摘要
提出一种基于迭加权相关权重矩阵(IR-CWM),使用XGBoost机器学习模型训练选取特征变化波段,训练预测变化检测结果的方法(IR-CWM-XGB).以国产高分五号(GF-5)高光谱遥感影像为数据源,经过预处理后,通过迭代加权得到迭加权相关权重矩阵(IR-CWM),然后经过XGBoost算法选取特征波段,选取随机样本进行模型训练,使用最终模型生成变化的结果.实验结果表明:本研究方法与变化矢量分析(CVA)变化检测方法、主成分变化矢量分析(PCA-CVA)变化检测方法、迭加权多元(IR-MAD)方法、卷积神经网络(CNN)等方法进行对比,本研究方法变化检测结果的Kappa系数和总体精度较高,误检率较低.
We proposed a method based on iteration reweighted correlation weight matrix(IR-CWM),used the XGBoost machine learning model to feature selection,and generated the change detection results(IR-CWM-XGB).After preprocessing the data source of GF-5 hyperspectral image,we obtained the iteration reweighted correlation weight matrix(IR-CWM),and feature selection based on XGBoost model,random samples for model training,and the change detection results were done by the final model.The experimental results show that,compared with the change vector analysis(CVA),convolutional neural network(CNN),principal component change vector analysis(PCA-CVA)and iteration reweighted multivariate alteration detection(IR-MAD).the proposed algorithm of this paper has the optimal results,which has high precision,high Kappa coefficient and low false detection rate.
作者
魏立飞
张杨熙
尹峰
黄庆彬
WEI Lifei;ZHANG Yangxi;YIN Feng;HUANG Qingbin(Faculty of Resources and Environmental Science, Hubei University, Wuhan 430062, China;Hubei Institute of Land and Resources, Wuhan 430071, China;Shenzhen Cadastral Surveying & Mapping Office, Shenzhen 518034, China)
出处
《湖北大学学报（自然科学版）》
CAS
2020年第4期398-403,410,共7页
Journal of Hubei University：Natural Science
基金
国家自然科学基金(41622107)
地球信息工程国家重点实验室开放基金(SKLGIE2018-M-3-3)
空间数据挖掘与信息共享教育部重点实验室开放基金(2018LSDMIS05)
测绘遥感信息工程国家重点实验室开放基金(18R02)资助。
