A Hybrid Method for Remote Sensing Change Detection

Remote Sensing Change Detection (RSCD) holds significant value across various fields, primarily aimed at identifying surface alterations through the analysis of remotely sensed images acquired at different times. The rapid advancement of Deep Learning (DL) and Machine Learning (ML) technologies has echoed in this domain as well, with DL and ML- based RSCD methods showcasing commendable performance, akin to their breakthroughs in other areas. However, existing methods either suffer from limited applicability or fail to strike a balance between Change Detection (CD) performance and efficiency due to constrained model design. In this paper, we propose a hybrid RSCD method combining Convolutional Neural Networks (CNN) and Graph Neural Networks (GNN). Broadly, we employ CNN to extract semantic features at various levels, and considering the characteristics of these multi-level semantic features, we utilize a GNN-based approach for feature interaction. Additionally, we introduce hierarchical feature supervision to train a more proficient model. The results on high-resolution 512 × 512 datasets for building change detection and agricultural change detection explicitly indicate that our method outperforms the current state-of-the-art CD methods.