基于微钻阻力与 LightGBM 的活立木木材密度区间分类预测模型

An interval classification prediction model for standing tree wood density based on micro-drilling resistance and LightGBM

摘要

【目的】微钻阻力法是活立木木材密度无损检测的主要技术之一，然而，现有基于该技术的回归预测模型易受仪器、操作及环境等多源干扰，存在稳定性差、通用性弱的问题，制约了其在工程实践（如木材分级、野外快速检测）中的应用，亦限制了科研中的可靠分析。为提升微钻阻力法在木材密度预测中的稳健性与实用性，本研究旨在探索一种能够包容测量不确定性的建模方法。【方法】通过转换建模思路，将连续密度预测问题重构为离散区间分类任务。以木荷为例，基于经数据预处理与异常值剔除后保留的 481 组微钻阻力－密度匹配数据，采用基于机器学习法（LightGBM）的区间分类方法构建预测模型，并与多种传统回归模型进行系统性对比。该方法通过密度区间输出包容测量波动，从而提升模型的适应能力。【结果】所构建的区间分类模型整体准确率达 0.876，宏平均 F₁ 值为 0.878，在木材密度分级任务中表现稳定；传统回归模型的最优 R² 为 0.771，更适用于连续密度预测。两类模型任务定位不同，分类指标与回归指标不宜直接比较。该区间分类模型在各密度区间上均表现稳健，精确率范围为 0.769 ~ 1.000；对于木材分级实践中常需明确区分的高、低密度两端区间，模型分类正确率均达到或超过 84%，显示出良好的判别能力。结果表明，区间分类策略能够提供比传统点估计更稳健的分级结果。【结论】所提出并验证的“区间分类”建模策略，通过将连续密度预测重构为区间分类任务，有效提升了微钻阻力法预测木材密度的实用性。该策略输出的密度等级直观适用于工程快速判断，其分类概率输出也为科学研究提供了结构化数据。研究为解决微钻阻力法应用中模型不稳定的共性问题，提供了一种侧重于稳健性与可用性的机器学习建模新视角。

Abstract

[Objective] The micro-drilling resistance method is one of the principal techniques for non-destructive evaluation of wood density in standing trees. However, existing regression-based prediction models built upon this technique are susceptible to multi-source disturbances arising from instrumentation, operators, and environmental conditions, leading to poor stability and limited generalizability. These shortcomings constrain its application in engineering practice (e.g., wood grading and rapid field testing) and hinder reliable analysis in scientific research. To enhance the robustness and practical applicability of the micro-drilling resistance method for wood density prediction, this study aims to explore a modeling approach that accommodates measurement uncertainty. [Method] By transforming the modeling concept, the continuous density prediction problem was reconstructed as a discrete interval classification task. TakingSchima superba as an example, based on 481 sets of micro-drilling resistance–density paired data retained after data preprocessing and outlier removal, a predictive model was developed using an interval classification approach based on machine learning (LightGBM), and systematically compared with several conventional regression models. By outputting density intervals, this method accommodates measurement fluctuations, thereby enhancing the model’s adaptability. [Result] The interval classification model achieved an overall accuracy of 0.876 and a macro-averageF₁-score of 0.878, showing stable performance in wood density grading. Its precision across density intervals ranged from 0.769 to 1.000. For the high- and low-density intervals that are often critical for timber grading, the classification accuracy exceeded 84%, indicating good discriminative ability. These results suggest that the interval classification strategy can provide more robust grading results than conventional point estimation. [Conclusion] The proposed and validated “interval classification” modeling strategy, by reformulating continuous density prediction as an interval classification task, effectively enhances the practical applicability of the micro-drilling resistance method for predicting wood density. The density grades output by this strategy are intuitively suited to rapid engineering judgment, while its classification probability outputs provide structured data for scientific research. This study offers a new machine learning modeling perspective focused on robustness and usability，for addressing the common problem of model instability in the application of the micro-drilling resistance method.

