应用特征选择和机器学习方法建立兴安落叶松单木树龄预测模型  

Establishing a Prediction Model for the Age of Larix gmelinii Individual Tree Using Feature Selection and Machine Learning Methods

作　　者：王晓楠 苏文浩 董灵波 Wang Xiaonan;Su Wenhao;Dong Lingbo(Ministry of Education Key Laboratory of Sustainable Forest Ecosystem Management(Northeast Forestry University),Harbin 150040,P.R.China)

机构地区：[1]森林生态系统可持续经营教育部重点实验室(东北林业大学),哈尔滨150040

出　　处：《东北林业大学学报》2024年第11期64-71,82,共9页Journal of Northeast Forestry University

基　　金：“十四五”国家重点研发计划项目(2022YFD2200502);黑龙江省头雁创新团队计划项目(森林资源高效培育技术研发团队)。

摘　　要：林木树龄是制约天然林生长、更新和演替的重要因素,但天然林中单木树龄分异明显且实测困难。为了探讨不同特征选择方法和机器学习算法的组合对单木树龄模型预测精度的影响,以大兴安岭地区兴安落叶松天然林为研究对象,以44块固定样地的280株单木树龄实测数据为基础,采用皮尔森(Pearson)相关系数(PEA)和距离相关系数(DCC)特征选择方法,对单木、林分、竞争、地形和气候因素的38个因子进行筛选,并采用随机森林(RF)、支持向量回归(SVR)和k-近邻(k-NN)算法建立兴安落叶松天然林单木树龄预测模型,分析制约单木树龄预测模型精度的关键因素。结果表明:(1)当决策树为1500、节点数量为4时,DCC-RF模型预测精度最高(确定系数(R^(2))为0.5550,均方根误差(R_(MSE))为10.1669 a,平均绝对误差(M_(AE))为8.4039 a);(2)相较于DCC-SVR和PEA-k-NN模型,DCC-RF模型的R^(2)分别提高了0.0030和0.1521,均方根误差分别降低了0.2568和2.0742 a,平均绝对误差分别降低了0.0148和1.3750 a;(3)影响树龄预测最重要的指标是胸径(重要值为43.85%),其次为树高(重要值为19.14%)和生长积温(重要值为13.41%)等。表明应用距离相关系数和随机森林算法,构建兴安落叶松天然林的单木树龄预测模型具有较好的适应性。The age of trees is a critical factor that restricts the growth,regeneration,and succession of natural forests.However,the age variation among individual tree in natural forests is pronounced and difficult to measure.To explore the impact of different feature selection methods and machine learning algorithms on the predictive accuracy of individual-tree age model,we focused on the natural forest of Larix gmelinii in the Great Xing’an Mountains.Using measured age data from 280 individual trees across 44 fixed sample plots,we employed Pearson correlation coefficient(PEA)and distance correlation coefficient(DCC)feature selection methods to select from 38 factors related to individual trees,stands,competition,terrain,and climatic conditions.We established age prediction models for L.gmelinii using Random Forest(RF),Support Vector Regression(SVR),and k-Nearest Neighbors(k-NN)algorithms,and analyzed the key factors influencing the accuracy of the individual-tree age prediction model.The results showed that:(1)The DCC-RF model showed the highest predictive accuracy when the number of decision trees was set to 1500 and nodes to 4(coefficient of determination(R^(2))=0.5550,root mean square error(R_(MSE))=10.1669 a,mean absolute error(M_(AE))=8.4039 a).(2)Compared to the DCC-SVR and PEA-k-NN models,the R^(2) of the DCC-RF model improved by 0.0030 and 0.1521 respectively,while the R_(MSE) decreased by 0.2568 and 2.0742 a,and the M_(AE) decreased by 0.0148 and 1.3750 a respectively.(3)The most significant predictor for age estimation was the diameter at breast height(DBH)with an importance value of 43.85%,followed by tree height(19.14%)and growth accumulated temperature(13.41%).These results demonstrate that the application of the distance correlation coefficient and Random Forest algorithm is suitable for constructing age prediction model for L.gmelinii in natural forests.

关 键 词：兴安落叶松 树龄预测 距离相关系数 随机森林 

分 类 号：S757[农业科学—森林经理学]