2024, 46(4): 1448-1457.
doi: 10.11999/JEIT230651
Abstract:
HyperSpectral Image (HSI) classification methods based on limited labeled samples have made significant progress in recent years. However, due to the specificity of hyperspectral images, redundant information and limited labeled samples pose great challenges for extracting highly discriminative features. In addition, owing to the uneven distribution of pixels in each category, how to strengthen the role of central pixels and attenuate the negative impact of surrounding pixels with different categories is also the key to improve the classification performance. To overcome the above limitations, an HSI classification method based on Multi-Scale Asymmetric Dense Network (MS-ADNet) is proposed. Firstly, a multi-scale sample construction module is proposed, which extracts multiple scale patches around each pixel and performs deconvolution and stitching to construct multiscale input samples that contain both detailed structural regions and large homogeneous regions. Next, an asymmetric densely connected structure is proposed to achieve kernel skeleton enhancement in joint spatial and spectral feature extraction, i.e., enhancement of features extracted from the central cross-skeleton portion of a square convolutional kernel, which effectively facilitates feature reuse. Moreover, to improve the discriminability of spectral features, a streamlined element spectral attention mechanism is proposed and placed at the front and back ends of the densely connected network. With only five samples per class used for network training, the proposed method achieves competitive classification results with overall accuracies of 77.66%, 84.54%, and 92.39% on the Indiana Pines, Pavia University, and Salinas datasets, respectively.