关键词

微钻阻力法 / 木材密度 / 活立木 / LightGBM / 区间分类 / 木荷 / 有序分类 / 木材分级

Key words

micro-drilling resistance method / wood density / standing trees / LightGBM / interval classification / Schima superba / ordinal classification / wood grading

参考文献

[1] Ramírez Correa J A, Marín Vélez A M, Torres—Andrade P, et al. Influence of tree diameter and height on the physical and mechanical properties of Retrophyllum rospigliosii wood [J].Forests, 2025, 16(4): 593.

[2] Pinkowski G, Piernik M, Wołpiuk M, et al. Effect of chip thickness and tool wear on surface roughness and cutting power during up—milling wood of different density [J].BioResources, 2024, 19(4): 9234-9248.

[3] Apungu J, Antwi K, Appiah—Kubi E, et al. Investigating physical and mechanical properties of mahogany root wood (Khaya ivorensis) for its utilization [J].Drvna Industrija, 2025, 76(2): 189-199.

[4] Ali A C. Physical—mechanical properties and natural durability of lesser used wood species from Mozambique[D]. Uppsala: Swedish University of Agricultural Sciences,2011.

[5] De Souza Cardoso F, José Vinha Zanuncio A, Deroco Martins G, et al. Innovative methodology for determining basic wood density using multispectral images and MAPIR RGNIR camera [J].IEEE Access, 2025, 13: 166581-166592.

[6] 孙燕良, 张厚江, 朱磊, 等. 木材密度检测方法研究现状与发展 [J].森林工程, 2011, 27(1): 23-26.

[7] Sun Y L, Zhang H J, Zhu L, et al. Research and development of the method for testing wood density [J].Forest Engineering, 2011, 27(1): 23-26.

[8] [ 7 ] GB/T 1927.5—2021 无疵小试样木材物理力学性质试验方法 第 5 部分: 密度测定 [S].

[9] GB/T 1927.5—2021 Test methods for physical and mechanical properties of small clear wood specimens: part 5: determination of density [S].

[10] Gaitan—Alvarez J, Moya R, Berrocal A. The use of X—ray densitometry to evaluate the wood density profile of Tectona grandis trees growing in fast—growth plantations [J].Dendrochronologia, 2019, 55: 71-79.

[11] Charria D L G, Rodríguez V N, Oviedo M B S, et al. Estimación indirecta de densidad por el uso de “ pilodyn” en Eucalyptus pellita f. Muell [J]. Colombia Forestal, 2014, 17(2): 181.

[12] 姚建峰, 郭旭展, 符利勇, 等. 微钻阻力法间接测量木材密度 [J].林业科学, 2022, 58(9): 138-147.

[13] Yao J F, Guo X Z, Fu L Y, et al. Indirect measurement of wood density by micro drill resistance method [J].Scientia Silvae Sinicae, 2022, 58(9): 138-147.

[14] Nickolas H, Williams D, Downes G, et al. Application of resistance drilling to genetic studies of growth, wood basic density and bark thickness in Eucalyptus globulus [J].Australian Forestry, 2020, 83(3): 172-179.

[15] Isik F, Li B L. Rapid assessment of wood density of live trees using the Resistograph for selection in tree improvement programs [J].Canadian Journal of Forest Research, 2003, 33(12): 2426-2435.

[16] Downes G M, Lausberg M, Potts B M, et al. Application of the IML Resistograph to the infield assessment of basic density in plantation eucalypts [J].Australian Forestry, 2018, 81(3): 177-185.

[17] Cheng L T, Dai J, Yang Z G, et al. Variation of larch wood property indexes based on nondestructive testing data [J].BioResources, 2020, 15(2): 2906-2923.

[18] Cheng L T, Dai J, Yang Z G, et al. Theoretical and experimental research on moisture content and wood property indexes based on nondestructive testing [J].BioResources, 2020, 15(1): 1600-1616.

[19] Yao J F, Zhao Y B, Lu J, et al. Research on the wood density measurement in standing trees through the micro drilling resistance method [J].Forests, 2024, 15(1): 175.

[20] Gendvilas V, Downes G M, Lausberg M, et al. Predicting wood density using resistance drilling: the effect of varying feed speed and RPM [J].Forests, 2024, 15(4): 579.

