Blur-ASKNet: An Effective Network for Remote Sensing Image Classification Capturing Joint Features

Remote sensing image classification faces two major challenges: the intra-class variability and inter-class similarity. Consequently, it is more challenging for classification neural networks to extract discriminant features. While many studies have proposed solutions, they struggle to balance model performance with discriminant feature extraction. In this paper, we propose the Blur-ASKNet. The network can effectively capture both local and global features, referred to as joint features. The network employs blurpooling to handle intra-class variability, incorporates spatial and channel attention mechanisms for multiscale adaptive feature fusion, and introduces ASKM to extract joint features at both local and global levels. Additionally, it can be deployed on small and medium-sized systems to address inter-class similarity. We conducted comparative and ablation experiments on four typical remote sensing datasets to explore the differences between remote sensing and traditional images and evaluate individual modules of the proposed model. The model performance was subsequently measured. Experimental results demonstrate that the network achieved higher recognition accuracy than the state-of-the-art methods, while still maintaining acceptable complexity.