基于 UNet、UNet++、Attention-UNet 网络的遥感影像建筑物提取研究

文章选择 UNet、UNet++ 和 Attention-UNet 3 种卷积神经网络，研究该系列网络在建筑物提取中的性能差异。基于 WHU 建筑物数据集，通过实验和定量评估深入分析上述网络模型在遥感影像建筑物提取中的有效性和适用性。结果表明：UNet++ 的提取精度最高，其次是 Attention-UNet、UNet，分别为 0.987、0.983、0.979；UNet++ 的边缘提取能力和建筑物正确分割比例优于其他 2 种网络；针对中小型建筑物，Attention-UNet 的提取效果显著，但易出现非建筑物误提的情况，UNet++ 的提取效果良好，偶尔出现漏提现象，UNet 的提取效果欠佳。

This study selects three convolutional neural networks(CNNs), namely UNet, UNet++, and Attention-UNet, aiming to investigate the performance differences of this network series in building extraction. Based on the WHU building dataset, through experiments and quantitative evaluations, this research explores the effectiveness and applicability of the aforementioned network models in remote sensing image building extraction. The experimental results indicate that UNet++ achieves the highest extraction accuracy, followed by Attention-UNet and UNet, with accuracies of 0.987, 0.983, and 0.979, respectively. UNet++ exhibits superior edge extraction capabilities and correct building segmentation ratios compared to the other two networks. For small to medium-sized buildings, Attention-UNet demonstrates significant extraction effectiveness, but it is prone to false positives. UNet++ performs well with occasional instances of missed extractions, while the extraction effectiveness of UNet is subpar.