Long Range CCTV-Based Context-Aware Petroglyphs Surface Change Classification Using Dual Attention-Guided Pyramid Siamese Network

Advancements in deep learning and image acquisition technologies have significantly enhanced intelligent remote monitoring capabilities across various applications. Despite these advancements, there is a notable lack of longitudinal studies utilizing long range CCTV images to analyze surface changes in cultural heritage sites. This paper introduces a novel context-aware method for classifying petroglyph surface changes, leveraging deep learning to address this gap. Our proposed method comprises two main components: the context-aware sample selector module (CSSM) and the dual attention-guided pyramid siamese network (DAP-SiamNet). The CSSM enhances data quality by eliminating poor-quality images and detecting anomalies based on environmental context, followed by image normalization and registration. DAP-SiamNet focuses on extracting and integrating critical global features related to surface changes through the sequential feature amplification module and the deep feature consolidation module. We evaluated our methodology using two newly created datasets: a petroglyph anomaly detection dataset and a petroglyph change classification dataset. The experimental results demonstrate that the proposed model achieves outstanding performance with an accuracy of 0.863, an F1 score of 0.875, and an AUC of 0.932, verifying its effectiveness through rigorous field testing. This work provides a reliable solution for long-term monitoring of surface changes in cultural heritage.