融合FFT和EMHSA的双时相光学遥感影像变化检测网络
|
王杏伟(1984-),男,硕士,高级工程师,主要从事铁路工程测量、遥感领域的工作。Email:sky_wxw@qq.com。 |
Copy editor: 张仙
收稿日期: 2024-07-31
修回日期: 2025-04-02
网络出版日期: 2026-06-03
基金资助
教育部人文社会科学研究项目“西藏茶马古道传统村落的保护机制与活化路径研究”(23YJA850003)
四川省科技计划项目(重点研发)“化工重大危险源事故监测预警及应急救援决策支撑关键技术研究与示范”(2023YFS0415)
Change detection network for dual-temporal optical remote sensing images integrating fast Fourier transform and efficient multi-head self-attention
Received date: 2024-07-31
Revised date: 2025-04-02
Online published: 2026-06-03
近年来,基于深度学习的遥感变化检测取得了飞速发展,但对于复杂场景的变化检测仍然存在识别不完整、误检率高的问题。该文在SNUnet的基础上,通过融合快速傅里叶变换(fast Fourier transform,FFT)和高效多头自注意力机制(efficient multi-head self attention,EMHSA),提出了FTUNet网络。网络中的FFT模块实现对两时相图像的风格统一,降低由于光照变化等外界因素产生的“伪变化”所引起的错误检测;而在网络的特征提取阶段,引入EMHSA,充分提取特征图的上下文信息,以提高变化目标分割结果的完整性。在LEVIR-CD和SYSU-CD这2个公开数据集上的实验结果表明,FTUNet的F1得分比SNUNet分别提升1.42和1.53百分点,交并比分别提升2.31和2.07百分点。
王杏伟 , 唐康其 , 刘燕 , 刘欢 . 融合FFT和EMHSA的双时相光学遥感影像变化检测网络[J]. 自然资源遥感, 2025 , 37(5) : 113 -121 . DOI: 10.6046/zrzyyg.2024261
The deep learning-based change detection of remote sensing images has seen rapid advances in the past few years. However,it still faces challenges for change detection in complex scenes,such as incomplete recognition and high false detection rates. In response to these challenges,this paper proposed the FTUNet,a network based on SNUnet that integrates the fast Fourier transform (FFT) and efficient multi-head self-attention (EMHSA). Specifically,the FFT module in the network enabled style unification of dual-temporal images,reducing false detection caused by “pseudo changes” due to external factors such as light variations. Additionally,the EMHSA was introduced in the feature extraction stage to fully extract the contextual information from the feature maps,thereby enhancing the segmentation integrity of target changes. Experiments on the LEVIR-CD and SYSU-CD public datasets showed that the FTUNet exhibited increases of 1.42 and 1.53 percentage points in F1 score,as well as increases of 2.31 and 2.07 percentage points in intersection over union (IoU),compared to the SNUNet.
表1 消融实验结果Tab.1 Results of ablation experiment (%) |
| 网络模型 | P | R | IoU | F1 |
|---|---|---|---|---|
| SNUNet(基准网络) | 89.91 | 87.21 | 79.44 | 88.54 |
| SNUNet+FFT | 90.81 | 87.86 | 80.69 | 89.31 |
| SNUNet+LRCFEM | 90.46 | 89.15 | 81.50 | 89.81 |
| FTUNet(本文网络) | 90.72 | 89.22 | 81.75 | 89.96 |
表3 LEVIR-CD数据集对比实验结果Tab.3 Comparison experiment results of LEVIR-CD dataset (%) |
| 网络模型 | P | R | IoU | F1 |
|---|---|---|---|---|
| FC-EF | 86.69 | 79.87 | 71.15 | 83.14 |
| FC-Siam-conc | 89.41 | 85.02 | 77.24 | 87.16 |
| FC-Siam-diff | 90.12 | 83.01 | 76.09 | 86.42 |
| SNUNet | 89.91 | 87.21 | 79.44 | 88.54 |
| BIT | 92.01 | 86.72 | 80.65 | 89.28 |
| FTUNet(本文网络) | 90.72 | 89.22 | 81.75 | 89.96 |
表4 LEVIR-CD数据集对比实验可视化结果Tab.4 Visualization results of LEVIR-CD dataset comparison experiment |
| 序号 | 影像A | 影像B | 标签 | FE-EF | FC-Siam- conc | FC-Siam- diff | BIT | SNUNet | FTUNet |
|---|---|---|---|---|---|---|---|---|---|
| a | ![]() | ||||||||
| b | ![]() | ||||||||
| c | ![]() | ||||||||
| d | ![]() | ||||||||
| e | ![]() | ||||||||
| 图例 | ![]() | ||||||||
表5 SYSU-CD数据集对比实验结果Tab.5 Comparison experiment results of SYSU-CD dataset (%) |
| 网络模型 | P | R | IoU | F1 |
|---|---|---|---|---|
| FC-EF | 78.60 | 69.55 | 58.47 | 73.80 |
| FC-Siam-conc | 81.11 | 70.76 | 60.75 | 75.59 |
| FC-Siam-diff | 90.35 | 49.33 | 46.86 | 63.81 |
| SNUNet | 81.90 | 74.37 | 63.87 | 77.95 |
| BIT | 81.71 | 73.49 | 63.11 | 77.38 |
| FTUNet(本文网络) | 79.95 | 79.01 | 65.94 | 79.48 |
表6 SYSU-CD数据集对比实验可视化结果Tab.6 Visualization results of SYSU-CD dataset comparison experiment |
| 序号 | 影像A | 影像B | 标签 | FE-EF | FC-Siam- conc | FC-Siam- diff | BIT | SNUNet | FTUNet |
|---|---|---|---|---|---|---|---|---|---|
| a | ![]() | ||||||||
| b | ![]() | ||||||||
| c | ![]() | ||||||||
| d | ![]() | ||||||||
| e | ![]() | ||||||||
| 图例 | ![]() | ||||||||
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
宋嘉鑫, 李轶鲲, 杨树文, 等. 基于后验概率空间变化向量分析的NSCT高分辨率遥感影像变化检测[J]. 自然资源遥感, 2024, 36(3):128-136.doi:10.6046/zrzyyg.2023079.
|
| [18] |
柴佳兴, 张云生, 杨振, 等. 联合无人机影像生成DSM和DOM的多层次建筑物变化检测[J]. 自然资源遥感, 2024, 36(2):80-88.doi:10.6046/zrzyyg.2023001.
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
/
| 〈 |
|
〉 |