Answers are grounded in this paper and cite their sources.
The combination of convolutional neural networks (CNNs) and attention mechanisms (AMs) effectively enhances feature representation capabilities in hyperspectral image (HSI) classification. However, most existing methods face great challenges in terms of parameter numbers and computational overhead, which hinders their applications when computing and storage resources are limited. To address these issues, we propose an ultralightweight multi-domain feature extraction network with cross spatial-spectral attention (ULMN-CS2A) for HSI classification, whichprimarily consists of three modules, i.e., the collaborative frequency-spatial-spectral feature extraction module (CFSS), Gaussian neighboring pixel ReLU activation (GNReLU), and cross spatial-spectral attention (CSSA). Firstly, the ultralightweight CFSS module is designed to replace traditional lightweight convolutional layers by independently extracting features from the frequency, spatial, and spectral domains. Secondly, the GNReLU module enhances the network's nonlinear fitting ability and improves inter-layer information transmission by aggregating neighboring pixels with Gaussian weights. Thirdly, the lightweight CSSA module captures the paired pixel-level spatial-spectral relationships and enhances the global context representation ability by simultaneously learning their interactions. Extensive experiments demonstrate that the proposed ULMN-CS2A method shows strong competitiveness compared to state-of-the-art (SOTA) lightweight methods in terms of model parameters, FLOPs, and classification performance under small sampling rates. Meanwhile, ULMN-CS2A-MSP achieves an excellent classification result of 82.31% in terms of open-OA on SA dataset for open-set HSI classification task.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: