<p>Buildings are the main carriers of human activities, and their accurate detection is of great significance in the fields of urban planning, topographic map updating, map mapping and so on. Due to with its powerful feature learning ability, deep learning technique provides a&#xa0;possibility to realize the automatic detection of building roof with high accuracy and robustness. Aiming at the challenging issues that the fuzzy boundaries might appear in building roof detection, this paper proposes a&#xa0;boundary-aware graph convolutional network for the building roof detection from high-resolution remote sensing imagery, which consists of a&#xa0;backbone network module, a&#xa0;boundary segmentation module, a&#xa0;region segmentation module and a&#xa0;boundary-aware graph convolution module. More specifically, the backbone network module adopts the truncated Res2Net for efficiently extracting the rich semantic information. Following this, both the boundary segmentation module and region segmentation module are responsible for extracting boundary and region features using shallow and deep features, respectively. Afterwards, these boundary and region features are embedded into a&#xa0;graph convolution network through a&#xa0;novel graph representation to construct a&#xa0;boundary-aware and input-dependent adjacency matrix, thus realizing more accurate building roof detection. The quantitative and qualitative analyses were conducted on the Vaihingen dataset and Potsdam dataset of the International Society for Photogrammetry and Remote Sensing (ISPRS) benchmark. By the comparative analyses between the proposed method and some mainstream methods, we conclude that the proposed method is more accurate in detecting the roofs of buildings, and the optimal accuracies have reached 97.64 and 99.17%, respectively. Moreover, the visualization of the detection results intuitively demonstrates the accuracy of the prediction of the proposed method.</p>