MHSI-Net.zip

Street identification in historical maps is essential for understanding urban historical landscapes and advancing heritage conservation. Traditional methods extract street characteristics through manually feature engineering, but limits the accuracy and generalizability. While deep learning methods have achieved automatic extraction of complex deep features, they still struggle to capture spatiotemporal variations in streets, resulting in limited performance and interpretability.To address this challenge, this study proposes a novel Multi-modal Historical Street Identification Network (MHSI-Net) that integrates street-related semantics to identify streets from specific construction periods in historical maps. It comprises three primary modules: multi-modal representation learning, multi-modal interactive fusion, and an interpretable attention-enhanced method. These modules are designed to extract features of street morphology from street images and network structure from street graph, facilitate cross-scale semantic fusion and capture local-global connections, and ultimately improve identification accuracy and interpretability.In the task of identifying Yuan streets (1267–1368) on the 1937 Beijing map, our model based on MHSI-Net demonstrated superior performance, outperforming existing benchmarks across various metrics, confirming the feasibility of multi-modal representation learning for fine-grained historical street identification. Additionally, the interpretable attention-enhanced method further improves both interpretability and accuracy regarding the morphological characteristics and spatial distribution of Yuan streets.