用于遥感影像建筑物变化检测的多尺度交叉对偶注意力网络
张建兵(1974— ),男,湖北洪湖人,博士,讲师,硕士生导师,主要研究方向为网格GIS、分布式计算、数字地球、空间无线信息搜索引擎、深度学习。E-mail: zhangjb@cup.edu.cn |
收稿日期: 2023-07-28
修回日期: 2023-10-06
网络出版日期: 2023-12-05
基金资助
国家自然科学基金项目(61972414)
国家重点研发计划项目(2016YFC0303707)
Multi-scale Cross Dual Attention Network for Building Change Detection in Remote Sensing Images
Received date: 2023-07-28
Revised date: 2023-10-06
Online published: 2023-12-05
Supported by
National Natural Science Foundation of China(61972414)
National Key Research and Development Program of China(2016YFC0303707)
本文提出了一种用于遥感影像变化检测的多尺度交叉对偶注意力网络MSCDAN(Multi-Scale Cross Dual Attention Network),该神经网络模型利用改进的ResNet18网络提取原始遥感影像中的多尺度低级特征,并通过结合交叉注意力和对偶注意力2种注意力机制的CDA(Cross Dual Attention)模块提取注意力特征信息。CDA模块可以加强输入数据中不同视角或特征图之间的关联、融合时空信息、捕捉地表变化的时间序列特征、识别周期性变化和持续性变化等时序相关的变化模式。MSCDAN模型通过全转置卷积上采样模块FTCUM(Fully Transposed Convolution Upsampling Module)对特征图中的每个点进行局部的特征融合,由神经网络判别变化边界,避免了像双线性插值等传统方法带来的模糊和锯齿等问题,且实现了端到端的训练和优化,从而能够更好地适应遥感影像变化检测任务需求。相较于现有主流方法DTCDSCN(Dual-Task Constrained Deep Siamese Convolutional Network),本文提出的方法在DSIFN数据集上的准确度提高了5.13%,在WHU-CD数据集上的准确度提高了1.3%。同时,本文方法在这两个数据集上的表现也优于现有方法ChangeNet以及LamboiseNet,在CDD数据集上的表现优于改进DeepLabv3+和SRCD-Net。这些结果表明本文方法在不同数据集上均具有良好的性能,对进一步研究遥感影像变化检测具有重要参考价值。
张建兵 , 严泽枭 , 马淑芳 . 用于遥感影像建筑物变化检测的多尺度交叉对偶注意力网络[J]. 地球信息科学学报, 2023 , 25(12) : 2487 -2500 . DOI: 10.12082/dqxxkx.2023.230432
Remote sensing image change detection is a crucial technique that utilizes remote sensing technology to analyze and compare image data captured at different time periods or scenes. In practice, features at varying scales encompass diverse representation ranges, enabling the extraction of more comprehensive and detailed information. This paper proposes a Multi-Scale Cross Dual Attention Network (MSCDAN) method for building change detection in remote sensing images using the multi-scale Cross Dual Attention (CDA) mechanism and residual convolution neural network architecture. The proposed method leverages the characteristics of a residual network to extract change features of different dimensions from remote sensing images. For each feature dimension, a CDA module is created, which utilizes both cross attention and dual attention mechanisms. It combines spatiotemporal information to capture time-series features of surface changes and identifies time-series related change patterns, such as periodic and persistent changes. In this way, the multi-scale CDA module enhances the correlation between different perspectives or feature maps within the input data, which facilitates the exchange and fusion of information in multiple dimensions and enhances the model capability for complex change scenes, leading to improved change detection performance. A Fully Transposed Convolutional Upsampling Module (FTCUM) is introduced to perform local feature fusion for each point in the feature map, and the change boundary is identified by the neural network. This avoids the problems of blurring and jaggedness brought by traditional methods like bilinear interpolation and allows for end-to-end training and optimization, making the method more effective in meeting the requirements of change detection tasks. Extensive experiments are conducted on two benchmark datasets, namely WHU-CD and DSIFN, to evaluate the performance of the proposed method. Compared to the mainstream method, i.e., DTCDSCN (Dual-Task Constrained Deep Siamese Convolutional Network), our proposed method increases the accuracy by 5.13% on the DSIFN dataset and by 1.3% on the WHU-CD dataset. Additionally, for other exiting methods, the proposed method is also better than the ChangeNet and LamboiseNet on the three datasets and outperforms the improved DeepLabv3+ and SRCD-Net on the CDD Dataset. These exceptional findings across various datasets confirm the effectiveness of the proposed method in detecting changes in remote sensing images. Through the application of residual networks and attention mechanisms, our approach achieves superior results in intricate scenarios. This study shows that our proposed method performs remarkably well on various datasets. It serves as a reference for further comprehensive research on remote sensing image change detection using multi-scale cross-pairwise attention networks.
