Joint Shifted Attention and Cross-Guided Feature Fusion for Remote Sensing Scene Classification

Remote sensing scene classification (RSSC) has become a research hotspot due to its widespread application in various fields. However, high-resolution remote sensing images generally exhibit high inter class similarity and intra class diversity, which poses significant challenges for remote sensing scene classification. This article proposes a novel remote sensing scene classification network, shifted attention and cross guided feature compensation network (SACGNet). Firstly, a shift self-attention (SSA) module was constructed, which enhances the model's local perception ability by introducing shift convolution, while guiding the attention mechanism to efficiently model long-distance dependencies. Secondly, a cross-guided feature compensation module (CGFC) was designed, which utilizes a gating mechanism to achieve bidirectional guidance and fine-grained fusion between the original features and enhanced features, thereby improving the class discrimination power in complex scenes. In addition, this article proposes a structure aware learning rate scheduling strategy SGLR, combined with the cosine decay stagnation disturbance mechanism (CASP), which adaptively adjusts the cosine learning rate using attention entropy, gradient sharpness, and feature structure index. Attenuation disturbance is applied during the performance stagnation stage to achieve tightly coupled optimization scheduling with feature structures, effectively helping the model get rid of local optima. Numerous experiments were conducted on three representative remote sensing scene classification datasets, UCM, AID, and NWPU. The experimental results showed that the proposed SACGNet achieved better classification performance than existing advanced methods at multiple training ratios, demonstrating the comprehensive advantages of the proposed method in classification performance and robustness.