Hyperspectral Images Ground Object Recognition Based on Split Attention

Abstract

Conventional hyperspectral image (HSI) recognition methods are based on artificially designed descriptors to extract superficial features that restrict ground object recognition capabilities. Convolutional neural networks have performed excellently in HSI recognition. However, traditional networks are expensive computational cost in the training process and usually ignore the channel discrimination of feature maps. In order to alleviate these problems, we propose a ground object recognition model that combines the principal component analysis (PCA) method and the residual split-attention network (ResNeSt) for HSI recognition. PCA is applied to minimize the spectral correlation of the HSI data by linear mapping. For the network module, we introduce grouped convolution to reduce network parameters, and the split attention module recalibrates the importance of the network feature map channels. Experimental results on three real HSI datasets demonstrate that the proposed model can effectively extract features from the HSI dataset and accelerate its convergence. Compared with other network modules, the ResNeSt module has fewer calculation parameters and obtained top-ranking accuracy for HSI recognition that demonstrate the module is efficient.