Correlated Mapping Attention Cooperative Network for Urban Remote Sensing Image Segmentation

In current remote sensing segmentation tasks, the difficulty of segmenting spectrally similar objects is a significant issue. Solving this problem is crucial for improving segmentation accuracy. Traditional image-domain segmentation methods rely on color and texture features, but spectrally similar objects have negligible color differences, leading to suboptimal segmentation results. To address this, we propose a network framework called Correlated Mapping Attention Cooperative Network (CMACNet) by extending the problem from the image domain to the feature domain. Image-domain methods depend on color and texture features, whereas feature-domain methods process higher-level abstract features, avoiding issues caused by color similarity. Specifically, CMACNet first employs an autoencoder structure. The autoencoder compresses the input data and attempts to reconstruct the original data, ensuring that the latent space representations capture essential and representative features of the input data, thereby extracting highly generalized and versatile features. Next, we introduce the correlated mapping attention mechanism, which adaptively adjusts the attention to different features based on their correlations, effectively addressing the challenge of segmenting spectrally similar objects. Furthermore, to efficiently establish global relationships among features, we design a cross global interaction layer for global feature remapping. Comprehensive experiments on the Vaihingen and Potsdam datasets demonstrate that CMACNet outperforms existing state-of-the-art methods, achieving mean intersection over union scores of 84.77% and 87.69%, respectively.