摘要
针对传统极化合成孔径雷达图像分类中特征提取不完整、特征表征性不够强、分类中干扰杂质较多等问题,该文提出了一种基于Deeplab模型的极化合成孔径雷达图像地物分类方法。实验通过在荷兰地区数据上对田野、植被、建筑区、水域、山体5类进行分类,然后在欧洲其他区域进行了算法评价。与传统的结合条件随机场的FCN-8s特征分类模型相比,该文方法能够提取更高效的底层特征,得到更高的分类精度、Kappa系数和总体精度。该方法不仅能在山体上提高10%左右的分类精度,而且能在这5类以外的类别掺杂情况,保证模型良好的鲁棒性。
Aiming at the problem of the feature extraction was incomplete for traditional polarization synthetic aperture rada image classification,the characteristics were not strong enough,and there were many interference impurities in the classification,this paper proposed a method for classification of PolSAR image based on Deeplab model.The experiment classified the fields,vegetation,building area,water area and mountain body by data in the Netherlands,and then carried out algorithm evaluation in other parts of Europe.Compared with the traditional CRF-based feature classification model of C R F,this method could extract more efficient underlying features and obtained higher classification accuracy,Kappa coefficient and overall accuracy.Not only could the classification accuracy of about 10%be improved on the mountain,but also the doping of the categories other than the five categories could ensure the good robustness of the model.
作者
王云艳
罗冷坤
王重阳
WANG Yunyan;LUO Lengkun;WANG Chongyang(School of Electrical and Eleetronic Engineering,Hubei University of Technology,Wuhan 430068,China)
出处
《测绘科学》
CSCD
北大核心
2020年第6期110-117,共8页
Science of Surveying and Mapping
基金
国家自然科学基金项目(41601394)
湖北工业大学博士启动基金项目(BSQD2016010)。
关键词
极化合成孔径雷达
Deeplab网络
空洞卷积
多孔空间金字塔
polarimetric synthetic aperture radar
Deeplab netw ork
atrous convolution
atrous spatial pyramid pooling