基于超像元词包特征和主动学习的高分遥感影像变化检测

为解决高分辨率遥感影像变化检测中存在底层特征缺乏语义信息、像元级的检测结果存在“椒盐”现象以及监督分类中样本标注自动化程度较低,本文提出一种基于超像元词包特征和主动学习的变化检测方法。首先采用熵率分割算法获取叠加影像的超像元对象;其次提取两期影像像元点对间的邻近相关影像特征（相关度、斜率和截距）和顾及邻域的纹理变化强度特征（均值、方差、同质性和相异性）,经线性组合作为像元点对的底层特征;然后基于像元点对底层特征利用BOW模型构建超像元词包特征,并采用一种改进标注策略的主动学习方法从无标记样本池中优选信息量较大的样本,且自动标注样本类别;最后训练分类器模型完成变化检测。通过选用2组不同地区的GF-2影像和Worldview-Ⅱ影像作为数据源进行实验,实验结果中2组数据集的F1分数分别为0.8714、0.8554,正确率分别为0.9148、0.9022,漏检率分别为0.1681、0.1868,误检率分别为0.0852、0.0978。结果表明,该法能有效识别变化区域、提高变化检测精度。此外,传统主动学习方法与改进标注策略的主动学习方法的学习曲线对比显示,改进的标注策略可在较低精度损失下,有效提高样本标注自动化程度。

High-Resolution Remote Sensing Imagery Change Detection based on Super-Pixel BOW Features and Active Learning

The following problems exist in the change detection of high-resolution remote sensing imagery: lack of the semantic information of low-level features, the "salt and pepper" phenomenon in the detection results based on pixel-level methods, and the low degree of sample labeling automation in supervised classification. In this paper, we proposed a change detection method based on super-pixel Bag-of-Words features and active learning. Firstly, we used the entropy rate segmentation algorithm to obtain the segmentation objects of superimposed images. Secondly, we extracted the features of the Neighborhood Correlation Images (correlation, slope, and intercept) and the change intensity features of texture (mean value, variance, homogeneity, and dissimilarity) while considering neighborhood context information between pixel pairs of the studied two phases of images, and then combined them as the low-level features of pixel pairs. Followingly, based on these low-level features, we constructed the expression of Bag-of-Words features in the super-pixel regions by the Bag-of-Words (BOW) model, and we adopted an improved annotation strategy to annotate automatically the samples with large information from the unlabeled sample pool. Finally, we conducted the change detection using the trained classification model. By choosing two groups in different parts of GF-2 imagery and Worldview-Ⅱ imagery as a data source for experiments, the experimental results show that the F1 scores of the two groups of data sets are 0.8714 and 0.8554, the precision is 0.9148 and 0.9022, the missed detection rate is 0.1681 and 0.1868, and the false detection rate is 0.0852 and 0.0978, respectively. The results demonstrate that our proposed method can effectively ide.pngy the variation area, improve the accuracy of change detection. In addition, the comparison of the learning curves between the traditional active learning method and the active learning method with improved annotation strategy shows that the improved annotation strategy can effectively improve the automation degree of sample annotation at a lower precision loss.