顾及边缘的多时相SAR图像半监督建筑区提取

Edge-Aware Semi-Supervised Built-up Area Extraction Using Multi-Temporal Synthetic Aperture Radar Images

针对合成孔径雷达(synthetic aperture radar,SAR)图像中建筑区域难以辨识与标注的问题,提出一种结合改进的伪标签技术和边缘增强策略的半监督建筑区提取新方法.首先,引入同一位置、不同时相的SAR图像作为自然数据增强手段,并通过多个不同时相图像的预测结果投票确定伪标签;其次,设计一种边缘增强辅助模块,通过特征图变形以修正建筑区主体特征,辅以跳跃连接改进边缘特征,并针对主体和边缘特征进行分离式监督;此外,构建一个包含2种传感器和2个城市区域的多时相SAR图像建筑区提取数据集,含1 000幅带标注图像和800组无标注时序图像,并基于该数据集进行实验验证.实验表明,在所构建测试集上,基线方法使用全量数据训练后交并比(intersection over union,IoU)为63.43%,而所提方法在使用10%和全量数据时IoU分别为63.46%和68.24%,仅利用10%的标注数据即可达到基线方法使用全量标注数据训练的精度.

To address the challenges of identifying and annotating built-up areas in synthetic aperture radar(SAR)images,a novel semi-supervised method for extracting built-up areas that combined improved pseudo-labeling techniques with an edge enhancement strategy was proposed.Initially,SAR images from the same location but at different time were introduced as a natural data augmentation method,and the pseudo-labels were determined by voting based on the prediction results of multi-temporal images.Subsequently,an edge-enhancement auxiliary module was designed,which corrected the body features of the built-up areas through feature map warping and improved edge features with skip connections.Separate supervision for the body and edge features was performed.Moreover,a dataset for extracting built-up areas in multi-temporal SAR images,which included two types of sensors and two urban areas,was constructed.This dataset contains 1,000 annotated images and 800 groups of unlabeled temporal images.Experimental validations based on this dataset have demonstrated that on the constructed test set,the baseline method trained with full data achieves an intersection over union(IoU) of63.43%,while the proposed method reaches an IoU of 63.46% and 68.24% when using 10% and full data,respectively.Remarkably,using only 10% of the annotated data,the proposed method can achieve the precision that the baseline method has obtained with full annotated data.