Remote Sensing for Natural Resources >
Pansharpening of hyperspectral remote sensing images based on feature enhancement and Three-Stream Transformer
Received date: 2024-06-21
Revised date: 2024-11-13
Online published: 2026-06-03
Pansharpening of remote sensing images refers to the fusion of panchromatic images (PAN) and low-spatial-resolution hyperspectral (or multispectral) images (LR-HSI/LRMS) to produce high-spatial-resolution hyperspectral (or multispectral) images (HR-HSI/HRMS). Currently, deep learning-based pansharpening methods have increasingly matured. However, pansharpening still faces several challenges, including inadequate feature extraction, insufficient guidance for information fusion, and oversimplified single-stage architectures, resulting in HR-HSI imagery with compromised spatial and spectral fidelity. To address these issues, this paper proposed a two-stage pansharpening method for hyperspectral images based on feature enhancement and a Three-Stream Transformer architecture. In the first stage, preliminarily enhanced hyperspectral images (HSI) were generated using a feature enhancement module and a multi-scale fusion module. Specifically, the feature enhancement module strengthened spatial and spectral information across multiple scales, while the multi-scale fusion module integrated the enhanced HSI at different scales. In the second stage, the initially enhanced HSI, PAN, and images resulting from their fusion were treated as three separate feature streams using the self-attention mechanism of the Transformer. Then, these streams were transformed into the Q(Query), K(Key), and V(Value) matrices via linear layers, followed by multi-head attention computation, which effectively guides the extraction and fusion of spatial and spectral information. Furthermore, the enhanced HSI and an additional fusion module were leveraged to refine image quality, yielding HR-HSI results with richer spatial and spectral details. Validation experiments were conducted on three classic hyperspectral datasets. The results demonstrate that the proposed method outperforms both conventional and existing deep learning-based approaches in terms of quantitative evaluation metrics. Considering qualitative evaluation results, it also preserves spectral information of the HSI and spatial details of the PAN images, producing more realistic HR-HSI images.
ZHANG Jie , WANG Hengyou , HUO Lianzhi . Pansharpening of hyperspectral remote sensing images based on feature enhancement and Three-Stream Transformer[J]. Remote Sensing for Natural Resources, 2025 , 37(6) : 97 -106 . DOI: 10.6046/gyzyyg.2024222
表1 Pavia Center 数据集上的定量实验结果(降低分辨率)Tab.1 Quantitative experimental results on the Pavia Center dataset (reduced-resolution) |
| 方法 | CC/%↑ | SAM↓ | RMSE/ %↓ | ERGAS↓ | PSNR↑ |
|---|---|---|---|---|---|
| PCA | 85.84 | 8.98 | 3.39 | 6.47 | 31.08 |
| BF | 92.43 | 9.60 | 3.51 | 6.78 | 30.81 |
| GS | 95.74 | 6.44 | 2.55 | 4.94 | 32.93 |
| MG | 95.62 | 6.54 | 2.18 | 4.47 | 34.35 |
| PLRDiff | 97.51 | 6.21 | 2.05 | 4.21 | 37.88 |
| PanNet | 96.82 | 6.35 | 1.92 | 3.89 | 35.61 |
| HyperPNN | 96.73 | 6.09 | 1.72 | 3.82 | 36.74 |
| GPPNN | 96.53 | 6.52 | 1.91 | 4.01 | 35.35 |
| HyperKite | 98.04 | 5.61 | 1.29 | 2.85 | 38.97 |
| HyperTransformer | 98.84 | 4.67 | 1.24 | 2.31 | 41.56 |
| 本文方法 | 98.92 | 3.79 | 0.88 | 2.07 | 42.82 |
