基于特征子空间解耦与迭代残差精炼的高分辨率遥感影像变化检测OA

With the development of high-resolution remote sensing earth observation technology,the rich texture details in images enhance information content but also introduce complex background noise caused by lighting,shadows,and seasonal phenological differences.To address the issues of pseudo-change misdetection caused by complex background noise and the loss of small target details caused by traditional upsampling in high-resolution remote sensing change detection,a network based on feature decoupling and iterative refinement(DIR-Net)is proposed.First,the pre-trained FastSAM is utilized as a visual prior encoder to extract multi-scale robust features.Next,a Feature Subspace Decoupling Module is designed to explicitly decompose bi-temporal features into a shared semantic subspace and a differential feature subspace through orthogonal projection and cross-recalibration strategies,suppressing environmental noise from the source.Finally,an Iterative Residual Refinement Module is proposed.By introducing a coordinate attention mechanism,the decoding process is modeled as a coarse-to-fine residual regression problem,gradually recovering the edge details of small targets in the resolution-maintained feature space.Experimental results on three public datasets,LEVIR-CD,WHU-CD,and SYSU-CD,demonstrate that the F1 scores of DIR-Net reached 91.33%,93.31%,and 86.29%,respectively.Compared with mainstream ChangeFormer and BIT algorithms,the F1 score improved by an average of approximately 5.0%.The proposed method significantly reduces the false alarm rate of pseudo-changes while maintaining a very high recall rate.This method effectively resolves the challenges of feature coupling and detail loss,demonstrating stronger robustness and higher boundary localization accuracy in complex scenes.