A new contrastive-learning framework enables hyperspectral classification with few labels. It masks redundant spatial and spectral information, mixes samples into positive pairs, and separates spatial and spectral features. On four benchmarks, using five labels per class, it improved key accuracy metrics, offering a practical route to cost-effective, robust land-cover mapping.
Hyperspectral images capture rich spatial and spectral information and support fine-grained land-cover analysis. Deep models are data-hungry, while hyperspectral annotation is expensive. Existing contrastive pretraining often aligns samples under spatial transformations to learn spatial invariance. Although effective, such methods struggle with spectral variation and build positive samples mainly from changes to the same object, overlooking joint spatial–spectral relationships. This limits learning from unlabeled data and accuracy when labels are scarce. Background clutter and spectral redundancy further complicate feature extraction. Robust, transferable features are therefore essential. Given these challenges, research is needed on label-efficient joint spatial–spectral representation learning for hyperspectral classification.
Researchers at Shenzhen University's College of Computer Science and Software Engineering and Wuhan University's School of Computer Science published (DOI: 10.34133/remotesensing.1074) the study in J ournal of R emote S ensing on 23 September 2026. They address a practical bottleneck: hyperspectral classifiers require many labeled pixels, but annotation is costly and spectral variability can confuse models trained mainly with spatial augmentations. The work targets accurate land-cover classification when only five labeled samples per class are available. It extracts transferable knowledge from unlabeled data before fine-tuning.
The Contrastive Learning with Spatial–Spectral Masking and Mixing (CLSMM) framework introduces spatial–spectral masking that removes redundant information while preserving semantics, cross-sample mixing that creates harder positive pairs, and a spatial–spectral decoupling network with separate spectral and spatial branches. With only five labels per class, CLSMM achieved overall accuracy (OA), average accuracy (AA), and Kappa coefficient of 92.62%, 95.30%, and 0.9051 on Pavia University; 98.26%, 96.14%, and 0.9754 on Pavia Center; 96.36%, 94.10%, and 0.9560 on Yan Cheng; and 89.68%, 92.98%, and 0.8855 on Salinas. Outperformed small-sample methods and produced sharper boundaries. Ablations confirmed that spatial and band-cluster masking are complementary.
During pretraining, 50% of unlabeled samples were used with 23×23 input patches. Superpixel segmentation with Simple Non-Iterative Clustering (SNIC) and spectral clustering into 10 band clusters defined structured masking units. A mask ratio removed selected superpixels within band clusters, and a mixing ratio blended positive samples from different objects. The student–teacher network used a negative-cosine loss and an exponential moving average (EMA) teacher. The spatial branch used 3×3 convolutions and residual modules; the spectral branch used the central pixel’s spectral vector and 1D convolutions; features were concatenated. Fine-tuning used five labels per class, averaged over ten random samplings. Combining superpixel and band-cluster masking gave the best results. On Pavia University, decoupling alone reached 72.46% OA, while adding contrastive pretraining raised it to 92.62%. The model used about 1.36 million parameters on Pavia University. This design reduced redundancy while preserving semantic cues and mitigated background interference. These choices balanced removal and preservation.
The team explains that the method forces the model to recognize semantically consistent ground objects despite missing spatial and spectral information. Decoupling prevents one information type from corrupting the other during feature extraction. Authors say the results support label-efficient hyperspectral mapping and plan to explore physics-informed augmentation and semi-supervised learning.
The study used four public benchmarks: Pavia University, Pavia Center, Yan Cheng, and Salinas. It pretrained on 50% unlabeled patches, then fine-tuned with five labeled samples per class and reported averages over ten random draws. Spatial–spectral masking combined SNIC superpixels with spectral band clustering; proportional mixing generated virtual positive pairs. A dual-branch convolutional neural network (CNN) extracted spatial context and central-pixel spectral vectors separately, then concatenated them. A teacher–student contrastive objective with EMA stabilized pretraining; cross-entropy fine-tuned the student branch.
The approach could support precision agriculture, environmental monitoring, mineral exploration, and large-area land-cover mapping where labels are scarce. Because pretraining uses unlabeled data and inference retains only the student branch, it may be practical for deployment. Future work aims to add physics-informed augmentation, fuse external and multimodal data, and adopt semi-supervised, self-supervised, and active-learning strategies to reduce annotation dependence further. These directions may broaden access to accurate hyperspectral analysis in data-limited real-world settings and improve robustness in complex scenes globally.
###
References
DOI
Original Source URL
https://spj.science.org/doi/10.34133/remotesensing.1074
Funding information
This work was supported by the National Natural Science Foundation of China (Grant Nos. U25A20802 and 62271327), Shenzhen Science and Technology Program (Grant Nos. JCYJ20240813142308012, RCJC20221008092731042, and KCXFZ20240903093100001), Project of Department of Education of Guangdong Province (Grant No. 2023KCXTD029), AI4S Program of Shenzhen University (Grant No. 2026AI4S11), Research Team Cultivation Program of Shenzhen University (Grant No. 2023JCT002), Open Fund of State Key Laboratory of Remote Sensing and Digital Earth (Grant No. OFSLRSS202510), Open Fund of National Earth Observation Data Center (Grant No. 2025C1), and Scientific Foundation for Youth Scholars of Shenzhen University (Grant No. 827-0001080).
About J ournal of R emote S ensing
J ournal of R emote S ensing , an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.
Journal of Remote Sensing
Not applicable
Contrastive Learning with Spatial–Spectral Masking and Mixing for Hyperspectral Image Classification
23-Sep-2026
The authors declare that they have no competing interests.