Segmenting of historic landscape system along Jiangnan Canal based on deep learning and multimodal geodata

As a living heritage, the Jiangnan Canal landscape features interwoven natural and artificial water networks. To accurately interpret its overall spatial composition and address heritage sustainability challenges, this study proposes Geo-SegFormer: a framework for the automated segmentation of historical landscapes by integrating deep learning with multimodal geospatial data. Using a self-constructed dataset, the proposed method achieves the first 1-meter-resolution reconstruction of the entire Jiangnan Canal landscape system across 35 categories, surpassing traditional manual methods in coverage, diversity, and precision. Quantitative analysis reveals the contribution weights to the segmentation: DEM 51.6%, hydrographic data 17.7%, and imagery data 30.7%, demonstrating that the canal landscape is structured upon natural terrain, with water networks as its spatial framework. This outcome establishes a critical data foundation for heritage conservation and interdisciplinary spatial quantitative research. Beyond this specific case, the developed methodology also possesses considerable potential for cross-regional application.