A Lightweight Dual-Branch Swin Transformer for Remote Sensing Scene Classification

Abstract

The main challenge of scene classification is to understand the semantic context information of high-resolution remote sensing images. Although vision transformer (ViT)-based methods have been explored to boost the long-range dependencies of high-resolution remote sensing images, the connectivity between neighboring windows is still limited. Meanwhile, ViT-based methods commonly contain a large number of parameters, resulting in a huge computational consumption. In this paper, a novel lightweight dual-branch swin transformer (LDBST) method for remote sensing scene classification is proposed, and the discriminative ability of scene features is increased through combining a ViT branch and convolutional neural network (CNN) branch. First, based on the hierarchical swin transformer model, LDBST divides the input features of each stage into two parts, which are then separately fed into the two branches. For the ViT branch, a dual multilayer perceptron structure with a depthwise convolutional layer, termed Conv-MLP, is integrated into the branch to boost the connections with neighboring windows. Then, a simple-structured CNN branch with maximum pooling preserves the strong features of the scene feature map. Specifically, the CNN branch lightens the LDBST, by avoiding complex multi-head attention and multilayer perceptron computations. To obtain better feature representation, LDBST was pretrained on the large-scale remote scene classification images of the MLRSN and RSD46-WHU datasets. These two pretrained weights were fine-tuned on target scene classification datasets. The experimental results showed that the proposed LDBST method was more effective than some other advanced remote sensing scene classification methods.