遥感影像样本自动生成与智能迭代分类方法  

Automatic samples generation and intelligent iterative classification of remote sensing images

作　　者：李欣 杨懿[2] 王宁 顾海燕 丁少鹏 李海涛[2] LI Xin;YANG Yi;WANG Ning;GU Haiyan;DING Shaopeng;LI Haitao(China Energy Engineering Group Jiangsu Power Design Institute Co.,Ltd.,Nanjing 211102,China;Chinese Academy of Surveying and Mapping,Beijing 100036,China)

机构地区：[1]中国能源建设集团江苏省电力设计院有限公司,南京211102 [2]中国测绘科学研究院,北京100036

出　　处：《测绘科学》2022年第8期197-203,共7页Science of Surveying and Mapping

基　　金：中央级公益性科研院所基本科研业务费项目(AR2123,AR2203,AR2313)

摘　　要：针对先验知识未能有效指导智能分类、智能分类与样本采集相对独立的问题,提出遥感影像样本自动生成与智能迭代分类方法。首先利用遥感影像及对应的历史解译数据构建样本数据集;其次利用深度卷积神经网络模型进行训练,得到预训练模型;再次利用预训练模型对即时遥感影像进行智能分类,得到智能分类结果;最后将校正后分类结果反馈到样本数据集,完成样本数据集的更新,利用更新后的样本数据集对智能分类模型进行迭代优化,形成模型与样本的优化闭环。试验结果表明:该方法通过样本自动生成与更新,以及模型迭代训练,能够提升智能分类模型精度,为解决先验知识未充分利用、智能分类与样本采集相对独立、分类结果未实时反馈等问题提供思路。Aiming at the problems that prior knowledge cannot effectively guide intelligent classification,intelligent classification and sample collection were relatively independent,a method of remote sensing image sample generation and intelligent iterative classification was proposed.Firstly,a sample data set was constructed using remote sensing images and corresponding historical interpretation data.Secondly,an intelligent classification model was obtained using a deep convolutional neural network model.Thirdly,the real-time remote sensing images were intelligently classified using the intelligent classification model,and the intelligent classification results were obtained for analysis and correction.At last,the corrected classification results were constructed to update the sample data set,which was fed back to the sample data set to complete the update of the sample data set.The intelligent classification model was iteratively optimized using the updated sample data set to form an optimization closed loop between the model and the sample.The experimental results showed that the method could improve the accuracy of the intelligent classification model,and provided ideas for solving the problems of underutilization of prior knowledge,relatively independent intelligent classification and sample collection,and no real-time feedback of classification results.

关 键 词：遥感影像分类 样本数据集 智能迭代分类 深度学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]