| 理想值 | 100.00 | 0.00 | 0.00 | 0.00 | ∞ |
表2 Botswana 数据集上的定量实验结果(降低分辨率)Tab.2 Quantitative experimental results on the Botswana dataset (reduced-resolution) |
| 方法 | CC/%↑ | SAM↓ | RMSE/ %↓ | ERGAS↓ | PSNR↑ |
|---|---|---|---|---|---|
| PCA | 94.81 | 2.38 | 1.97 | 2.31 | 39.95 |
| BF | 92.93 | 2.46 | 1.86 | 2.37 | 39.91 |
| GS | 95.10 | 2.31 | 1.95 | 2.19 | 40.21 |
| MG | 96.04 | 2.07 | 1.51 | 1.74 | 41.85 |
| PLRDiff | 97.14 | 1.93 | 1.33 | 1.82 | 42.07 |
| PanNet | 93.36 | 2.17 | 1.53 | 2.71 | 40.41 |
| HyperPNN | 97.28 | 1.74 | 1.18 | 1.44 | 43.45 |
| GPPNN | 96.21 | 1.90 | 1.41 | 1.65 | 42.05 |
| HyperKite | 98.13 | 1.46 | 1.03 | 1.22 | 44.97 |
| HyperTransformer | 98.04 | 1.33 | 0.94 | 1.15 | 45.74 |
| 本文方法 | 98.54 | 1.25 | 0.92 | 1.11 | 45.91 |
| 理想值 | 100.00 | 0.00 | 0.00 | 0.00 | ∞ |
图4 Pavia Center 数据集上的可视化实验结果(降低分辨率)Fig.4 Visualization experiment results on the Pavia Center dataset (reduced-resolution) |
图5 Pavia Center数据集对应图像的残差可视化结果(降低分辨率)Fig.5 Residual visualization results of images corresponding to the Pavia Center dataset (reduced-resolution) |
图6 Botswana 数据集上的可视化实验结果(降低分辨率)Fig.6 Visualization experiment results on the Botswana dataset (reduced-resolution) |
图8 Pavia Center(100,100)数据集上不同网络方法全色锐化结果与地面真值之间的差异Fig.8 Differences between panchromatic sharpening results and ground truth values using different network methods on the Pavia Center (100,100) dataset |
表3 Pavia Center数据集上运行时间以及网络参数总量对比Tab.3 Comparison of runtime and total network parameters on the Pavia Center dataset |
图10 Chikusei 数据集上的可视化实验结果(全分辨率)Fig.10 Visualization experimental results on the Chikusei dataset (full resolution) |
表4 Chikusei 数据集定量实验结果(全分辨率)Tab.4 Quantitative experimental results (full-resolution) on the Chikusei dataset |
| 方法 | ${D}_{\mathrm{\lambda }}$ | Ds | QNR |
|---|---|---|---|
| PCA | 0.053 4 | 0.061 5 | 0.888 4 |
| BF | 0.048 7 | 0.057 4 | 0.896 7 |
| GS | 0.055 8 | 0.080 1 | 0.868 6 |
| MG | 0.041 4 | 0.066 5 | 0.894 9 |
| PLRDiff | 0.034 5 | 0.053 6 | 0.913 7 |
| PanNet | 0.025 7 | 0.032 3 | 0.942 8 |
| HyperPNN | 0.021 5 | 0.029 7 | 0.949 4 |
| GPPNN | 0.027 9 | 0.035 7 | 0.937 4 |
| HyperKite | 0.012 5 | 0.024 5 | 0.962 3 |
| HyperTransformer | 0.015 7 | 0.016 4 | 0.968 2 |
| 本文方法 | 0.013 8 | 0.015 1 | 0.971 3 |
| 理想值 | 0.000 0 | 0.000 0 | 1.000 0 |
表5 Transformer模块里的多头数量(N)消融实验研究(Pavia Center)Tab.5 Multiple head count (N) ablation experimental study in the Transformer module (Pavia Center) |
| N | CC/%↑ | SAM↓ | RMSE/ %↓ | ERGAS↓ | PSNR↑ |
|---|---|---|---|---|---|
| 4 | 97.28 | 4.96 | 1.47 | 2.31 | 40.98 |
| 8 | 98.31 | 4.35 | 1.19 | 2.10 | 42.40 |
| 12 | 98.92 | 3.79 | 0.88 | 2.07 | 42.82 |
| 16 | 98.40 | 3.95 | 1.13 | 2.12 | 42.52 |
| 20 | 97.56 | 4.37 | 1.40 | 2.27 | 41.33 |
表6 FEM和Transformer模块的消融实验研究(Pavia Center)Tab.6 Experimental study on ablation of FEM and Transformer modules (Pavia Center) |
| 方法 | CC/%↑ | SAM↓ | RMSE/ %↓ | ERGAS↓ | PSNR↑ |
|---|---|---|---|---|---|
| B/L | 92.07 | 6.85 | 3.07 | 5.76 | 32.54 |
| B/L+FEM | 96.46 | 5.34 | 2.65 | 4.18 | 35.12 |
| B/L+Transformer | 97.96 | 4.51 | 1.72 | 3.09 | 38.90 |
| B/L+FEM+ Transformer | 98.92 | 3.79 | 0.88 | 2.07 | 42.82 |
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
路琨婷, 费蓉蓉, 张选德. 融合卷积神经网络的遥感图像全色锐化[J]. 计算机应用, 2023, 43(9):2963-2969.
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
|
| [34] |
|
| [35] |
|
| [36] |
|
| [37] |
|
/
| 〈 |
|
〉 |