[21] Gendvilas V, Lee D J, Kain D P, et al. Predicting wood density using resistance drilling: the effect of instrument and operator [J].Forests, 2024, 15(1): 157.

[22] Kabir S, Wagner C, Ellerby Z. Toward handling uncertainty—at—source in AI: a review and next steps for interval regression [J].IEEE Transactions on Artificial Intelligence, 2024, 5(1): 3-22.

[23] Ke G, Meng Q, Finley T, et al. LightGBM: A Highly Efficient Gradient Boosting Decision Tree[C]//Advances in Neural Information Processing Systems 30 (NIPS 2017). Curran Associates Inc., 2017: 3146-3154.

[24] 王云鹏, 张蕊, 周志春, 等. 10 年生木荷生长和材性性状家系变异及选择 [J].南京林业大学学报 (自然科学版), 2020, 44(5): 85-92.

[25] Wang Y P, Zhang R, Zhou Z C, et al. A variation and selection of growth and wood traits for 10—year—old Schima superba [J].Journal of Nanjing Forestry University (Natural Sciences Edition), 2020, 44(5): 85-92.

[26] 彭徐剑, 何晨, 詹庆斌. 3 种南方防火树种理化性质的试验研究 [J].林产工业, 2021, 58(3): 27-31.

[27] Peng X J, He C, Zhan Q B. Experimental research on physical and chemical properties of three fire—prevention tree species in South China [J].China Forest Products Industry, 2021, 58(3): 27-31.

[28] 韦昌幸. 48 年生木荷人工林生产力及经济效益评价研究 [D].长沙: 中南林业科技大学, 2014.

[29] Wei C X. Productivity analysis and economic benefit evaluation of48a Schima superba plantation [D]. Changsha: Central South University of Forestry & Technology,2014.

[30] 彭建, 鲁鑫钥, 余小晏, 等. 基于 Pilodyn 的楠木木材密度预测模型的建立 [J].西北林学院学报, 2023, 38(6): 269-275.

[31] Peng J, Lu X Y, Yu X Y, et al. Establishment of wood density prediction model of Phoebe zhennan based on pilodyn data [J].Journal of Northwest Forestry University, 2023, 38(6): 269-275.

[32] Fundova I, Funda T, Wu H X. Non—destructive wood density assessment of Scots pine (Pinus sylvestris L. ) using Resistograph and Pilodyn [J].PLoS One, 2018, 13(9): e0204518.

[33] Trachta M A, Acosta J J, Walker T D. Removing friction and bark improves drill resistance data for selection on wood density in Pinus taeda plantations [J].Forest Products Journal, 2024, 75(1): 26-33.

[34] Kobayashi Y. New precise model of studentized principal components [J].Communications in Statistics—Theory and Methods, 2024, 53(2): 487-504.

[35] Garthwaite P H, Elfadaly F G, Crawford J R. Modified confidence intervals for the Mahalanobis distance [J].Statistics & Probability Letters, 2017, 127: 131-137.

[36] 周飞燕, 金林鹏, 董军. 卷积神经网络研究综述 [J].计算机学报, 2017, 40(6): 1229-1251.

[37] Zhou F Y, Jin L P, Dong J. Review of convolutional neural network [J].Chinese Journal of Computers, 2017, 40(6): 1229-1251.

[38] 谈询滔, 顾依依, 阮彤, 等. 数据集分类可用性评估的置信区间方法 [J].计算机科学, 2019, 46(1): 78-85.

[39] Tan X T, Gu Y Y, Ruan T, et al. Confidence interval method for classification usability evaluation of data sets [J].Computer Science, 2019, 46(1): 78-85.

[40] DS/EN 338: 2016 Structural timber—strength classes [S].

基金资助

广东省林业局自然资源事务管理专项资金−生态林业建设 (省级组织实施) 项目(GDLZXY20220102)

广东省林业局自然资源事务管理专项资金−生态林业建设 (省级组织实施) 项目(GDLZXY20230102)

广东省林业局 2021 年省级生态公益林效益补偿资金省统筹经费项目()