优化随机森林的高光谱图像降维及分类算法  

Optimized random forest algorithm for hyperspectral image dimensionality reduction and classification

作　　者：李淑英 程磊 彭柏栋 张强 LI Shuying;CHENG Lei;PENG Baidong;ZHANG Qiang(School of Automation,Xi’an University of Posts and Telecommunications,Xi’an 710121,China)

机构地区：[1]西安邮电大学自动化学院,陕西西安710121

出　　处：《西安邮电大学学报》2022年第5期67-76,共10页Journal of Xi’an University of Posts and Telecommunications

摘　　要：对传统随机森林(Random Forest,RF)特征降维方法存在的去冗余能力不足及随机森林分类器缺乏有效的寻优方法获取最优超参数的问题进行研究,提出一种优化随机森林的高光谱图像降维及分类(Optimized Random Forest Algorithm for Hyperspectral Image Dimensionality Reduction and Classification,ORFDRC)算法。该算法通过增强型随机森林降维(Enhanced Random Forest Dimension Reduction,ERFDR)的预处理,使用网格-爬山(Grid-Hill Climbing,GHC)算法对RF模型进行超参数优化,并构建优化后的网格-爬山随机森林(Grid-Hill Climbing Random Forest,GHC-RF)分类器。最后,将降维后得到的新特征集输入到GHC-RF分类器中进行分类。在多个数据集上与多种降维算法及分类器进行实验对比,验证结果表明,ORFDRC算法在完成高光谱图像分类任务的同时,具有更高的分类精度。For the problems that the traditional random forest(RF)feature dimension reduction method has insufficient redundancy capability,and the random forest classifier lacks effective optimization method to obtain optimal hyperparameters,an optimized random forest dimensionality reduction and classification algorithm(ORFDRC)for hyperspectral image is proposed.It is based on the enhanced random forest dimension reduction(ERFDR)preprocessing and the grid-hill climbing(GHC)algorithm for hyperparameter optimization of RF.The grid-hill climbing random forest(GHC-RF)classifier is constructed.Finally,the new feature set obtained after dimension reduction is input into the GHC-RF classifier for classification.The experimental comparison with multiple dimension reduction algorithms and classifiers on multiple data sets is carried out.It shows that ORFDRC algorithm has higher classification accuracy for hyperspectral image classification.

关 键 词：高光谱图像 随机森林 超参数寻优 遥感图像分类 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]