基于特征选择和梯度提升算法的高光谱遥感地物识别  

Hyperspectral Image Classification based on Feature Selection and Gradient Boosting

作　　者：张梓童 王恒 高世臣 张维[2] ZHANG Zi-tong;WANG Heng;GAO Shi-chen;ZHANG Wei(School of Science,China University of Geosciences(Beijing),Beijing 100083,China;No.1 Oil Production Plant,PetroChina Changqing Oilfield Company,Yan'an 716000,China)

机构地区：[1]中国地质大学(北京)数理学院,北京100083 [2]中国石油长庆油田分公司第一采油厂,陕西延安716000

出　　处：《数学的实践与认识》2019年第22期187-196,共10页Mathematics in Practice and Theory

摘　　要：高光谱遥感数据波段数目较多,且波段之间的相关性高,影响到敏感波段在地物识别中的作用,并造成大量冗余计算,降低时效.提出了一种随机森林结合递归特征消除的敏感特征选择方案,以提高高光谱遥感地物识别的精度与效率.通过RF-RFE特征选择方法得到最优特征组合,并运用LightGBM和XGBoost等提升算法来提高分类精度.在江苏省常州的茶树数据集上进行分类实验时,在原始数据上的分类精度达到了94.27%和94.45%;在特征选择出的最优特征子集上进行实验时,分类精度达到了94.40%和94.36%.实验结果表明,该方案的分类精度要优于决策树和朴素贝叶斯等传统分类算法,同时大幅减少了运算量,取得了较好的识别效果,具有一定的推广和应用价值.Hyperspectral remote sensing data has a large number of bands,and the correlar tion between the bands is high,which affects the role of sensitive bands in feature recognition,and causes a large number of redundant calculations and reduces aging.A sensitive fear ture selection scheme based on random forest combined with recursive feature elimination is proposed to improve the accuracy and efficiency of hyperspectral remote sensing object recognition.The RF-RFE feature selection method is used to obtain the optimal feature combination,and the lifting algorithms such as LightGBM and XGBoost are used to improve the classification accuracy.When the classification experiments were carried out on the tea tree dataset in Changzhou,Jiangsu Province,the classification accuracy on the original data reached 94.27% and 94.45%.When the experiment was performed on the optimal feature subset selected by the feature,the classification accuracy reached 94.40%.And 94.36%.The experimental results show that the classification accuracy of this scheme is better than that of traditional classification algorithms such as decision tree and naive Bayes.At the same time,the computational complexity is greatly reduced,and a better recognition effect is obtained,which has certain promotion and application value.

关 键 词：高光谱遥感 特征选择 梯度提升算法 随机森林 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]