基于熵权与集成学习的半监督小样本树种分类研究  

Research on Semi Supervised Small Sample Tree Species Classification Based on Entropy Weight and Ensemble Learning

作　　者：王静 李静 WANG Jing;LI Jing(Information Engineering College,Henan Mechanical and Electrical Vocational College,Zhengzhou 451191,China;College of Information Engineering,Henan University of Science and Technology,Luoyang 471000,China)

机构地区：[1]河南机电职业学院信息工程学院,郑州451191 [2]河南科技大学信息工程学院,河南洛阳471000

出　　处：《森林工程》2025年第1期151-161,共11页Forest Engineering

基　　金：国家自然科学基金项目(62071171);黄炎培职业教育思想研究规划课题(ZJS2022Zd33);中西部地区本科层次职业教育理论与实践研究(22GDZY0229);面向复杂场景的小目标检测与识别关键技术研究(242102210071)。

摘　　要：针对传统半监督自训练分类方法易导致数据集混乱,影响后续小样本树种分类精度这一问题,基于熵权法(en-tropy weight,EW)与集成学习(ensemble learning,EL)提出EW-EL的半监督小样本树种分类方法。EW-EL在传统半监督自训练分类方法的理论上引入EL的思想,以熵权法作为基础理论设计按基分类器当前训练周期下的F1分数计算的信息熵作为计算权重因子,再依信息熵越大基分类器越不稳定思想设计权重,使集成分类器分类概率更集中,减少集成分类器偏向性。结果显示,EW-EL较传统半监督自训练方法能更有效地均衡数据分布,使新加入数据的伪标签样本类别更准确。EW-EL所得到的小样本树种分类总精度(OA)为0.97、召回率(Recall)为0.96及Kappa系数为0.97,3种指标均优于监督分类、传统半监督自训练方法及利用传统EL机制所构建的半监督自训练方法。其中,EW-EL方法较融合软投票机制的半监督自训练方法,OA与Recall均提升了1%。EW-EL联合简单线性迭代聚类所制成的树种图在所选测试区内达到了94%。此外,进一步分析证明,EW-EL能通过集成诸多分类器,来实现更佳的小样本树种分类结果,更适用于低成本下的相关部门进行林业资源统计的工作。To address the issue that traditional semi-supervised self-training classification methods can lead to dataset confusion,affecting the accuracy of subsequent small-sample tree species classification,an EW-EL(entropy weight and ensemble learning) semi-supervised small-sample tree species classification method is proposed based on the entropy weight method(EW) and ensemble learn-ing(EL).EW-EL introduces the concept of EL into the theoretical framework of traditional semi-supervised self-training classification methods,using the entropy weight method as a foundational theory.It calculates the information entropy based on the F1 score of base classifiers in the current training cycle as a weight factor.Then,design the weights according to the idea that the larger the information entropy,the more unstabel the base classifier will be.This will make the classification probabilities of the ensemble classifier more concentrated and reduce the bias of the ensemble classifier.The findings demonstrate that,in contrast to conventional semi-super-vised self-training techniques,EW-EL can efficiently balance data distribution,producing more precise pseudo-label sample catego-ries for recently added data.With a recall of 0.96 and a Kappa coefficient of 0.97,the overall accuracy(OA) of the EW-EL method for small-sample tree species classification is 0.97.All three indicators are superior to supervised classification,conventional semi-su-pervised self-training techniques,and semi-supervised self-training techniques built using conventional EL mechanisms.In particu-lar,the EW-EL approach outperforms semi-supervised self-training techniques that incorporate a soft voting mechanism in terms of OA and recall by 1%.Furthermore,in the chosen test area,the tree species map produced with EW-EL in combination with basic linear iterative clustering reached 94% accuracy.Moreover,extra analyses show that EW-EL can integrate several classifiers to provide bet-ter small-sample tree species classification results,which makes it more appropriate f

关 键 词：无人机影像 熵权法 深度学习 集成学习 半监督小样本分类 树种分类 树种制图 EW-EL 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]