Hyperspectral Image Classification Using a Superpixel–Pixel–Subpixel Multilevel Network

Abstract

Hyperspectral images (HSIs) often contain irregular ground cover with mixed spectral features and noise, which makes it challenging to identify the ground cover using only pixel features, superpixel features, or a combination of both. To alleviate the above problem, this paper proposes a superpixel-pixel-subpixel multilevel network (SPSM), which compensates for the insufficiencies of the different levels and decrease the information loss. For arbitrary irregular regions, superpixel features are simulated as network nodes using a graph convolutional network (GCN) to capture the spatial texture structure of the HSI, which improves the smooth classification results of local regions while facilitating the identification of different vegetation features in the region. Additionally, the global attention module (GAM) learns local regular regions based on pixel-level features to extend the global interactive representation capability and reduce information loss. To overcome spectral mixing and enhance material discrimination, the normalized attention module (NAM) is used to suppress unimportant subpixel information and identify and remove irrelevant details, thereby improving the identification of critical features that differentiate different materials. Finally, the three features are fused to build a SPSM classification framework to improve robustness to overfitting, reduce computational complexity, and facilitate target recognition. Experimental results on four HSI datasets demonstrate that the method is more capable of recognizing detailed features than other advanced comparison methods.