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

doi:10.1109/TGRS.2025.3608942

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)

DOI

Also indexed in: Semantic Scholar · OpenAlex · dblp · IEEE Xplore

TL;DR. 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.

Quick facts

Method name	S2CMEN; S²CMEN (spelling in the published paper)
Authors	Mengxin Cao, Yongmin Li, Xu Zhang, Guixin Zhao, Guohua Lv, Aimei Dong, Jinyong Cheng, Wei Li, Xiangjun Dong
DOI	10.1109/TGRS.2025.3608942
Task	Hyperspectral image (HSI) classification
Core idea	Superpixel segmentation and classification enhance each other; the two branches are jointly optimized via a unified loss (the abstract calls it a unit loss)
GSAM	Global spatial adaptive module: an adaptive spectral-superpixel network (ASSN) plus a graph convolutional network (GCN) that captures global spatial dependencies by adaptively deriving superpixels
SSFM	Spatial-spectral fusion module: a Spectral-Swin Transformer (SW-Attention and SWS-Attention, convolutional Q/K/V) extracts long-range spectral features and fuses them with the global spatial features
MES	Mutual enhancement strategy: under the unit loss, the superpixel segmentation loss and the classification loss mutually enhance each other
Datasets	Indian Pines (IP), Pavia University (PU) and Houston
Headline result	Overall accuracy (OA): 97.38% on Indian Pines, 92.33% on Pavia University, 91.38% on Houston
Ablation	Fig. 8 reports GSAM ablation experiments on the IP, PU and Houston datasets
Venue	IEEE Transactions on Geoscience and Remote Sensing, vol. 63 (2025), Art. no. 5524714, pp. 1-14; IEEE Xplore document 11159546
Published online	11 September 2025 (IEEE Xplore)
Access	Not open access on IEEE Xplore
BibTeX key	`cao2025mutual`

Main results

Setting	Result	Compared with
Indian Pines (IP)	Overall accuracy (OA) = 97.38% “achieves overall classification accuracies of 97.38%, 92.33%, and 91.38% on Indian Pines (IPs), Pavia University (PU), and Houston, respectively”	—
Pavia University (PU)	Overall accuracy (OA) = 92.33% “achieves overall classification accuracies of 97.38%, 92.33%, and 91.38% on Indian Pines (IPs), Pavia University (PU), and Houston, respectively”	—
Houston	Overall accuracy (OA) = 91.38% “achieves overall classification accuracies of 97.38%, 92.33%, and 91.38% on Indian Pines (IPs), Pavia University (PU), and Houston, respectively”	—

Quoted text is verbatim from the paper.

Key points

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.

Abstract

Most existing hyperspectral image (HSI) classification methods primarily focus on capturing subtle spectral variations by leveraging local spectral–spatial cues derived from patch-level representations. However, limited attention has been given to exploring the global spatial contextual correlations among pixels of HSI. In this study, we propose the superpixel segmentation and classification mutual enhancement network (S²CMEN), a novel framework that integrates global spatial correlations with spectral information through the mutual enhancement of superpixel segmentation and classification. Specifically, a global spatial adaptive module (GSAM) is designed to obtain the direct correlation of the global classes in HSI. It consists of an adaptive spectral-superpixel network (ASSN) and a graph convolutional network (GCN), forming a synergistic architecture that effectively captures global spatial relationships by adaptively deriving superpixel results from HSIs. Notably, GSAM offers a transferable global spatial representation for HSI tasks, enabling integration with other spectral feature extraction models. Furthermore, we develop a spatial–spectral fusion module (SSFM) to obtain comprehensive spectral features and fuse them with the extracted global spatial features. Finally, under the constraint of a unit loss, the mutual enhancement strategy (MES) can make the superpixel segmentation loss and the classification loss mutually enhance each other for better performance. We conducted extensive experiments on three public datasets. The proposed S²CMEN achieves overall classification accuracies of 97.38%, 92.33%, and 91.38% on Indian Pines (IPs), Pavia University (PU), and Houston, respectively, consistently surpassing existing state-of-the-art methods.

Frequently asked questions

Which paper introduced S2CMEN, and how is the name written?

S2CMEN was introduced in "A Mutually Enhancement Network for Superpixel Segmentation and Classification of Hyperspectral Image" (IEEE Transactions on Geoscience and Remote Sensing, 2025). The paper writes the name with a superscript 2, S²CMEN, and expands it as superpixel segmentation and classification mutual enhancement network.

What problem does S2CMEN address?

Most existing HSI classification methods rely on local spectral-spatial cues from patch-level representations, and limited attention has been given to global spatial contextual correlations among pixels. S2CMEN adds them through a global spatial adaptive module (GSAM), in which an adaptive spectral-superpixel network (ASSN) and a graph convolutional network (GCN) capture global spatial relationships by adaptively deriving superpixels. GSAM offers a transferable global spatial representation that can be integrated with other spectral feature extraction models.

