摘要
研究识别山地冰川面积对预测全球气候变化和洪水灾害具有重要意义。在总结国内外山地冰川识别方法的基础上,以各拉丹冬冰川及其周围冰川群作为研究对象,基于有关遥感数据对比分析了目前计算冰川面积的9种主要方法,包括比值法、雪盖指数法(NDSI)、非监督分类方法、监督分类方法、面向对象分类方法和神经网络分类方法,并基于混淆矩阵对这些分类方法进行了精度评价。结果表明:神经网络分类方法的总体分类精度为99.372%,比值法和NDSI的总体分类精度分别为99.370%和99.359%,最大似然分类方法、SVM分类方法和面向对象分类方法的总体分类精度均高于98%,最小距离分类方法的漏分率最高为34.51%,非监督分类方法的漏分率和错分率分别为11.07%和11.31%。试验结果表明:神经网络分类方法的整体冰川提取效果好,可以区分水体、积雪和冰;比值法和NDSI识别裸冰效果好,但无法区分水体和冰。最大似然分类方法、SVM和面向对象分类方法的整体冰川提取效果较好;最小距离分类方法容易将部分冰川区域漏分为非冰川区域;非监督分类方法容易错分冰川区和非冰川区。
Monitoring the area of glaciers is important for predicting global climate change and flood disasters.Based on summarizing the methods for identification of mountain glacier at home and abroad,the current nine major methods for calculating the area of glacier,including the ratio method,the method of normalized difference snow index(NDSI),the methods of unsupervised classification and supervised classification,the method of object-oriented classification and the method of neural net classification,are comparatively analyzed herein on the basis of the relevant remote sensing data by taking Geladandong Glacier and the glacier group around it as the study cases,and then the accuracy evaluations are made on these classification methods based on confusion matrix.The study result shows that the overall accuracy of the neural net classification method is 99.372%,the overall accuracies of the ratio method and NDSI are 99.370%and 99.359% respectively,the overall accuracies of the maximum likelihood classification,the method of support vector machines(SVM)and the method of object-oriented classification are all greater than 98%,while the classification missing rate of the method of the minimum distance classification is the highest,which is 34.51%,the classification missing rate and the classification error rate of the method of the unsupervised classification are 11.07%and 11.31%respectively.The relevant experiment shows that the overall image extraction effect of the method of neural net classification is better,from which water body,accumulated snow and ice can be divided,while the effects of both the ratio method and NDSI are better for identifying bare ice,but cannot divide water and ice.The overall effects of the method of the maximum likelihood classification,the method of SVM and the method of object-oriented classification are better for extracting glacier,while a part of glacial regions are prone to be missed and then divided into non-glacial regions and both the glacial regions and the non-glacial regions are prone to be erroneously divided by the method of unsupervised classification as well.
作者
范慧颖
董武
康宝生
张大奇
FAN Huiying;DONG Wu;KANG Baosheng;ZHANG Daqi(School of Information Science and Technology,Northwest University,Xi’an 710127,Shaanxi,China;School of Water Resources and Electric Power,Qinghai University,Xining 810016,Qinghai,China)
出处
《水利水电技术》
北大核心
2020年第5期47-58,共12页
Water Resources and Hydropower Engineering
基金
国家自然科学基金项目(51769027)。
关键词
遥感图像
山地冰川
识别
对比分析
深度学习
remote sensing image
mountain glacier
identification
comparative analysis
deep-learning
作者简介
范慧颖(1996-),女,硕士研究生,主要从事冰川遥感监测方面的研究。E-mail:1395632010@qq.com;通信作者:张大奇(1976-),男,副研究员,博士,主要从事人工智能、图像处理、视频/图像检测与分析技术及其应用研究。E-mail:dqzhang2005@aliyun.com。
