Remote Sensing for Natural Resources >
A classification network of hyperspectral images with multi-scale feature fusion
Received date: 2024-02-02
Revised date: 2024-05-07
Online published: 2026-06-03
The classification of hyperspectral images faces challenges like ineffective extraction of multi-scale features and easy loss of pose information. Considering these challenges, this study proposed a classification network of hyperspectral images with multi-scale feature fusion-the hierarchical multi-scale concatenation net (HMC-Net). Initially, multi-scale convolution kernels were applied for parallel computing to extract multi-level features. Meanwhile, the 1×1 convolutional kernels were employed to reduce input-output dimensions, balancing computational complexity. These operations enabled efficient feature extraction without significantly increasing the overall computational burden. Subsequently, independent capsule networks were used for parallel processing of features at various scales. The max pooling was improved via dynamic routing to enhance the translation invariance of features, thereby reducing the loss of pose information. Finally, the concatenate operation integrated feature maps of different scales, thereby achieving a precise analysis of multi-level information in the classification of hyperspectral images. Comparative experimental results demonstrate that the HMC-Net achieved an overall accuracy of 94%, 98%, and 99% on the Kennedy Space Center, University of Pavia, and Salinas datasets, respectively. Compared to the latest classification model of hyperspectral images, the HMC-Net exhibited significant performance advantages, validating its effectiveness.
WEI Lin , RAN Haoxiang , YIN Yuping . A classification network of hyperspectral images with multi-scale feature fusion[J]. Remote Sensing for Natural Resources, 2025 , 37(3) : 113 -122 . DOI: 10.6046/zrzyyg.2024060
表1 HMC-Net模型组合结构Tab.1 Combination structure of HMC-Net |
| 模型名称 | 模型介绍 |
|---|---|
| HMC-0 | 只使用单尺度卷积核的胶囊网络模型 |
| HMC-1 | 使用最大池化层的多尺度卷积网络模型 |
| HMC-Net(本文) | 多尺度网络模型+改进最大池化层的网络模型 |
表2 消融实验结果Tab.2 Results of ablation experiment(%) |
| 模型名称 | OA | AA | Kappa |
|---|---|---|---|
| HMC-0 | 97.42 | 97.01 | 98.02 |
| HMC-1 | 92.43 | 91.96 | 92.36 |
| HMC-Net(本文) | 99.31 | 98.86 | 99.23 |
表3 肯尼迪航天中心数据集定量对比实验结果Tab.3 Quantitative comparative experimental results on the Kennedy space center dataset(%) |
| 量化 指标 | SPP | DCNN | 3-D CNN | SPL- SR | CNN_ HSI | Spectral- NET | HMC-Net (本文) |
|---|---|---|---|---|---|---|---|
| AA | 92 | 92 | 86 | 91 | 89 | 85 | 92 |
| OA | 91 | 93 | 93 | 92 | 93 | 87 | 94 |
| F1分数 | 95 | 94 | 86 | 92 | 89 | 85 | 93 |
| 召回率 | 93 | 93 | 93 | 93 | 93 | 86 | 94 |
表4 帕维亚大学数据集定量对比实验结果Tab.4 Quantitative comparative experimental results on the Pavia university dataset(%) |
| 量化 指标 | SPP | DCNN | 3-D CNN | SPL- SR | CNN_ HSI | Spectral- NET | HMC-Net (本文) |
|---|---|---|---|---|---|---|---|
| AA | 95 | 88 | 94 | 88 | 93 | 89 | 97 |
| OA | 93 | 92 | 94 | 84 | 94 | 91 | 98 |
| F1分数 | 94 | 89 | 92 | 86 | 92 | 92 | 98 |
| 召回率 | 96 | 90 | 94 | 89 | 96 | 90 | 97 |
表5 萨利纳斯数据集定量对比实验结果Tab.5 Quantitative comparative experimental results on the Salinas dateset(%) |
| 量化 指标 | SPP | DCNN | 3-D CNN | SPL- SR | CNN_ HSI | Spectral- NET | HMC- Net(本文) |
|---|---|---|---|---|---|---|---|
| AA | 81 | 93 | 98 | 96 | 97 | 94 | 99 |
| OA | 76 | 89 | 96 | 93 | 95 | 89 | 99 |
| F1分数 | 83 | 92 | 97 | 95 | 98 | 94 | 99 |
| 召回率 | 78 | 87 | 99 | 94 | 96 | 92 | 99 |
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
SabourS, FrosstN,Hinton G E. Dynamic routing between capsules[C]// Proceedings of the 31st International Conference on Neural Information Processing Systems.Curran Associates,Inc., 2017:3859-3869.
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
|
| [34] |
|
| [35] |
|
| [36] |
|
/
| 〈 |
|
〉 |