基于弱监督分类的城市绿网覆盖用地语义分割方法

A Semantic Segmentation Method for Urban Green Plastic based on Weakly Supervised Classification

摘要:

城市建设过程中,裸土的地表必须要由绿网覆盖,以减轻扬尘污染,所以绿网覆盖用地一定程度上表征了城市建设进程,也能够帮助识别违章建筑。当前基于深度学习的高分辨率遥感影像地表覆盖分类方法已经取得较好效果,但对大量标注样本的需求限制了其进一步广泛应用。本文针对城市绿网覆盖用地的识别需求,提出了一种基于弱监督分类的提取方法；该方法只需要获取图像级标签,然后通过类激活图获取粗略像素级标签,再经过标签改正获取精细的像素级标签,最后提出一种边界感知的联合损失函数实现绿网覆盖用地分割。经过谷歌地图不同分辨率的遥感图像验证,绿网用地的场景识别精度优于87%,语义分割精度优于83%,证明本文方法能够有效实现城市绿网覆盖用地的提取。

Abstract:

During urban construction, exposed soil surfaces must be covered with green plastic to mitigate dust pollution. Consequently, green plastic can characterize the progress of urban development and aid in identifying illegal structures to some extent. While deep learning-based high-resolution remote sensing image land cover classification methods have achieved promising results, the requirement for large annotated datasets limits their broader application. This paper addresses the identification needs for urban green plastic by proposing a weak supervision classification-based extraction method. This approach requires only image-level labels. It then obtains coarse pixel-level labels through class activation maps, refines these into precise pixel-level labels via label correction, and finally proposes a boundary-aware joint loss function to achieve segmentation of green plastic. Validations using Google images remote sensing imagery at various resolutions demonstrated scene recognition accuracy exceeding 87% and semantic segmentation accuracy surpassing 83% for green plastic. These results show that the proposed method effectively extracts urban green plastic.