南京林业大学学报(自然科学版) ›› 2023, Vol. 47 ›› Issue (1): 31-37.doi: 10.12302/j.issn.1000-2006.202104014

所属专题: 智慧林业之森林参数遥感估测

• 专题报道Ⅰ:智慧林业之森林参数遥感估测 • 上一篇    下一篇

基于机器学习算法的樟子松立木材积预测

孙铭辰(), 姜立春()   

  1. 东北林业大学林学院,森林生态系统可持续经营教育部重点实验室, 黑龙江 哈尔滨 150040
  • 收稿日期:2021-04-12 接受日期:2021-08-04 出版日期:2023-01-30 发布日期:2023-02-01
  • 通讯作者: 姜立春
  • 基金资助:
    国家自然科学基金项目(31570624);黑龙江省应用技术研究与开发计划(GA19C006);黑龙江省头雁创新团队计划,国家林业和草原科学数据中心黑龙江子平台(2005DKA32200-OH)

Standing volume prediction of Pinus sylvestris var. mongolica based on machine learning algorithm

SUN Mingchen(), JIANG Lichun()   

  1. Key Laboratory of Sustainable Forest Ecosystem Management, Ministry of Education, School of Forestry, Northeast Forestry University, Harbin 150040, China
  • Received:2021-04-12 Accepted:2021-08-04 Online:2023-01-30 Published:2023-02-01
  • Contact: JIANG Lichun

摘要: 【目的】 通过非线性和多种机器学习算法构建并对比不同的立木材积模型,为樟子松(Pinus sylvestris var. mongolica)立木材积的精准预测提供理论依据。【方法】 以大兴安岭图强林业局184株樟子松伐倒木数据为基础,建立非线性二元材积模型(NLR),并通过十折交叉检验和袋外数据(OOB)误差检验的方法得到3种最优机器学习算法,包括:反向神经网络(BP)、ε-支持向量回归(ε-SVR)和随机森林(RF)。对比分析不同模型间的差异,得到最优立木材积模型。【结果】 机器学习算法在立木材积的拟合和预测中均优于传统二元材积模型,具体拟合结果排序为RF>BP>ε-SVR> NLR。其中RF的决定系数(R2)比传统模型的提高了2.00%,均方根误差(RMSE)、相对均方根误差(RMSE%)、平均绝对误差(MAE)分别降低了22.90%、22.93%、36.34%,且与真实值相比平均相对误差(MRB)的绝对值更低,证明了RF在立木材积预测中的优越性。【结论】 机器学习算法作为一种新兴的建模方法可以有效地提高立木材积的预测精度,为森林资源的精准调查和经营管理提供新的解决方案。

关键词: 樟子松, 二元材积模型, BP神经网络, ε-支持向量回归(ε-SVR), 随机森林(RF)

Abstract:

【Objective】 Using various, nonlinear machine learning algorithms, different volume models were constructed and compared to provide a theoretical basis for the accurate prediction of the volume of Pinus sylvestris var. mongolica.【Method】 A total of 184 felled Pinus sylvestris var. mongolica trees in the Tuqiang Forestry Bureau of the Greater Khingan Mountains were used to establish a nonlinear binary volume model (NLR). Three optimal machine learning algorithms were obtained using the K-fold cross test and OOB error test, including back propagation neural network (BP), ε-support vector regression (ε-SVR), and random forest (RF). An optimal volume model was obtained by comparing and analyzing the differences between the different models. 【Result】 The results showed that the machine learning algorithm was superior to the traditional binary volume model in the fitting and prediction of standing volume, and the specific order was RF > BP > ε-SVR > NLR. Compared with the traditional model, the R2 of RF increased by 2.00%; the RMSE, RMSE% and MAE decreased by 22.95%, 22.93% and 36.34%, respectively; and the absolute value of MRB was lower than the real value, which proved the superiority of RF in volume prediction. 【Conclusion】 Machine learning algorithms can effectively improve the accuracy at which standing volume can be predicted, providing a new solution for the accurate investigation and management of forest resources.

Key words: Pinus sylvestris var. mongolica, binary volume model, BP neural network, ε-support vector regression(ε-SVR), random forest(RF)

中图分类号: