Sep 2026· 2026 IEEE 1st International Conference on Artificial Intelligence Implementation & Applications (ICAIIA)· pp. 268-273· 0 citations· 18 references
TL;DR
Under 5-shot evaluation, the best SCRM variants improve OA over the strongest baseline by 2.21 percentage points, which support SCRM as a lightweight supervised strategy for label-scarce hyperspectral image classification.
Abstract
Hyperspectral image classification (HSIC) remains challenging when only a few labeled pixels are available for each class. Under such label-scarce conditions, deep models easily overfit the limited supervision and often fail to learn perturbation-invariant spatial representations from local patches. To address this issue, this paper proposes a Spatial Consistency Regularization Model (SCRM) for few-shot HSIC. SCRM is built upon a lightweight PCA-based 2D-CNN backbone and introduces a dual-view spatial perturbation mechanism to explicitly regularize prediction and feature stability. For each labeled patch, weakly and strongly perturbed views are generated and jointly optimized. Prediction-level consistency is imposed through a symmetric Kullback-Leibler divergence loss, while feature-level consistency is enforced by a normalized mean squared error constraint. Under 5-shot evaluation, the best SCRM variants improve OA over the strongest baseline by 2.39, 0.75, 1.61, and 2.21 percentage points on Indian Pines, Pavia University, Salinas, and Kennedy Space Center, respectively, while remaining competitive in most 10-shot settings. These results support SCRM as a lightweight supervised strategy for label-scarce hyperspectral image classification.
Few-shot hyperspectral image classification (FS-HSIC) remains highly challenging because the limited number of labeled samples not only makes traditional supervised learning models prone to overfitting but also restricts their ability to learn generalizable spatial–spectral representations. Moreover, existing methods a...
Joint hyperspectral and light detection and ranging (LiDAR) data classification shows great promise for object recognition and land-cover mapping. However, most existing approaches assume clean labels—an assumption that may not hold in practical remote sensing applications—leaving noisy-label learning for real-world hy...
Xia Yue, An-Feng Liu, Ning Chen et al.· IEEE Transactions on Geoscie...· 0 citations
Hyperspectral image (HSI) classification requires the representation of contiguous spectral signatures and spatial context under limited labeled supervision. Vision test-time training (TTT) offers an alternative to fixed-weight token mixing by performing self-supervised inner-model updates on the input sequence, but it...
RS-HCFormer is proposed, a hierarchical conditional encoding Transformer for hyperspectral image classification that substantially outperforms state-of-the-art methods in both accuracy and efficiency.
Qiang Li, Hao-Yue Xu, Jia-Ning Zhang et al.· 3rd International Conference...· 0 citations
The adaptive basis spectral network (ABSNet), a lightweight domain generalization (DG) framework that learns adaptive low-rank spectral bases for constructing sample-specific low-rank adapters that achieves accuracy comparable to state-of-the-art (SOTA) methods with only a small fraction of the parameters of the compar...
Jian Qiu, Meng-Qi Chai, Shu-Ming Xiong et al.· IEEE Transactions on Geoscie...· 0 citations
Hyperspectral image classification requires joint modeling of spectral signatures and spatial structure. In patch-based classification, each input patch is assigned the label of its center pixel, although neighboring pixels may belong to different land-cover classes. This setting motivates coordinating shallow spatial...
Long Xian, Heng-Zhuo Ma, Lianlei Shan et al.· Remote Sensing· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.