表1 不同方法在LEVIR-CD、WHU-CD和DSIFN数据集上的精确率(Pre.)、召回率(Rec.)、IoU、 分数与准确率(OA)对比Tab. 1 Comparison of Pre., Rec., IoU, score and OA of different methods on LEVIR-CD, WHU-CD and DSIFN dataset (%) |
模型 | LEVIR-CD | WHU-CD | DSIFN | |||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Pre. | Rec. | IoU | F1 | OA | Pre. | Rec. | IoU | F1 | OA | Pre. | Rec. | IoU | F1 | OA | ||||
ChangeNet | 69.81 | 50.15 | 68.71 | 78.30 | 96.62 | 81.60 | 58.66 | 72.15 | 83.45 | 97.51 | 78.16 | 85.69 | 71.45 | 81.75 | 86.20 | |||
LamboiseNet | 80.51 | 94.14 | 77.37 | 85.66 | 97.12 | 65.24 | 83.22 | 60.68 | 69.99 | 93.54 | 77.54 | 78.35 | 64.32 | 77.83 | 79.93 | |||
DTCDSCN | 88.53 | 86.83 | 78.05 | 87.67 | 98.77 | 63.92 | 82.30 | 56.19 | 71.95 | 97.42 | 53.87 | 77.99 | 46.76 | 63.72 | 84.91 | |||
MSCDAN | 89.68 | 84.49 | 86.17 | 89.30 | 98.81 | 88.82 | 81.41 | 77.52 | 84.95 | 98.69 | 85.61 | 84.37 | 78.66 | 84.99 | 89.27 |
表2 不同方法在CDD数据集上的精确率(Pre.)、召回率(Rec.)、IoU、 分数与准确率(OA)对比Tab. 2 Comparison of Pre., Rec., IoU, score and OA of different methods on CDD dataset (%) |
模型 | Pre. | Rec. | IoU | F1 | OA |
---|---|---|---|---|---|
改进Deep Labv3+ | 87.30 | 90.20 | — | 88.40 | 96.40 |
SRCD-Net | 92.55 | 93.34 | 86.81 | 92.94 | — |
MSCDAN | 93.68 | 92.80 | 88.74 | 93.18 | 97.92 |
[1] |
张良培, 武辰. 多时相遥感影像变化检测的现状与展望[J]. 测绘学报, 2017, 46(10):1447-1459.
[
|
[2] |
张钰屏, 刘籽琳, 陈天其. 遥感影像变化检测研究进展[J]. 电脑与信息技术, 2023, 31(2):15-18.
[
|
[3] |
杨洋, 李轶鲲, 杨树文, 等. 基于空间模糊C均值聚类和贝叶斯网络的高分一号遥感影像变化检测[J]. 测绘与空间地理信息, 2023, 46(4):34-37,42.
[
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
黄维, 黄进良, 王立辉, 等. 基于PCA的变化向量分析法遥感影像变化检测[J]. 国土资源遥感, 2016, 28(1):22-27.
[
|
[10] |
|
[11] |
黄晓君, 颉耀文, 卫娇娇, 等. 基于变化检测-CART决策树模式自动识别沙漠化信息[J]. 灾害学, 2017, 32(1):36-42.
[
|
[12] |
田青林, 秦凯, 陈俊, 等. 基于注意力金字塔网络的航空影像建筑物变化检测[J]. 光学学报, 2020, 40(21):2110002.
[
|
[13] |
|
[14] |
|
[15] |
|
[16] |
刘英, 何雪, 李单阳, 等. 注意力引导全尺度连接网络的高分辨率影像变化检测[J]. 遥感学报, 2022.
[
|
[17] |
赵祥, 王涛, 张艳, 等. 基于改进DeepLabv3+孪生网络的遥感影像变化检测方法[J]. 地球信息科学学报, 2022, 24(8):1604-1616.
[
|
[18] |
陈婕, 刘纪平, 徐胜华. 增强边缘信息的全卷积神经网络遥感影像建筑物变化检测[J]. 测绘通报, 2023(6):61-67.
[
|
[19] |
高建文, 管海燕, 彭代锋, 等. 基于局部-全局语义特征增强的遥感影像变化检测网络模型[J]. 地球信息科学学报, 2023, 25(3):625-637.
[
|
[20] |
宋文宣, 彭代锋. 一种改进全卷积网络的遥感影像变化检测[J]. 遥感信息, 2022, 37(6):130-136.
[
|
[21] |
代云锋, 刘丽娜. 结合改进混合卷积模型的遥感影像变化检测[J]. 遥感信息, 2022, 37(6):53-59.
[
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
边继龙, 王厚博, 李金凤. 基于多尺度注意力网络的立体匹配方法[J]. 北京邮电大学学报, 2021, 44(3):27-34.
[
|
[27] |
|
[28] |
|
[29] |
刘春娟, 乔泽, 闫浩文, 等. 基于多尺度互注意力的遥感图像语义分割网络[J]. 浙江大学学报(工学版), 2023, 57(7):1335-1344.
[
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
|
/
〈 | 〉 |