改进UNet+ +与联合损失函数的建筑物提取方法

Building Extraction Method Based on ImprovedUNet + + and Joint Loss Function

针对现有遥感影像建筑物提取方法中存在的特征提取不充分等问题,提出了一种改进UNet++与联合损失函数的建筑物提取方法。通过引入ResNeSt50编码器以增强对建筑物复杂结构的特征提取能力,采用融合Sigmoid激活函数和二进制交叉熵损失的联合损失函数,以平衡模型的准确性和训练的收敛速度,进而提高模型的整体性能。在制作的赣州建筑物数据集上进行了验证,Precision、Recall和F1-Score值分别为0.945、0.926、0.935,对比FPN、PAN、PSPNet等主流网络精度较高,在WHU全球卫星建筑物数据集上进一步验证了模型的泛化能力,Precision、Recall和F1-Score值分别达到了0.894、0.870、0.882。

To address the deficiencies in feature extraction observed in existing methodologies for the extraction of buildings from remote sensing imagery,this study introduces an advanced ap-proach that synergizes an augmented UNet++architecture with a composite loss function,The integration of a ResNeSt5o cncoder scrves to cnhance the extraction of complex structural fea-tures of buildings,while a hybrid loss function combining the Sigmoid activation function with binary cross-cntropy loss is employed to optimize the balance between model accuracy and the spccd of convergence during training,thercby clevating the overall performanc of the model.Validation conducted on a spccially compiled Ganzhou building datasct yielded Prccision,Recall,and F1-Score metrics of 0.945,0.926,and 0.935,respectively,indicating superior accuracy over conventional networks such as the FPN,PAN,and PSPNet.Further validation on the WHU global satellite building dataset affirmed the model's robust gencralization capabilities,with Pre-cision,Recall,and F1-Score achieving 0.894,0.870,and 0.882,respectively.