Intelligent identification of silo-cave traditional residences in the Chinese Loess Plateau based on deep learning

Silo-caves are a unique traditional cave-dwelling architecture on the Loess Plateau in the middle Yellow River region, representing precious material evidence of the construction wisdom of Chinese civilization. However, such valuable cultural heritage is under the dual threats of natural erosion and human destruction, resulting in the rapid disappearance of many silo-caves. Owing to their wide and scattered distribution, traditional field surveys and manual mapping are time-consuming and inefficient, and are restricted by complex terrain and potential collapse risks, which can hardly meet the demand for large-scale and efficient information collection. This paper compares the applicability of mainstream deep learning object-detection models for silo-cave identification using Google Earth images. Experiments show that MobileNet has obvious advantages over Faster R-CNN, SSD, RetinaNet, and YOLO in model size, inference speed, and detection accuracy. A Cascading Tile Inference Algorithm (CTIA) is proposed to improve recognition in edge regions of remote sensing images. Silo-cave settlements display a spatial pattern of overall dispersion and local clustering, coexisting harmoniously with modern above-ground buildings. This study verifies the feasibility of deep learning in silo-cave heritage detection and provides a technical reference for the efficient investigation and conservation of this distinctive traditional architectural heritage.