基于Deeplab v3+的高分辨率遥感影像地物分类研究  

High-resolution Remote Sensing Image Land Use Classification Based on Deeplab v3+

作　　者：陆妍如 毛辉辉 贺琰 宋现锋 LU Yanru;MAO Huihui;HE Yan;SONG Xianfeng(College of Resources and Environment,University of Chinese Academy of Sciences,Beijing 100049,China;Institute of Geographic Sciences and Natural Resources Research,Chinese Academy of Sciences,Beijing 100101,China)

机构地区：[1]中国科学院大学资源与环境学院,北京100049 [2]中国科学院地理科学与资源研究所,北京100101

出　　处：《地理空间信息》2022年第6期1-6,共6页Geospatial Information

基　　金：国家重点研发计划(2020YFC1807103,2017YFB0503702);中国科学院大学科教结合协同育人专项(117900M002);企业委托项目(E041050601)。

摘　　要：采用Deeplab v3+语义分割网络模型开展了高精度地物分类研究,并与FCN、UNet、SegNet等网络模型对比分析,结果表明Deeplab v3+提取的耕地、植被、建筑用地、道路、水系等地物图斑,其分类总体精度与Kappa系数在各项指标上均优于FCN、UNet、SegNet。此外,Deeplab v3+在路网道路、建筑物等线性特征显著、形状或边界变化剧烈的地物目标上,对图像纹理及空间几何特征的识别,都具有更高的有效性和适用性。We explored land use classification based on Deeplab v3+ semantic segmentation network model to avoid reducing the loss of spatial features. In the test of classifying popular UAV aerial images and GF-2 satellite images, Deeplab v3+ models achieved the best performances on overall accuracy, Kappa coefficient and other indicators on cultivated land, vegetation, building block, road and water body in comparison with traditional FCN, UNet, and SegNet models. Moreover, Deeplab v3+ shows a super capability on pattern recognition of building blocks and linear features(roads) and is not affected by adjacent land parcels.

关 键 词：Deeplab v3+ 高空分辨率遥感影像 语义分割 地物分类 

分 类 号：P237[天文地球—摄影测量与遥感]