Hyperspectral image classification: cross-domain few-shot learning and superpixel mutual enhancement

Two IEEE Transactions on Geoscience and Remote Sensing papers (Cao et al., 2024; Cao et al., 2025) study hyperspectral image (HSI) classification.

S3CFSL (2024) targets cross-domain few-shot classification, adding feature denoising and a semantic-enhanced domain alignment so that knowledge transfers from a labelled source domain to a new target scene with few labels. S2CMEN (2025) lets superpixel segmentation and classification enhance each other so that the classifier uses global spatial context, and reports overall accuracies of 97.38% on Indian Pines, 92.33% on Pavia University, and 91.38% on Houston. The two papers report overall accuracy under different settings, so their Houston figures are not directly comparable: S3CFSL uses Chikusei as the source domain and five labelled target samples per class, while S2CMEN trains and tests on splits of each dataset.

Papers compared

S2CMEN: A Mutually Enhancement Network for Superpixel Segmentation and Classification of Hyperspectral Image

Mengxin Cao, Yongmin Li, Xu Zhang, Guixin Zhao, Guohua Lv, Aimei Dong, Jinyong Cheng, Wei Li, Xiangjun Dong · IEEE Transactions on Geoscience and Remote Sensing, vol. 63, article 5524714 (2025)

S²CMEN (S2CMEN) is a hyperspectral image classification network in which superpixel segmentation and classification enhance each other through a unified loss, combining superpixel-based global spatial context with Spectral-Swin Transformer spectral features. It reaches 97.38% overall accuracy on Indian Pines.

• Treats superpixel segmentation and classification as mutually enhancing tasks: an adaptive spectral-superpixel network (ASSN) derives superpixels from the image, and the segmentation and classification branches are jointly optimized via a unified loss.
• The global spatial adaptive module (GSAM) combines ASSN with a graph convolutional network (GCN) to capture global spatial dependencies; its global spatial representation can be integrated with other spectral feature extraction models.
• The spatial-spectral fusion module (SSFM) uses a Spectral-Swin Transformer with spectral window attention (SW-Attention) and spectral window shift attention (SWS-Attention), computing query, key and value matrices with convolutional operations.
• Reaches overall classification accuracies of 97.38% on Indian Pines, 92.33% on Pavia University and 91.38% on Houston; Fig. 8 reports a GSAM ablation on all three datasets.

S3CFSL: Spatial-Spectral–Semantic Cross-Domain Few-Shot Learning for Hyperspectral Image Classification

Mengxin Cao, Xu Zhang, Jinyong Cheng, Guixin Zhao, Wei Li, Xiangjun Dong · IEEE Transactions on Geoscience and Remote Sensing, vol. 62, article 5525315 (2024)

S3CFSL classifies new hyperspectral scenes from five labelled samples per class by transferring knowledge from the labelled Chikusei scene. It combines a cross-spatial–spectral transformer, Gaussian feature denoising and semantic-enhanced domain alignment, and reports 98.52% overall accuracy on Pavia Centre, 88.79% on Salinas and 77.63% on Houston.

• With Chikusei as the source domain and five labelled samples per target class, S3CFSL reaches overall accuracy (OA) of 98.52% on Pavia Centre, 88.79% on Salinas and 77.63% on Houston.
• OA, AA and Kappa of S3CFSL exceed all eight baselines on every target dataset, but the OA margins over the closest competitor, GCC-FSL, are small: 0.07, 0.46 and 0.19 points.
• Module ablation (OA on Pavia Centre / Salinas / Houston): plain ViT 94.33/81.46/69.77; SSDC + CSST extractor 96.47/86.39/75.16; plus feature denoising 97.25/87.44/75.69; plus SEDA (full model) 98.52/88.79/77.63.
• Feature denoising adds Gaussian-filtered features back to the unfiltered ones. Gaussian filtering beats median, mean and no filtering on all three datasets; mean filtering even falls below no filtering on Houston.

