evaluations with a detection accuracy of 86.3 % AUC, a recognition accuracy of 83.7 % AUC, and an overall identification accuracy of 88.1 % AUC on the test dataset demonstrate that the proposed method possesses accurate and reliable debris flow warning capability. Thus, further precautions can be taken in advance to reduce the damage to human settlements and infrastructure caused by debris flows.