Construction of an intelligent interpretation and change detection system based on UAV remote sensing for landscape architectural heritage conservation

Landscape architectural heritage, comprising classical gardens, imperial complexes, and culturally significant cultivated landscapes, demands continuous fine-grained monitoring that conventional decadal surveys cannot deliver. This study proposes an integrated UAV remote sensing system that couples acquisition, interpretation, and change reasoning within a single conservation-oriented workflow. A multi-scale attention semantic segmentation network, built on a modified U-Net backbone with channel–spatial attention, atrous spatial pyramid pooling, and gated cross-level fusion, addresses the extreme scale disparity among garden elements. A temporal Siamese change detection algorithm with phenology-aware pseudo-change suppression distinguishes seasonal variation from genuine alteration. The two components are embedded in a six-layer architecture spanning data acquisition, preprocessing, intelligent interpretation, change detection, knowledge base, and visualization. Experiments on three heritage sites of contrasting typology—Humble Administrator’s Garden, Chengde Summer Resort, and Gulangyu villa cluster—show that the proposed model reaches 91.78% overall accuracy and 81.25% mIoU, outperforming FCN, SegNet, U-Net, DeepLabV3+ , and Swin-UNet. The change detection algorithm holds false alarm rate below 8% under cross-season pairings, against 15–34% for representative baselines. An eighteen-month deployment yielded mean early-warning lead time of 27.3 days and processing throughput of 1.93 ha/h, supporting the system’s practical value for preventive conservation, routine inspection, and post-disaster damage assessment of landscape heritage.