Automatic Extraction of Road Networks by Using Teacher–Student Adaptive Structural Deep Belief Network and Its Application to Landslide Disaster
To address the challenge of rapid post-disaster road accessibility assessment (e.g., after landslides), this paper proposes an adaptive Deep Belief Network (DBN) method based on a teacher–student learning framework for fully automatic road network extraction and passability classification from aerial/satellite imagery. The method innovatively integrates Restricted Boltzmann Machines (RBMs), a neuron birth–death algorithm, and layer-wise generative strategies to dynamically optimize both network architecture and parameters, while incorporating lightweight design for real-time inference on edge devices. Evaluated on a benchmark dataset comprising satellite/aerial imagery from seven major cities, the approach achieves a substantial improvement in mean road detection accuracy—from 40.0% to 89.0%. It was successfully deployed to analyze pre- and post-rainfall disaster satellite imagery in Japan, accurately identifying traversable road segments. The method significantly enhances robustness, generalization capability, and practical deployability under complex, real-world conditions.