Abstract: Due to the high similarity between background and buildings in high spatial resolution remote sensing images, which makes it difficult for the network to take into account buildings of different sizes, the pixels in the building boundary region are confused with the background, and the building boundaries are easily missed. In order to solve the above problems, the building extraction algorithm (UUNet) for remote sensing images fusing partial convolution and residual refinement is proposed. Using U-Net as the baseline network, firstly, this paper improves the encoder. It adds two Conv4×4 at the front end of the encoder to expand the sensing field at the beginning and capture more remote sensing image feature information. It utilizes the PC module constructed by partial convolution (PConv3×3) to enhance the ability of the encoder to extract multi-scale building features, and downsamples twice with Conv2×2 to reduce the loss of building feature information. Secondly, this paper reduces the number of parameters. It crops the three-layer structure of the U-Net network decoder to a UUNet network decoder. Lastly, it adds an improved residual refinement module. It constructs a U-shaped residual refinement module cropped to a three-layer structure at the output of the decoder, to further purify the rough building feature maps output from the decoder, so as to make the edge information of the buildings clearer. Decoder is jump-connected to the encoder of the U-shaped residual refinement module to preserve the initial features, and SimAM is embedded in the refinement module to improve the building focus, optimize the network to improve the boundary blurring, and enhance the quality of target boundary extraction. In the ablation experiment conducted on the Satellite dataset II (East Asia), UUNet shows improvements over U-Net, with IoUBuilding, IoUBackground, F1, OA and mIoU increased by 2.78 percentage points, 0.12 percentage points, 1.91 percentage points, 0.19 percentage points and 1.45 percentage points, respectively, indicating that UUNet outperforms the baseline network. Furthermore, comparative experiments on both Satellite dataset II (East Asia) and WHU dataset demonstrate that UUNet performs better than existing mainstream algorithms, significantly enhancing building extraction in high-resolution remote sensing images.

Key words: high resolution remote sensing imagery, building extraction, boundary smoothing, multiscale features, U-Net, Partial convolution

摘要： 由于高空间分辨率遥感图像中背景与建筑物对象的相似度高，导致网络难以兼顾不同大小的建筑物，建筑边界区域的像素与背景混淆，建筑边界很容易被漏检。为解决上述问题，提出融合Partial卷积与残差细化的遥感影像建筑物提取算法（UUNet）。以U-Net为基线网络，首先，改进编码器。在编码器前端加入两个Conv4×4，在最初扩大感受野，捕捉更多遥感影像特征信息，利用Partial卷积（PConv3×3）构造的PC模块，增强编码器提取多尺度建筑物特征的能力，用Conv2×2进行两倍下采样，减少建筑物特征信息丢失。其次，减少参数量。裁剪U-Net网络解码器三层结构为UUNet网络解码器。最后，增加改进的残差细化模块。在解码器输出端构造裁剪到三层结构的U型残差细化模块，对解码器输出的粗糙建筑物特征图进行进一步提纯，使建筑物边缘信息更加清晰，网络解码器与U型残差细化模块编码器进行跳跃连接，保留最初特征，将SimAM嵌入细化模块中，提高建筑物关注度，优化网络改善边界模糊，提升目标边界提取质量。在Satellite dataset Ⅱ（East Asia）数据集上进行消融实验，UUNet比U-Net的IoUBuilding、IoUBackground、F1、OA和MIoU分别提高2.78个百分点、0.12个百分点、1.91个百分点、0.19个百分点、1.45个百分点，表明UUNet网络优于基线网络；在Satellite dataset Ⅱ （East Asia）数据集和WHU数据集上做对比实验，UUNet相较于现有的主流算法更优，能够显著地提升高分辨率遥感影像中建筑物提取的效果。

关键词: 高分辨率遥感影像, 建筑物提取, 边界平滑, 多尺度特征, U-Net, Partial卷积