Multiscale parallel network with coordinate attention for hyperspectral image classification

Hyperspectral image classification (HIC) is the foundation of hyperspectral remote sensing applications, and the design of efficient and high-precision hyperspectral image classification models has been a hot direction. In this paper, to improve the poor performance caused by the restricted receptive field of the traditional CNN structure, from the perspective of improving the multi-scale feature extraction capability, we propose a multi-scale parallel network with coordinate attention (MSPCA) for HIC. Specifically, the model first uses multiple 3-D convolution kernels with different receptive fields to extract multi-scale features, and then the extracted features are merged in a fully connected manner. Next step, the resulting features are fed into three branches. Each branch is composed of two 2-D convolutions, and Coordinate Attention (CA) is used before each 2-D convolution to enhance the extraction capability of the model for salient features. MSPCA achieves excellent classification results on all three representative datasets, which proves that increasing the model's multi-scale feature extraction capability are of significant significance for the enhancement of classification performance.