面向对象卷积神经网络的耕作梯田提取  

Extraction of Farming Terraces Using Object-based Convolutional Neural Networks from Very High Resolution Satellite Images

作　　者：周珏 李蒙蒙 汪小钦[1] 吴思颖 金时来 ZHOU Jue;LI Mengmeng;WANG Xiaoqin;WU Siying;JIN Shilai(Key Laboratory of Spatial Data Mining and Information Sharing of Ministry of Education,Fuzhou University,Fuzhou 350116,China;Soil and Water Conservation Experimental Station of Fujian Province,Fuzhou 350003,China)

机构地区：[1]福州大学空间数据挖掘与信息共享教育部重点实验室,福州350116 [2]福建省水土保持试验站,福州350003

出　　处：《遥感信息》2022年第2期138-144,共7页Remote Sensing Information

基　　金：国家自然科学基金项目(42001283);福建省高校产学研重点项目(2017Y4010)。

摘　　要：针对目前高分辨率遥感影像耕作梯田提取方法普遍精度不高的问题,提出一种面向对象与卷积神经网络相结合的方法。以福建省南平市为例,构建面向对象卷积神经网络,利用高分辨率GF-2和ZY-3遥感数据进行耕作梯田精细提取,并对比分析深度学习与传统方法、不同分辨率数据源以及不同分类器对提取效果的影响。结果表明:该方法总体精度达到87.1%,Kappa系数为0.76,与采用低层次特征的随机森林分类对比,总体精度提高了10.2%;分别结合深层次特征与随机森林、XG Boost和Ada Boost分类器,总体精度差异小于2%;该方法基于GF-2影像的提取精度较ZY-3提高了4.6%。此方法可有效表征高分辨率影像梯田对象的深层图像特征,并顾及影像中梯田的边界信息,实现了梯田的精细提取。This paper deals with the difficulties of extracting farming terraces from high resolution remote sensing images.It proposes an object-based convolutional neural network(CNN)by combining object-based image analysis and a CNN to extract farming terraces from Chinese high-resolution GF-2 and ZY-3 remote sensing data.It conducts the extraction in Nanping city,Fujian province,and compares extraction results between the proposed method and other existing methods,i.e.,random forest,XG Boost and Ada Boost,and between different high resolution remote sensing images.The results show that:(1)This method extracts farming terraces with an overall accuracy of 87.1%and a Kappa coefficient of 0.76.Compared with random forest classification using low-level image object features,the obtained overall accuracy is improved by 10.2%.(2)Based upon high-level image features,the extraction results of farming terraces via different classifiers are similar,referring to an accuracy difference less than 2%.(3)Comparing between GF-2 and ZY-3 images,the extraction accuracy using the GF-2 dataset is 4.6%higher than that of the ZY-3 dataset.The proposed object-based CNN method can effectively extract terraces from high resolution remote sensing images when using deep image features and considering the boundary information of farming terraces.

关 键 词：高分辨率遥感 卷积神经网络 面向对象分析 梯田提取 迁移学习 

分 类 号：S127[农业科学—农业基础科学]