An urban change detection method based on multimodal data and knowledge graph technology

Urban change detection faces critical challenges in capturing comprehensive transformations across morphological, environmental, social, and economic dimensions. Knowledge graphs demonstrate exceptional compatibility with multimodal geospatial data, providing a novel approach for change detection. However, existing knowledge graphs are predominantly static and lack deep fusion between features, limiting their direct application to change detection. To address these limitations, this study proposes an urban change detection method based on multimodal data and knowledge graph technology. First, the study develops a multimodal bitemporal urban knowledge graph (MBUKG) that integrates multimodal geographical data. Second, the study proposes a dual cross-attention knowledge representation learning (DCKRL) framework to derive knowledge graph entity vectors. Finally, the study constructs change rate indicators based on cosine similarity to quantify the extent of changes in grid entities between 2017 and 2023, thereby enabling urban change detection. The results demonstrated the effectiveness of the proposed framework, achieving an F1 score of 0.917. The DCKRL framework exhibits robust performance with a Hit@10 value of 0.670. The findings reveal that MBUKG successfully integrates multimodal data with different temporal attributes, while DCKRL effectively captures intricate relationships among entities. The proposed method can provide scientific support for urban planning.