基于主成分分析下贝叶斯优化卷积神经网络模型人工林树种识别的研究  

Study on Tree Species Identification of Planted Forests Based on PCA-BO-CNN Model

作　　者：王晓红 辛守英 张薇 焦琳琳 WANG Xiaohong;XIN Shouying;ZHANG Wei;JIAO Linlin(College of Mining Engineering,North China University of Science and Technology,Tangshan 063210,China;Hebei Industrial Technology Institute of Mine Ecological Remediation,Tangshan 063210,China;Mine Green Intelligent Mining Technology Innovation Center of Hebei Province,Tangshan 063210,China;Hebei Geological Engineering Exploration Institute,Baoding 071051,China;Hebei University of Water Resources and Electric Engineering,Cangzhou 061001,China)

机构地区：[1]华北理工大学矿业工程学院,河北唐山063210 [2]河北省矿区生态修复产业技术研究院,河北唐山063210 [3]河北省矿山绿色智能开采技术创新中心,河北唐山063210 [4]河北省地质工程勘查院,河北保定071051 [5]河北水利电力学院,河北沧州061001

出　　处：《森林工程》2025年第2期298-311,共14页Forest Engineering

基　　金：中央引导地方科技发展资金项目(246Z5901G);河北省自然科学基金青年基金(D2019209322,2019209317);唐山市科技局应用基础研究计划(20130202b);华北理工大学博士专项经费(BS201818)。

摘　　要：为探究基于主成分分析(Principal Component Analysis,PCA)下贝叶斯优化(Bayesian Optimization,BO)卷积神经网络(Convolutional Neural Network,CNN)算法(PCA-BO-CNN)模型对人工林树种识别的方法,以提高遥感技术在人工林树种识别中的准确率和鲁棒性。以塞罕坝机械林场为研究区域,利用Sentinel-1遥感数据、Sentinel-2遥感数据、数字高程模型(digital Elevation Model,DEM)数据及森林资源二类调查数据和PCA-BO-CNN算法模型结合,并与其他不同算法模型对比分析,以提高人工林树种识别的准确性。结果表明,1)相比PCA算法处理前,PCA算法处理后多源数据特征的PCA1—PCA39共计39个特征的标准差和特征间的区分性明显提升。因此,PCA算法处理有利于提升对华北落叶松、白桦、樟子松、蒙古栎和云杉主要优势树种及非林地的识别精度;2)在PCA算法处理前,BO-随机森林(random forest,RF)算法模型对主要优势树种及非林地识别的总体准确度(OA)和Kappa系数精度,分别为81.87%,0.7545。在PCA算法处理后,PCA-BO-CNN算法模型对主要优势树种及非林地识别的OA和Kappa系数精度相对提高,分别为83.10%,0.7703;3)相比PCA算法处理前的BO-RF算法模型,PCA算法处理后的PCA-BO-CNN算法模型对塞罕坝林场主要优势树种及非林地识别的调和平均数(F1)、OA和Kappa系数的整体精度相对较高。具体,相比BO-RF算法模型PCA-BO-CNN算法模型的OA提升了1.24%,且相比PCA算法处理前PCA-BO-CNN算法模型OA提升了3.71%。与其他算法模型相比,基于PCA-BO-CNN算法模型的人工林树种识别方法具有很强的准确性和鲁棒性,为掌握塞罕坝林场人工林的树种分布,进而了解森林碳储量、森林对气候变化的响应、制定碳减排政策以及推动森林可持续发展提供重要的理论依据。To explore the identification method of tree species in planted forests based on the Bayesian optimization convolutional neural network(PCA-BO-CNN)algorithm model based on principal component analysis,to improve the accuracy and robustness of remote sensing technology in tree species identification in planted forests.In this study,the Sahanba Mechanical Forest Farm was selected as the study area.Sentinel-1 remote sensing data,Sentinel-2 remote sensing data,DEM data,and the forest resource category 2 survey data were combined with the PCA-BO-CNN algorithm model,and compared with other algorithm models,to improve the accuracy of tree species identification in planted forests.The results showed that:(1)Compared with the pre-PCA algorithm,the standard deviation of 39 features of PCA1-PCA39 of multi-source data features after PCA algorithm processing and the differentiation among features were significantly improved.Therefore,PCA was beneficial to improve the identification accuracy of the dominant species of Larix gmelinii var.principis-rupprechtii(Mayr)Pilger,Betula platyphylla Sukaczev,Pinus sylvestris var.mongholica Litv.,Quercus mongolica Fisch.ex Ledeb.and Picea asperata Mast.as well as non-forest land.(2)Before PCA algorithm,the overall accuracy(OA)and Kappa coefficient accuracy of the BO-random forest(BO-RF)algorithm model for the identification of dominant tree species and non-forest land were 81.87%and 0.7545,respectively.After PCA algorithm processing,the accuracy of OA and Kappa coefficients of the PCA-BO-CNN algorithm model for the identification of dominant tree species and non-forest land was relatively improved,which were 83.10%and 0.7703,respectively.(3)Compared with the BO-RF algorithm model before PCA algorithm processing,the overall accuracy of the PCA-BO-CNN algorithm model after PCA algorithm processing was relatively higher for the identification of the F1,OA and Kappa coefficients of the main dominant tree species and non-forest land in Saihanba Forest Farm.Specifically,compared with the BO-RF

关 键 词：PCA-BO-CNN模型 塞罕坝林场 人工林 遥感技术 树种识别 

分 类 号：S771.8[农业科学—森林工程]