基于DS证据融合生成伪标签的SAR图像变化检测

SAR image change detection based on pseudo labels generated from DS evidence fusion

合成孔径雷达(Synthetic Aperture Radar,SAR)图像变化检测是一种检测两张SAR图像中变化区域的技术。在基于神经网络的无监督变化检测方法中,伪标签的质量影响到了检测结果的精度。为了提高精度,提出一种基于登普斯特-沙弗(Dempster-Shafer,DS)证据融合理论生成伪标签的方法,再结合多层次模糊C类均值聚类和卷积小波神经网络实现变化检测。该方法通过对三种不同差异图聚类生成的伪标签进行DS证据融合得到融合伪标签,使训练样本更加准确,然后利用网络对像素进行分类,得到变化检测结果。经过实验,该方法在渥太华、越南红河和黄河数据集上的检测精度分别达到了98.48%、97.95%和96.18%。

Synthetic aperture radar(SAR)image change detection technology aims to detect the change areas in two different SAR images.In such unsupervised methods based on neural network,the quality of pseudo labels restricts the accuracy of existing detection results.In order to improve the accuracy,we propose a method based on DS evidence fusion in order to generate pseudo labels and then we use Hierarchical Fuzzy C-Means Clustering and Convolutional-Wavelet Neutral Network to realize change detection.This method obtains fusion pseudo labels by DS evidence fusion on three different pseudo labels clustered by three difference Images,which makes the training samples more accurate,and then uses network to obtain the change detection results.Experiments on the Ottawa,Vietnam Red River and Yellow River datasets show the change detection accuracy of our method reaches 98.48%,97.95 and 96.18 respectively.