Spatial-Frequency Aggregation Transformer for Remote Sensing Scene Classification

Due to its powerful global feature learning ability, the transformer has been widely applied in remote sensing image scene classification. However, for remote sensing scene images with complex scenes, it is crucial to reasonably construct a global-local feature learning framework while balancing global feature learning in order to effectively describe the scene images. In this learning framework, this paper incorporates a spatialfrequency feature learning mechanism to capture the local features of remote sensing scene images from both the spatial and frequency domains, proposing the spatial-frequency aggregation transformer (SFAT). The proposed network effectively integrates the multi-head mechanism into spatial convolution features and frequency domain wavelet feature learning. Specifically, consistent with classical wavelet transforms, after extracting wavelet features, the proposed method utilizes <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1 \times 1$</tex> convolutions to automatically learn the main feature information of the high and low-frequency components from the four subcomponents, thus completing the wavelet inverse transform more effectively. Additionally, the network integrates a grouped aggregation module to enable feature interaction among the multi-head features. Experimental results on two remote sensing image scene classification datasets demonstrate the effectiveness of SFAT.