Semantic segmentation of historic landscape system along Jiangnan Canal based on deep learning and multi-modal geodata

Abstract The Jiangnan Canal historical landscape system has sustained socioeconomic development for millennia. To better extract heritage landscape value from historical imagery, a Geo-SegFormer historical landscape semantic segmentation method integrating deep learning with multi-modal geospatial data is proposed. Firstly, the panchromatic satellite imagery is enhanced into three differentiated bands using the CLAHE algorithm. Subsequently, based on the formation mechanism of canal-associated landscapes, topographic and hydrographic data are introduced as auxiliary features and fused with enhanced images to generate five-channel inputs. Additionally, dynamic channel scaling, dynamic weight loading, and parameter migration mechanisms are designed to enable the SegFormer model to automatically adapt to tasks with arbitrary input channels and class counts while retaining pretrained weights. Finally, using the self-constructed dataset, the proposed method restores 35 fine-grained landscape categories at 1-meter resolution across the entire Jiangnan Canal region. Ablation experiments and XGBoost-SHAP interpretation further confirm the model’s effectiveness. This research advances historical panchromatic imagery application for interpreting the Jiangnan Canal landscape system while providing a flexible framework for incorporating additional multi-modal data in future semantic segmentation tasks. This research not only advances the application of historical panchromatic imagery to interpret the Jiangnan Canal landscape system but also provides a flexible framework adaptable to additional multi-modal data for diverse future segmentation tasks.