基于Transformer和小波变换卷积的SAR影像无监督变化检测方法OA
An unsupervised change detection method for synthetic aperture radar images based on Transformer and wavelet transform convolution
无监督变化检测是合成孔径雷达(synthetic aperture radar,SAR)影像信息提取中的研究热点,近年来受到广泛关注.尽管已有研究在该领域取得一定进展,但现有方法大多仅关注空间域特征的利用,在空间-频率双域特征融合利用方面的探索研究较少.因此,文章设计了一种基于 Transformer 和小波变换卷积的 UNet 变化检测模型,通过联合对数比值法和模糊 C 均值法获取伪标签,实现 SAR 影像变化信息的无监督提取.该模型的编码部分包含 2个分支,空间域分支利用 Transformer 提取空间域全局特征,并与卷积网络提取的空间域局部特征相融合;频率域分支利用小波变换卷积提取多尺度频率域特征,增强模型对承载影像主体信息的低频分量和承载边界细节的高频分量的响应能力.在解码器部分实现多尺度跨域互补信息在 UNet 架构中的有机融合,并在 2 个数据集上进行实验来验证所提方法的有效性.结果表明:与对比方法中最好的变化检测结果相比,该文方法在 2 个数据集上的 F1分数分别提高了 0.030 和 0.017,Kappa 系数分别提高了 0.032 和 0.018,有效提高了变化检测结果的可靠性.
Unsupervised change detection represents a hot research topic in information extraction from synthetic aperture radar(SAR)images,having attracted wide attention in recent years.Although existing studies have made certain advances in this field,available methods for unsupervised change detection primarily focus on the utilization of spatial-domain features,leading to limited explorations into the fusion of spatial-frequency dual-domain features.In this context,this study designed a UNet model for change detection based on Transformer and wavelet transform convolution.Using pseudo-labels obtained by combining log-ratio operation with fuzzy C-means clustering,the UNet model enables the unsupervised information extraction of changes from SAR images.The encoder of the model consists of two branches:spatial-and frequency-domain branches.Among these,the spatial-domain branch extracts global spatial features using Transformer and fuses them with local spatial features extracted using the convolutional network.The frequency-domain branch extracts multi-scale frequency-domain features using wavelet transform convolution,enhancing the model's responses to both low-frequency components carrying the main information of images and high-frequency components carrying boundary details.Finally,the decoder facilitates the organic fusion of multi-scale cross-domain complementary information within the UNet architecture.The effectiveness of the proposed method was verified using experiments on two datasets.The results indicate that,compared to the most accurate change detection results derived using the comparative methods,the detection results determined using the proposed method showed increases of 0.030 and 0.017 in the F1 score and increases of 0.032 and 0.018 in the Kappa coefficient on the two datasets.Therefore,the proposed method can effectively enhance the reliability of change detection results.
张懂庆;张祥;王银伟;张宇;庄会富
甘肃省地质矿产勘查开发局第一地质矿产勘查院,天水 741020||甘肃省矿产资源综合勘查利用与保护工程研究中心,天水 741020||甘肃省地质矿产勘查开发局第一地质矿产勘查院环境损害司法鉴定所,天水 741020自然资源部国土卫星遥感应用中心,北京 100048甘肃省地质矿产勘查开发局第一地质矿产勘查院,天水 741020||甘肃省矿产资源综合勘查利用与保护工程研究中心,天水 741020中国矿业大学自然资源部国土环境与灾害监测重点实验室,徐州 221116中国矿业大学自然资源部国土环境与灾害监测重点实验室,徐州 221116
信息技术与安全科学
无监督变化检测SAR影像Transformer小波变换卷积特征融合
unsupervised change detectionsynthetic aperture radar(SAR)imageTransformerwavelet trans-form convolutionfeature fusion
《自然资源遥感》 2026 (4)
11-20,10
甘肃省地矿局一勘院地质工匠培育计划项目"基于空天地一体的矿山变形监测方法研究"(编号:YD202413)和国家自然科学基金项目"无标注样本下光学与SAR异源影像自监督学习变化检测方法研究"(编号:42401539)共同资助.
评论