基于特征过滤法和Stacking集成学习的无人机影像作物精细分类  

UAV image fine crop classification based on feature filtering method and Stacking ensemble learning

作　　者：刘朝辉 杨风暴[1] 张琳 LIU Zhaohui;YANG Fengbao;ZHANG Lin(School of Information and Communication Engineering,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学信息与通信工程学院,山西太原030051

出　　处：《现代电子技术》2025年第7期1-10,共10页Modern Electronics Technique

基　　金：国家自然科学基金面上项目(61972363);中央引导地方科技发展资金项目(YDZJSX2021C008)。

摘　　要：针对目前多种典型作物分类中特征冗余导致同科作物混淆、分类精度低的问题,文中提出一种结合特征过滤法筛选特征和Stacking集成学习的作物精细分类方法。首先,结合敏感波段构造新型植被指数并进行阈值分割,实现作物区域提取;然后,提取不同作物的颜色和纹理特征,进而计算单类作物特征系数和作物间特征差异系数,实现各典型作物的分类特征过滤法优选;最后,构建融合多种机器学习算法的Stacking集成学习作物分类模型,其中第一层的基学习器选择随机森林、支持向量机、K⁃最近邻算法,第二层的元学习器选择逻辑回归模型,实现多种典型作物精细分类。实验结果表明,所提方法对7种典型作物的总体分类精度和Kappa系数分别为85.2%和83.34%,相比于未进行特征选择的分类结果分别提升了2.18%和3.68%,具有较高的分类精度,为多种典型作物的精细分类提供了新方法。The feature redundancy in multiple typical crop classifications at present leads to confusion and low classification accuracy of crops of the same family,so this paper proposes a crop fine classification method that combines the feature filtering method for feature screen and Stacking ensemble learning.A new type of vegetation index is constructed by combining sensitive bands,and the threshold value is segmented,so as to realize crop region extraction.The color and texture features of different crops are extracted,and then the feature coefficients of a single type of crop and the coefficients of feature differences among crops are calculated,so as to realize the classification feature filtering method preference for each typical crop.Finally,a Stacking ensemble learning crop classification model that integrates multiple machine learning algorithms is constructed.Among them,the random forest(RF),support vector machine(SVM)and K⁃nearest neighbor(K⁃NN)algorithms are selected for the base learner in the first layer,and the logistic regression model is selected for the meta⁃learner in the second layer,so that the various typical crops are classified finely.The experimental results show that the overall classification accuracy and Kappa coefficient of the proposed method for the seven typical crops are 85.2%and 83.34%,respectively,which are 2.18%and 3.68%higher than the classification results without feature selection.To sum up,the proposed method has high classification accuracy,and can be used as a new method for the fine classification of multiple typical crops.

关 键 词：作物分类 特征选择 Stacking集成学习 植被指数 阈值分割 衍生特征 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP751[电子电信—信息与通信工程]