How does S2CMEN extract spectral features?

Its spatial-spectral fusion module (SSFM) uses a Spectral-Swin Transformer to extract long-range, detailed spectral features, then fuses them with the global spatial features. The Spectral-Swin Transformer uses spectral window attention (SW-Attention) and spectral window shift attention (SWS-Attention). It computes the query, key and value matrices with convolutional operations to better process high-dimensional spectral features.

How are superpixel segmentation and classification trained together in S2CMEN?

The segmentation and classification branches are jointly optimized via a unified loss, which the abstract calls a unit loss. Under this constraint, the mutual enhancement strategy (MES) makes the superpixel segmentation loss and the classification loss mutually enhance each other.

Which datasets is S2CMEN evaluated on, and what accuracy does it reach?

It is evaluated on three public hyperspectral datasets: Indian Pines (IP), Pavia University (PU) and Houston. Its overall classification accuracies are 97.38%, 92.33% and 91.38%, respectively, and the abstract reports that it consistently surpasses existing state-of-the-art methods.

Where is S2CMEN published, and is it open access?

It appears in IEEE Transactions on Geoscience and Remote Sensing, vol. 63 (2025), Art. no. 5524714, pp. 1-14, DOI 10.1109/TGRS.2025.3608942 (IEEE Xplore document 11159546). The IEEE version is not open access.

How do I cite S2CMEN?

Cite it as: Cao, M., Li, Y., Zhang, X., Zhao, G., Lv, G., Dong, A., Cheng, J., Li, W., & Dong, X. (2025). A Mutually Enhancement Network for Superpixel Segmentation and Classification of Hyperspectral Image. IEEE Transactions on Geoscience and Remote Sensing, 63, Article 5524714. https://doi.org/10.1109/TGRS.2025.3608942 BibTeX is available at https://codezx6.github.io/papers/mutual-enhancement-hsi.html#cite and in https://codezx6.github.io/publications.bib.

中文摘要

用于高光谱图像超像素分割与分类的互增强网络（S2CMEN）

S2CMEN（S²CMEN）通过超像素分割与分类的互增强，将像素间的全局空间相关性与光谱信息结合，用于高光谱图像分类。全局空间自适应模块 GSAM 由自适应光谱-超像素网络 ASSN 与图卷积网络 GCN 组成，自适应地导出超像素以捕获全局空间关系；空间-光谱融合模块 SSFM 用 Spectral-Swin Transformer 提取长程光谱特征，并与全局空间特征融合；互增强策略 MES 在统一损失约束下使超像素分割损失与分类损失相互促进。在 Indian Pines、Pavia University、Houston 上总体分类精度分别为 97.38%、92.33%、91.38%。

关键词：高光谱图像分类；超像素分割；特征融合；Transformer；图卷积网络；互增强

Keywords

hyperspectral image classification superpixel segmentation feature fusion Transformer graph convolutional network mutual enhancement global spatial context remote sensing

Also referred to as: S²CMEN; Superpixel Segmentation and Classification Mutual Enhancement Network; GSAM (global spatial adaptive module); SSFM (spatial-spectral fusion module); MES (mutual enhancement strategy); Spectral-Swin Transformer.

Research area map and related search terms

Field path (broad to narrow): artificial intelligence › computer vision › remote sensing › Earth observation › hyperspectral imaging › hyperspectral image classification › superpixel-based HSI classification › joint segmentation and classification › S2CMEN

Task

hyperspectral image classification HSI classification superpixel segmentation remote sensing image classification

Method

superpixel-based classification graph convolutional network (GCN) Swin Transformer spatial-spectral fusion global spatial context multi-task mutual enhancement

Datasets

Indian Pines Pavia University Houston

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

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

Applications

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

Related benchmarks (not used in this paper)

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

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

Terms are grouped by role. "Related areas", "Applications" and "Related benchmarks (not used in this paper)" describe the surrounding field, not results of this paper. See the site-wide research area map.

Related papers

Hyperspectral image classification

S3CFSL: Spatial-Spectral–Semantic Cross-Domain Few-Shot Learning for Hyperspectral Image Classification (IEEE Transactions on Geoscience and Remote Sensing, vol. 62, article 5525315, 2024)

Cite this paper

@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}
}

Cao, M., Li, Y., Zhang, X., Zhao, G., Lv, G., Dong, A., Cheng, J., Li, W., & Dong, X. (2025). A Mutually Enhancement Network for Superpixel Segmentation and Classification of Hyperspectral Image. IEEE Transactions on Geoscience and Remote Sensing, 63, Article 5524714. https://doi.org/10.1109/TGRS.2025.3608942