Key concepts

Hyperspectral image (HSI) classification

Assigning a land-cover class to each pixel of an image with hundreds of narrow spectral bands. Papers: S3CFSL, S2CMEN

Cross-domain few-shot learning

Learning on a labelled source scene and classifying a different target scene from only a few labelled samples per class. Papers: S3CFSL

Superpixel segmentation

Grouping neighbouring pixels into homogeneous regions, used to add global spatial context to pixel classification. Papers: S2CMEN

Benchmarks

Indian Pines, Pavia University, Pavia Centre, Salinas, Houston and Chikusei are standard HSI scenes. Papers: S3CFSL, S2CMEN

Frequently asked questions

Which method handles cross-domain few-shot hyperspectral classification?

S3CFSL (IEEE TGRS 2024). It combines spatial and spectral dual channels with a cross-spatial–spectral transformer, feature-denoising operations, and semantic-enhanced domain alignment (SEDA).

Which hyperspectral datasets are used?

S3CFSL uses Chikusei as the source domain and Pavia Centre, Salinas and Houston as targets; S2CMEN uses Indian Pines, Pavia University and Houston.

Fields, tasks, datasets and related search terms

Task

hyperspectral image classification HSI classification superpixel segmentation remote sensing image classification cross-domain few-shot learning few-shot hyperspectral classification domain adaptation for remote sensing

Method

superpixel-based classification graph convolutional network (GCN) Swin Transformer spatial-spectral fusion global spatial context multi-task mutual enhancement meta-learning (episodic training) feature denoising semantic-aware domain alignment spatial-spectral Transformer spatial-spectral feature extraction

Datasets

Indian Pines Pavia University Houston Chikusei Pavia Centre Salinas

中文

高光谱图像分类 高光谱遥感 超像素分割 图卷积网络 空谱融合 遥感图像分类 跨域小样本学习 小样本学习 域适应 元学习

Fields (broad to narrow)

artificial intelligence computer vision remote sensing Earth observation hyperspectral imaging hyperspectral image classification superpixel-based HSI classification joint segmentation and classification few-shot hyperspectral classification cross-domain few-shot HSI classification

Related areas

image segmentation superpixels (SLIC) graph neural networks graph convolutional networks Vision Transformers Swin Transformer multi-task learning spectral-spatial feature fusion global context modeling land cover classification few-shot learning meta-learning transfer learning domain adaptation domain generalization label-efficient learning spectral-spatial feature learning denoising multispectral imaging

Applications

land-cover mapping precision agriculture environmental monitoring urban mapping geological survey mineral exploration

Related benchmarks (not used in this paper)

Salinas Pavia Centre Kennedy Space Center (KSC) Botswana WHU-Hi Chikusei Indian Pines Pavia University

中文 (扩展)

遥感 对地观测 高光谱成像 图像分割 超像素 图神经网络 视觉Transformer 多任务学习 空谱特征融合 全局上下文建模 地物分类 土地覆盖制图 小样本高光谱分类 迁移学习 域泛化 标签高效学习 空谱特征学习 去噪 环境监测

Compared with

DFSL DCFSL CMFSL Gia-FSL GCC-FSL DMCM

Cite these papers

@article{cao2025mutual,
  title        = {A Mutually Enhancement Network for Superpixel Segmentation and Classification of Hyperspectral Image},
  author       = {Cao, Mengxin and Li, Yongmin and Zhang, Xu and Zhao, Guixin and Lv, Guohua and Dong, Aimei and Cheng, Jinyong and Li, Wei and Dong, Xiangjun},
  journal      = {IEEE Transactions on Geoscience and Remote Sensing},
  year         = {2025},
  volume       = {63},
  pages        = {5524714},
  doi          = {10.1109/TGRS.2025.3608942},
  issn         = {0196-2892},
  url          = {https://doi.org/10.1109/TGRS.2025.3608942}
}

@article{cao2024s3fsl,
  title        = {Spatial-Spectral–Semantic Cross-Domain Few-Shot Learning for Hyperspectral Image Classification},
  author       = {Cao, Mengxin and Zhang, Xu and Cheng, Jinyong and Zhao, Guixin and Li, Wei and Dong, Xiangjun},
  journal      = {IEEE Transactions on Geoscience and Remote Sensing},
  year         = {2024},
  volume       = {62},
  pages        = {5525315},
  doi          = {10.1109/TGRS.2024.3434484},
  issn         = {0196-2892},
  url          = {https://doi.org/10.1109/TGRS.2024.3434484}
}