建筑物作为人类生产活动的重要载体，准确快速地提取建筑物可在自然资源管理领域发挥重要作用。基于卷积神经网络（convolutional neural network, CNN）在遥感影像建筑物提取方面取得了重大进展，但构建的网络模型在特征提取和特征融合方面仍有待优化。因此，提出了一种坐标注意力及卷积增强的全尺度融合建筑物提取网络（coordinate attention and convolutional enhanced full-scale fusion building extraction network, CCFNet）。所构建的模型由坐标注意力及卷积增强的残差编码器和全尺度融合解码器组成。编码器使用坐标注意力构建通道间的依赖关系并捕获的全局信息，其使用的非对称卷积增强地物边缘特征提取，并对旋转、翻转扭曲及纵横比不均匀的地物有更强的鲁棒性。解码器使用的全尺度融合方法则有助于建筑物的重建。在中国典型城市建筑物实例数据集实验结果表明，相比于其他建筑物提取网络，本文构建的CCFNet模型在Accuracy、F₁、IOU和MIOU共4种分割评价指标分别取得了93.84%、84.08%、72.53%和82.59%的最优实验精度。结果表明，该模型能够有效地提取建筑物区域。

Buildings are important carriers of human production activities. Accurate and fast extraction of building areas can play an important role in the field of natural resource management. Although significant progress has been made in building extraction from remote sensing images based on CNN(convolutional neural network), the constructed network model still needs to be optimized in feature extraction and feature fusion. Therefore, a coordinate attention and CCFNet(convolutional enhanced full-scale fusion building extraction network) was proposed. The constructed model consists of a residual encoder enhanced by coordinate attention and convolution and a full-scale fusion decoder. Coordinate attention was used in the encoder to build inter-channel dependencies and capture global information. The asymmetric convolution was used to enhance the edge feature extraction of ground objects, and it is more robust to rotation, flip distortion and uneven aspect ratio of ground objects. The full-scale fusion method used in the decoder helps to reconstruct the buildings. The experimental results on the dataset of typical Chinese city buildings show that compared with other building extraction networks, The CCFNet model constructed in this paper achieves the best experimental Accuracy of 93.84%, 84.08%, 72.53% and 82.59% in the four segmentation evaluation indicators of accuracy, F₁, IOU and MIOU, respectively. Experimental results show that the model can effectively extract building regions.