基于脚点热度图和双向连接图的遥感影像建筑物提取方法

目前，大多数基于深度学习的遥感影像建筑物提取方法采用语义分割的方式，对遥感影像进行二分类预测。然而，该类方法没有考虑建筑物的几何特性，难以进行精确提取。为了更精确地提取建筑物，该文引入建筑物的几何信息，提出一种基于脚点热度图和双向连接图的建筑物轮廓提取方法。该方法为一种多分支的深度卷积网络，分别对建筑物的脚点以及脚点间的连接性进行预测。在其中一个分支中，预测建筑物的脚点热度图，并用非极大抑制算法得到建筑物的脚点像素坐标；利用另外两个分支预测脚点之间的正向连通性和反向连通性，并通过这种双向连接图对脚点间是否具有连接性进行判断，在将具有连通性的脚点进行连接后，可得到最终的建筑物轮廓。该文算法在Buildings2Vec数据集上进行了验证，结果表明该方法在遥感影像建筑物提取中具有一定的优越性。

Building Extraction from Satellite Imagery Based on Footprint Map and Bidirectional Connection Map

Most state-of-the-art building extraction from satellite imagery are based on binary segmentation. However, the geographic information has not been considered in these methods, thus, it is difficult to extract building accurately. To consider fully the geographic information on feature extraction, a building extraction convolutional neural network based on footprint map and bidirectional connection is proposed. The proposed method is a multi-branch network, which is designed to predict the footprint and bidirectional connection map, respectively. This paper predicts the footprint heatmap of buildings and uses the Non-Maximum Suppression (NMS) algorithm to obtain the pixel coordinates. Another two branches are used to predict positive connectivity and negative connectivity between footprints. Each pair of nodes is connected according to the bidirectional connectivity map to obtain the final building outline. Experiments on the Buildings2Vec dataset demonstrate that the proposed method outperforms various previous work, which illustrate the superiority in building extraction from satellite imagery.