A Spectral Group-Wise Gated CNN–Mamba Network with Cross-Stage Mutual Distillation for Hyperspectral Image Classification.
Saved in:
| Title: | A Spectral Group-Wise Gated CNN–Mamba Network with Cross-Stage Mutual Distillation for Hyperspectral Image Classification. |
|---|---|
| Authors: | Zhang, Yan1 (AUTHOR), Cao, Xianghai1 (AUTHOR) caoxh@xidian.edu.cn |
| Source: | Remote Sensing. Jun2026, Vol. 18 Issue 11, p1814. 28p. |
| Subjects: | Artificial neural networks, Knowledge transfer, Remote sensing, Image recognition (Computer vision), Feature extraction |
| Abstract: | Highlights: What are the main findings? A CNN–Mamba spectral group-wise gating block (CMSB) is proposed. The CMSB partitions spectral channels into sub-groups and learns per-group gating weights to adaptively balance local CNN features and long-range Mamba context, improving the discrimination of similar land-cover classes. A progressive deep supervision strategy with temperature-regulated cross-stage mutual distillation and uncertainty-based dynamic loss weighting is proposed, alleviating vanishing gradients in shallow layers and enabling bidirectional knowledge transfer across stages. What are the implications of the main finding? On three benchmark HSI datasets (Indian Pines, Pavia University, and Houston 2013), the proposed Spectral Group-Wise Gated CNN–Mamba Network (SGGCMNet) achieves state-of-the-art overall accuracy (OA), average accuracy (AA) and kappa, e.g., 94.96% OA on Indian Pines with only 3% labeled samples, and remains the best across training ratios from 1% to 20%. The proposed CMSB and the cross-stage mutual-distillation strategy are generic, lightweight, and particularly effective in label-scarce remote sensing scenarios, providing a useful design choice for hybrid CNN–state-space models in HSI classification. Hyperspectral image (HSI) classification enables precise classification of land-cover types from rich spectral and spatial information. Recent methods combine convolutional neural network (CNN) and Mamba branches to exploit their complementary local and global modeling capabilities for HSI classification. However, most of these methods treat all spectral channels uniformly in feature fusion, failing to account for the discriminability differences across spectral bands. Moreover, most methods rely on a single classification head at the final layer, which may lead to vanishing gradients in shallow layers. To address these limitations, a spectral group-wise gated CNN–Mamba network with cross-stage mutual distillation, called SGGCMNet, is proposed. To address the first limitation, a CNN–Mamba spectral group-wise gating block (CMSB) is designed at the feature-fusion level. Specifically, the CMSB partitions channels into multiple sub-groups along the spectral dimension. Each sub-group learns its own fusion weights that balance local spectral–spatial cues produced by a CNN pathway with long-range context produced by a Mamba pathway. To address the second limitation, two loss-level optimization strategies are proposed jointly: A progressive deep supervision strategy with uncertainty-based dynamic weighting is proposed to attach classification heads at all network stages. A temperature-regulated cross-stage mutual-distillation mechanism is further designed to enable bidirectional knowledge transfer among classification heads at different stages. On three benchmark HSI datasets, SGGCMNet achieves state-of-the-art accuracy. [ABSTRACT FROM AUTHOR] |
| Copyright of Remote Sensing is the property of MDPI and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Highlights: What are the main findings? A CNN–Mamba spectral group-wise gating block (CMSB) is proposed. The CMSB partitions spectral channels into sub-groups and learns per-group gating weights to adaptively balance local CNN features and long-range Mamba context, improving the discrimination of similar land-cover classes. A progressive deep supervision strategy with temperature-regulated cross-stage mutual distillation and uncertainty-based dynamic loss weighting is proposed, alleviating vanishing gradients in shallow layers and enabling bidirectional knowledge transfer across stages. What are the implications of the main finding? On three benchmark HSI datasets (Indian Pines, Pavia University, and Houston 2013), the proposed Spectral Group-Wise Gated CNN–Mamba Network (SGGCMNet) achieves state-of-the-art overall accuracy (OA), average accuracy (AA) and kappa, e.g., 94.96% OA on Indian Pines with only 3% labeled samples, and remains the best across training ratios from 1% to 20%. The proposed CMSB and the cross-stage mutual-distillation strategy are generic, lightweight, and particularly effective in label-scarce remote sensing scenarios, providing a useful design choice for hybrid CNN–state-space models in HSI classification. Hyperspectral image (HSI) classification enables precise classification of land-cover types from rich spectral and spatial information. Recent methods combine convolutional neural network (CNN) and Mamba branches to exploit their complementary local and global modeling capabilities for HSI classification. However, most of these methods treat all spectral channels uniformly in feature fusion, failing to account for the discriminability differences across spectral bands. Moreover, most methods rely on a single classification head at the final layer, which may lead to vanishing gradients in shallow layers. To address these limitations, a spectral group-wise gated CNN–Mamba network with cross-stage mutual distillation, called SGGCMNet, is proposed. To address the first limitation, a CNN–Mamba spectral group-wise gating block (CMSB) is designed at the feature-fusion level. Specifically, the CMSB partitions channels into multiple sub-groups along the spectral dimension. Each sub-group learns its own fusion weights that balance local spectral–spatial cues produced by a CNN pathway with long-range context produced by a Mamba pathway. To address the second limitation, two loss-level optimization strategies are proposed jointly: A progressive deep supervision strategy with uncertainty-based dynamic weighting is proposed to attach classification heads at all network stages. A temperature-regulated cross-stage mutual-distillation mechanism is further designed to enable bidirectional knowledge transfer among classification heads at different stages. On three benchmark HSI datasets, SGGCMNet achieves state-of-the-art accuracy. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 20724292 |
| DOI: | 10.3390/rs18111814 |
