Remote Sensing for Natural Resources >
A method combining the siamese inverted residual structure with self-attention enhancement for change detection based on remote sensing images
Received date: 2023-12-15
Revised date: 2024-03-22
Online published: 2026-06-03
Change detection based on remote sensing images holds significant application potential in land source survey updating, and urban development monitoring and planning. Concerning the challenges of change detection based on remote sensing images in practical applications, this study proposed a lightweight change detection method combining the siameseinverted residual structure with self-attention enhancement. Instead of the traditional convolutional neural network structure, the siamese improved inverted residual structure was used as the backbone network to fully extract the feature information and significantly reduce the network complexity. The self-attention enhancement module was employed to improve the network's ability to pay attention to global information. Edge weights were added to the loss function to precisely optimize the details of the extraction results. The multilevel hopping residual connections were applied to fully integrate the global and local features. Finally, the performance of the proposed method was tested on the public and prepared datasets of remote sensing images for change detection, respectively. The results indicate that compared to other change detection methods, the proposed method significantly reduced network parameters and computational complexity while maintaining the detection accuracy, contributing to lightweight models of change detection based on remote sensing images.
ZHANG Qiao , CAO Zhicheng , SHEN Yang , WANG Zhoufeng , WANG Chengwu , XU Jiaxin . A method combining the siamese inverted residual structure with self-attention enhancement for change detection based on remote sensing images[J]. Remote Sensing for Natural Resources, 2025 , 37(3) : 85 -94 . DOI: 10.6046/zrzyyg.2023388
表1 不同模型在LEVIR-CD和SAMPLE-CD数据集上的精度对比Tab.1 Comparison of accuracy evaluation of different model on LEVIR-CD and SAMPLE-CD(%) |
| 模型 | LEVIR-CD | SAMPLE-CD | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| OA | Precision | Recall | F1 | IoU | OA | Precision | Recall | F1 | IoU | |
| FC-EF | 98.39 | 86.91 | 80.17 | 83.40 | 71.35 | 88.54 | 72.53 | 64.85 | 59.34 | 48.63 |
| FC-Siam-diff | 98.67 | 89.53 | 83.31 | 86.31 | 75.92 | 86.72 | 64.89 | 66.92 | 63.81 | 52.26 |
| FC-Siam-conc | 98.49 | 91.99 | 76.77 | 83.69 | 71.96 | 87.19 | 68.38 | 68.31 | 61.48 | 54.11 |
| BIT | 98.92 | 89.24 | 89.37 | 89.31 | 80.68 | 90.65 | 83.15 | 81.51 | 82.29 | 71.74 |
| Changeformer | 99.04 | 92.05 | 88.80 | 90.40 | 82.48 | 89.25 | 80.86 | 77.51 | 79.02 | 67.78 |
| Siam-MViT | 98.96 | 91.11 | 92.16 | 91.04 | 84.35 | 91.15 | 83.98 | 82.80 | 83.37 | 73.13 |
表2 LEVIR-CD数据集的变化检测可视化结果Tab.2 Visualization results of change detection on LEVIR-CD dataset |
| 序号 | 前时相 | 后时相 | 标签 | FC-EF | FC-Siam-diff | FC-Siam-conc | BIT | Changefomer | Siam-MViT |
|---|---|---|---|---|---|---|---|---|---|
| 1 | ![]() | ||||||||
| 2 | ![]() | ||||||||
| 3 | ![]() | ||||||||
| 4 | ![]() | ||||||||
| 5 | ![]() | ||||||||
| 6 | ![]() | ||||||||
| 图例 | ![]() |
表3 SAMPLE-CD数据集的变化检测可视化结果Tab.3 Visualization results of change detection on SAMPLE-CD dataset |
| 序号 | 前时相 | 后时相 | 标签 | FC-EF | FC-Siam-diff | FC-Siam-conc | BIT | Changefomer | Siam-MViT |
|---|---|---|---|---|---|---|---|---|---|
| 1 | ![]() | ||||||||
| 2 | ![]() | ||||||||
| 3 | ![]() | ||||||||
| 4 | ![]() | ||||||||
| 5 | ![]() | ||||||||
| 6 | ![]() | ||||||||
| 图例 | ![]() |
表4 模型计算效率统计Tab.4 Model calculation efficiency statistics |
| 模型 | Param/106个 | GFLOPS | 预测用时/ (ms/组) |
|---|---|---|---|
| FC-EF | 1.35 | 3.58 | 6.37 |
| FC-Siam-diff | 1.35 | 4.73 | 8.01 |
| FC-Siam-conc | 1.55 | 5.33 | 7.84 |
| BIT | 12.41 | 10.65 | 16.32 |
| Changeformer | 41.03 | 202.79 | 28.97 |
| Siam-MViT | 0.82 | 3.36 | 5.92 |
表5 消融实验结果Tab.5 Ablation experimental results |
| 模型 | 特征强化模块 | 改进型倒残差模块 | 边缘损失 函数 | LEVIR-CD F1/% | SAMPLE-CD F1/% | Param/106个 | GFLOPS |
|---|---|---|---|---|---|---|---|
| Siam-MViT | √ | √ | √ | 91.04 | 83.37 | 0.82 | 3.36 |
| A | √ | √ | × | 90.40 | 81.34 | 0.82 | 3.36 |
| B | √ | × | × | 89.26 | 78.63 | 0.58 | 1.70 |
| C | × | × | × | 88.63 | 78.36 | 0.15 | 1.03 |
| [1] |
李德仁. 利用遥感影像进行变化检测[J]. 武汉大学学报(信息科学版), 2003, 28(s1):7-12.
|
| [2] |
|
| [3] |
|
| [4] |
眭海刚, 冯文卿, 李文卓, 等. 多时相遥感影像变化检测方法综述[J]. 武汉大学学报(信息科学版), 2018, 43(12):1885-1898.
|
| [5] |
任秋如, 杨文忠, 汪传建, 等. 遥感影像变化检测综述[J]. 计算机应用, 2021, 41(8):2294-2305.
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
刘春亭, 冯权泷, 刘建涛, 等. DeepLabv3+语义分割模型的济南市防尘绿网提取及时空变化分析[J]. 遥感学报, 2022, 26(12):2518-2530.
|
| [12] |
沙苗苗, 李宇, 李安. 改进Faster R-CNN的遥感图像多尺度飞机目标检测[J]. 遥感学报, 2022, 26(8):1624-1635.
|
| [13] |
|
| [14] |
|
| [15] |
袁洲, 郭海涛, 卢俊, 等. 融合UNet++网络和注意力机制的高分辨率遥感影像变化检测算法[J]. 测绘科学技术学报, 2021, 38(2):155-159.
|
| [16] |
吴纹辉, 慎利, 董新丰, 等. 面向高分辨率遥感影像建筑物变化检测的边缘感知网络[J]. 地理与地理信息科学, 2021, 37(3):21-28.
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
田永林, 王雨桐, 王建功, 等. 视觉Transformer研究的关键问题:现状及展望[J]. 自动化学报, 2022, 48(4):957-979.
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
王振庆, 周艺, 王世新, 等. IEU-Net高分辨率遥感影像房屋建筑物提取[J]. 遥感学报, 2021, 25(11):2245-2254.
|
| [33] |
王子浩, 李轶鲲, 李小军, 等. 基于空间模糊C均值聚类和贝叶斯网络的抗噪声遥感图像变化检测[J]. 自然资源遥感, 2023, 35(4):96-104.doi: 10.6046/zrzyyg.2022260.
|
/
| 〈 |
|
〉 |