摘要
在面向土壤侵蚀的遥感影像地物提取过程中,存在严重的“椒盐”现象以及对微小地物不敏感的缺点,针对该问题,结合误差反向传播(back propagation,BP)、UNet++神经网络和面向对象思想,构建了一种二分支神经网络的深度学习方法,该方法能对在不同复杂度地区面向土壤侵蚀的地表覆盖类型进行自适应提取。以深圳市GF-2(Gaofen-2 satellite)遥感影像为例,进行面向对象分割和最优特征指标子集筛选,将分割结果以及最优特征子集作为二分支神经网络输入数据,获得地表覆盖类型提取结果。将所提方法与面向对象k最邻近(k-nearest neighbor,KNN)算法、BP和UNet++神经网络算法进行对比,结果表明:二分支神经网络对地物提取的总体精度达到91.73%,对各类别地物识别效果的综合评价指标F值(F-measure)均超过0.8,能有效抑制“椒盐”现象和微小地物的误分,评价结果优于对照组。
In the extraction process of soil erosion-oriented remote sensing image ground surface object,there are serious "pepper and salt" phenomenon and insensitivity to small surface objects.In order to solve this problem,a two-branch neural network deep learning method was constructed based on back propagation(BP),UNet++ neural network and object-oriented thought.The method can be used to adaptively extract soil erosion-oriented land cover types in different complexity regions.Taking the GF-2(Gaofen-2 satellite) remote sensing image of Shenzhen as an example,object-oriented segmentation and optimal feature index subset screening were carried out.The segmentation result and the optimal feature subset were used as the input data of the two-branch neural network to extract land cover types.The proposed method was compared with the object-oriented k-nearest neighbor(KNN),BP and UNet++ neural network algorithms.The results show that the overall accuracy of the two-branch neural network for ground object extraction reaches 91.73%,and the comprehensive evaluation index F measure of all kinds of ground object recognition effect is more than 0.8.It can effectively inhibit the phenomenon of "pepper and salt" and the misclassification of small ground objects,and the evaluation results are better than those of the control group.
作者
王书韬
史明昌
陈春阳
陈靖涛
WANG Shutao;SHI Mingchang;CHEN Chunyang;CHEN Jingtao(College of Water and Soil Conservation,Beijing Forestry University,Beijing 100083,China)
出处
《武汉大学学报（工学版）》
CAS
CSCD
北大核心
2024年第1期121-128,共8页
Engineering Journal of Wuhan University
基金
国家重点研发计划项目(编号:2017YFC0505504)。
