结合无人机遥感数据和U-Net++模型的建筑物足迹提取方法  

A Building Footprint Extraction Method Combining UAV Remote SensingData and the U-Net++ Model

作　　者：曾飞 ZENG Fei(Guangzhou Transportation Design&Research Institute Co.,Ltd.,Guangzhou 511430,Guangdong,China)

机构地区：[1]广州市交通设计研究院有限公司,广东广州511430

出　　处：《广东交通职业技术学院学报》2025年第1期52-56,119,共6页Journal of Guangdong Communication Polytechnic

基　　金：广州市交通设计研究院有限公司项目(编号:BLX-SJY-2023-001);自然资源部大湾区地理环境监测重点实验室开放基金(编号:No.2019002)。

摘　　要：针对复杂城市场景下的建筑物足迹提取问题,使用了更适合于捕捉高分辨率遥感图像细节特征的语义分割U-Net++模型。为了让模型表现出良好的性能和降低模型过拟合化风险,在训练前期通过warmup逐渐提高学习率,当学习率达到特定的值后选用LambdaLR策略。融合了加权交叉熵损失函数,以此来解决场景中建筑物目标样本不平衡、建筑物和道路目标的相似度较高的问题。在无人机遥感数据集UDDv6.0上完成了模型训练,并在UDDv6.0测试集图像上完成了量化指标的评测实验,实验结果表明,结合warmup和LambdaLR策略及整合带权重的损失函数后,U-Net++模型在建筑物足迹提取结果上获得了更好的性能。To address the challenge of building footprint extraction in complex urban environments,a semantic segmentation U-Net++model was used,as it is particularly effective at capturing fine-grained details in high-resolution remote sensing images.To enhance model performance and reduce the risk of overfitting,the learning rate was gradually increased during the early stages of training through a warm-up process.Once the learning rate reached a predefined value,the LambdaLR strategy was implemented.Additionally,a weighted cross-entropy loss function was employed to tackle the issues of class imbalance—due to the underrepresentation of building samples—and the high similarity between buildings and roads.The model was trained on the UAV remote sensing dataset UDDv6.0,and quantitative evaluation experiments were carried out on the UDDv6.0 test set images.The experimental results demonstrate that,by combining the warm-up and LambdaLR strategies with the weighted loss function,the U-Net++model achieved significantly improved performance in building footprint extraction.

关 键 词：无人机 遥感 建筑物足迹 卷积神经网络 

分 类 号：P237[天文地球—摄影测量与遥感]