Comprehensive Attribute Difference Attention Network for Remote Sensing Image Semantic Understanding

In the task of semantic understanding of remote sensing images, most current research focuses on learning contextual information through attention mechanisms or multiple inductive biases. However, these methods are limited in capturing fine-grained differences within the same attribute, are susceptible to background noise interference, and lack effective modeling capabilities for spatial relationships and long-range dependencies between different remote sensing attributes. To address these issues, we specifically focus on the homogenous and heterogenous differences between attributes in remote sensing images. Thus, we propose an innovative comprehensive attribute difference attention network (CADANet) to enhance the performance of understanding remote sensing images. Specifically, we design two key modules: the attribute feature aggregation (AFA) module and the context attribute-aware spatial attention (CAASA) module. The AFA module primarily focuses on global and local domain attribute modeling, reducing the impact of homogenous attribute differences through fine-grained feature extraction and global context information. The CAASA module integrates pixel-level global background information and relative position priors, employing a self-attention mechanism to capture long-range dependencies, thus addressing heterogenous attribute differences. Extensive experimental results conducted on the widely-used Vaihingen, Potsdam and WHDLD datasets effectively demonstrate that our proposed method outperforms other recent approaches in performance. Our code is available at https://github.com/lzp-lkd/CADANet.