Improved CNN and multisource remote sensing fusion for great wall weathering erosion monitoring

Weathering is eroding the Great Wall of China faster than heritage inventories can track, and two problems keep any remote-sensing–based response honest: reliably picking the physical wall body out of spectrally similar background, and turning what is picked out into a spatially explicit account of how erosion is unfolding along the corridor. We address both in one framework. Building on ResNet-50, GW-ResNet couples dilated convolution branches, dual channel–spatial attention modules, and a multi-scale feature pyramid, and it sits inside a three-stage progressive fusion pipeline that combines optical and SAR data at the pixel, feature, and decision levels for wall-ontology recognition; LiDAR-derived elevation is deliberately routed downstream to the erosion metrics rather than into the recognition fusion, so that the two problems remain separately auditable. A cross-modal attention mechanism enables selective information exchange between heterogeneous data streams, while Dempster-Shafer evidence theory manages residual inter-source conflicts at the decision stage. Applied to a representative rammed-earth Wall corridor in the Ningxia–Gansu region of northwestern China, the model achieved an overall accuracy of 94.6%, an F1 score of 92.7%, and an IoU of 86.5%, surpassing DeepLabV3 + by 4.7 and 7.9% points in F1 and IoU, respectively. Multi-source fusion yielded a 5.8-point F1 improvement over optical-only input, with SAR data proving particularly effective for discriminating weathered earthen walls from spectrally similar loess terrain. Spatiotemporal erosion analysis over 2015–2023 revealed that critically deteriorated segments expanded from 14.2% to 21.7% of the corridor length, with persistent hotspots concentrated along exposed ridgelines and loess bluff edges. Geographically weighted regression identified spatially varying erosion drivers, with aeolian abrasion dominating in western arid sectors and precipitation-induced gullying prevailing in central loess terrain. The proposed framework provides both a technical advancement in heritage feature extraction and an evidence-based tool for spatially differentiated conservation planning.