利用SegNet的光学卫星影像居民区要素提取  

Residential area extraction from optical satellite images using SegNet

作　　者：秦进春 刘松林[1,2] 宋昊磊 刘钊 QIN Jinchun;LIU Songlin;SONG Haolei;LIU Zhao(Xian Research Institute of Surveying and Mapping,Xi'an 710054,China;State Key Laboratory of Geo-Information Engineering,Xi'an 710054,China;Xi'an Aerospace Remote Sensing Data Technology Co.Ltd,Xi'an 710100,China)

机构地区：[1]西安测绘研究所,陕西西安710054 [2]地理信息工程国家重点实验室,陕西西安710054 [3]西安航天天绘数据技术有限公司,陕西西安710100

出　　处：《测绘科学与工程》2019年第4期18-23,共6页Geomatics Science and Engineering

摘　　要：本文针对2米空间分辨率光学卫星影像中居民区要素,提出了一种利用SegNet网络的要素提取方法。首先,通过人工判读在影像中标注感兴趣的地物要素,构建样本库;然后,使用样本库对SegNet网络进行训练;最后,设计了重叠裁切策略,利用训练好的SegNet网络对模型进行推理,并结合全连接条件随机场进行后处理,完成地物要素提取。该方法的特点在于对大范围图像的重叠裁切策略和融合使用了深度神经网络与条件随机场理论,提高了语义分割结果的精度。通过在天绘一号和高分一号卫星数据集上进行的对比实验结果表明:本文方法能够快速地提取居民区要素,具有校好的准确性和鲁棒性。A feature extraction method based on SegNet is proposed for residential area extraction from optical satellite images with 2-meter spatial resolution.Firstly,the features of interest in the image is marked through manual interpretation and the sample library is built.Then the SegNet is trained using the sample library*.Finally,the overlapping cutting strategy is designedt and the model is deduced using the trained SegNet.In addition,the model is post-processed combined with the fully connected conditional random fields to extract features.The innovation of this paper lies in the application of deep neural network and conditional random fields theory to the overlap cutting strategy and fusion of large-scale images,which improves the accuracy of semantic segmentation results.The results of comparative experiments based on the satellite data sets of TH-1 and Gaofen-1 show that the method in this paper can quickly extract residential areas with good accuracy and robustness.

关 键 词：SegWet网络 要素提取 光学彫像 语义分割 条件随机场 

分 类 号：P236[天文地球—摄影测量与遥感]