Multi-scale Convolutional Attention Fuzzy Broad Network for Few-Shot Hyperspectral Image Classification

Hyperspectral image (HSI) classification is a challenging research hotspot in the field of hyperspectral remote sensing. Due to the limited number of labeled samples, most existing classification models cannot fully utilize the spatial-spectral features of HSI to improve classification performance under few-shot conditions. In this paper, we propose a multi-scale convolutional attention fuzzy broad network (MCAFBN) for few-shot HSI classification. First, we design a multi-scale feature extraction module with convolution and self-attention to extract deep local and global features, in which an active learning (AL) training strategy is adopted. Then, fuzzy broad learning system (FBLS) can not only use a small number of fuzzy rules to learn the complex mapping relationship between the extracted fusion features and HSI labels but also use the enhancement layer to capture more nonlinear feature intersections to improve the fitting ability of the model. We use the classification results of FBLS to generate initial probability maps. Finally, we use guided filter to further correct misclassified samples in the initial probability map, which can better utilize spatial-spectral features to improve classification accuracy under few-shot conditions. Experimental results on three public HSI datasets show that the proposed model achieves state-of-the-art classification performance compared with eight popular models.