Research on Remote Sensing Image Feature Interaction and Fusion Change Detection Method Based on Axial Attention Mechanism

Remote sensing image change detection (RS-CD) has emerged as a prominent research focus in geospatial analysis. While current deep learning approaches predominantly employ Siamese encoders for feature extraction followed by decoding to identify change features, such architectures still face critical limitations. Specifically, conventional encoders process dual-temporal images independently, neglecting interactive feature learning, while simplistic fusion strategies in the decoding stage often lead to substantial information loss. To address these problems, a change detection network with feature interaction and fusion based on axial attention mechanisms is proposed. A feature interaction module is designed in the encoder, which uses the axial attention mechanism to model remote dependence on feature information and introduces attention weighted features into each other concurrently to improve the feature extraction capability of the encoder. In the feature fusion stage in the decoder, a feature fusion module is designed to perform remote dependence modeling using the columnwise axial attention mechanism after certain processing of the dual-time-phase features, aiming at retaining more information and highlighting changing features in the fusion stage. The experimental results show that the $F1$ scores of the proposed method on the publicly available datasets LEVIR-CD and WHU-CD are 91.36% and 92.50%, respectively, which are better than about eight state-of-the-art methods compared in this letter, proving the effectiveness of the proposed method.