基于特征增强和三流Transformer的高光谱遥感图像全色锐化
|
张杰(1999-),男,硕士研究生,主要研究方向为遥感图像处理等。Email: z1752678k@163.com。 |
Copy editor: 陈昊旻
收稿日期: 2024-06-21
修回日期: 2024-11-13
网络出版日期: 2026-06-03
基金资助
国家自然科学基金项目“基于高阶TV正则化低秩矩阵重构的深度神经网络鲁棒性增强研究”(62072024)
国家自然科学基金项目“退耕还林区森林时空格局变化的循环神经网络建模关键技术研究”(41971396)
建大杰青项目“基于数据驱动的偏微分方程学习及其在图像重建中的应用研究”(JDJQ20220805)
北京建筑大学研究生科研基金项目“基于深度学习网络的真实场景下遥感图像全色锐化研究”(PG2024149)
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
遥感图像的全色锐化是将全色图像(panchromatic images,PAN)与低空间分辨率的高光谱遥感图像(low spatial resolution hyperspectral images,LR-HSI)融合,生成高空间分辨率的高光谱遥感图像(high spatial resolution hyperspectral images,HR-HSI)。当前,基于深度学习的全色锐化方法逐渐成熟,但是依然存在一些不足,例如特征信息提取不够充分、信息融合缺乏指导、全色锐化阶段单一等,导致生成的HR-HSI图像空间和光谱信息不完善。针对以上问题,该文提出一种基于特征增强和三流Transformer的高光谱遥感图像全色锐化方法,采用双阶段实现全色锐化解决单阶段方法的不足。首先,采用特征增强模块与多尺度信息融合模块生成初步增强的高光谱遥感图像,特征增强模块主要从多个尺度上对HSI图像进行空间与光谱信息的增强,多尺度信息融合模块将增强后的HSI图像在不同尺度进行融合; 其次,采用Transformer中注意力思想,将初步增强的HSI,PAN以二者融合生成的图像作为3种特征流,通过线性层将其分别转换为Q、K、V,进而进行多头注意力计算,有效指导空间信息与光谱信息的提取与融合; 再次,利用增强的HSI图像与融合模块进一步提高图像质量,使生成的HR-HSI图像具有更加丰富的空间与光谱信息; 最后,在3个经典的高光谱数据集上进行了实验验证。实验结果表明,该文所提出的全色锐化方法在定量评价指标上优于传统方法和现有的深度学习方法,在定性结果上能够更好地保留高光谱图像的光谱信息和全色图像的空间信息,融合生成更加真实的HR-HSI图像。
关键词: 高光谱遥感图像; 全色锐化; 特征增强; 三流Transformer; 多尺度融合
张杰 , 王恒友 , 霍连志 . 基于特征增强和三流Transformer的高光谱遥感图像全色锐化[J]. 自然资源遥感, 2025 , 37(6) : 97 -106 . DOI: 10.6046/gyzyyg.2024222
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.
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] |
|
/
| 〈 |
|
〉 |