基于Attention Gates和R2U-Net的遥感影像建筑物提取方法  

Building Extraction from Remote Sensing Images Based on the R2U-Net Model and Attention Gates

作　　者：于文玲 刘波[1,2] 刘华 杜梓维 邹时林 苏友能[1] 刘娜娜 YU Wen-ling;LIU Bo;LIU Hua;DU Zi-wei;ZOU Shi-lin;SU You-neng;LIU Na-na(Faculty of Geomatics,East China University of Technology,Nanchang 330013;Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake,Ministry of Natural Resources,Nanchang 330013,China)

机构地区：[1]东华理工大学测绘工程学院,江西南昌330013 [2]自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,江西南昌330013

出　　处：《地理与地理信息科学》2022年第3期31-36,42,共7页Geography and Geo-Information Science

基　　金：国家自然科学基金项目(42161064、42001411);江西省自然科学基金项目(20202BABL212014、20212BAB204003);东华理工大学江西省数字国土重点实验室开放研究基金项目(DLLJ202004)。

摘　　要：针对深度语义分割算法提取遥感影像建筑物时易产生建筑物边缘分割不明确、提取精度不高等问题,该文提出一种基于Attention Gates(AG)和R2U-Net的遥感影像建筑物提取方法(AGR2U-Net)。该方法将R2U-Net模型每一层输出的特征图与其相邻层的特征图输入至改进的AG模型中,得到与输入影像大小一致的特征图,以提高R2U-Net模型的多尺度泛化能力,从而增强该模型对建筑物特征的响应及灵敏度,最终提升遥感影像建筑物提取精度。利用WHU卫星影像数据集和WHU航空影像数据集,对该方法与U-Net、Improved U-Net、SegU-Net和R2U-Net方法进行对比实验验证,结果表明,该方法的交并比、像素准确率和召回率均最高,且提取的建筑物边缘更准确、内部信息更完整、误检和漏检情况更少。In the process of building extraction from remote sensing images,deep learning semantic segmentation algorithms are prone to cause the shortcomings,such as unclear edge segmentation of buildings and low accuracy of building extraction in the recurrent residual U-Net(R2U-Net)model.In order to improve aforementioned shortcomings,this paper proposed a building extraction method from remote sensing images by combining Attention Gates(AG)and the R2U-Net model.In this method,the output feature map of each layer and the feature maps of its adjacent layers from the R2U-Net model were input into the improved AG model,to obtain the feature maps with the same size as the input images and improve the multi-scale generalization of the R2U-Net model,which can enhance the response and sensitivity of the model to the building features,and ultimately improve the accuracy of building extraction from remote sensing images.Finally,the proposed method was verified with other methods by using WHU satellite dataset and WHU aerial imagery dataset.The extraction results of the buildings showed that the mean values of the proposed method are the highest in the three evaluation indexes of intersection over union,pixel accuracy and recall,compared to those of the U-Net,Improved U-Net,SegU-Net and R2U-Net models.Moreover,this method can extract more accurate edges in the building extraction from remote sensing images and more complete internal information of buildings with lower false and missing rates.

关 键 词：遥感影像 Attention Gates R2U-Net模型 AGR2U-Net模型 建筑物提取 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]