林业科学研究2025，Vol.38Issue(6)：68-79,12.DOI:10.12403/j.1001-1498.20240520
基于AutoGluon-Tabular的不同龄组杉木人工林碳储量最优估算模型研究
Optimal Carbon Stock Estimation Models in Cunninghamia lanceolata Plantations of Different Age Groups Using AutoGluon-Tabular
摘要
Abstract
[Objective]Traditional methods for estimating forest carbon stocks often suffer from inefficien-cies,poor model adaptability,and lack of age-structure responsiveness,To address these limitations,this study investigated multi-aged Cunninghamia lanceolata plantations in Fujian Province,by combing un-manned aerial vehicle Light Detection and Ranging(LiDAR)point cloud data with Automated Machine Learning(AutoML)technology.The objective was to develop age-group-specific carbon stock estimation models to support high-precision and efficient carbon sink dynamic monitoring.[Methods]UAV LiDAR was used to acquire point cloud data,and individual trees were segmented with features extracted using the Dalponte2016 algorithm.The optimal carbon stock estimation models for different age groups were ex-plored using the AutoGluon-Tabular framework under various feature combinations,and their accuracy was compared with traditional Random Forest and Gradient Boosting Regression Tree(GBRT)models.[Results](1)The Dalponte2016 algorithm demonstrated good segmentation performance across all age groups,with mean values of tree detection rate,segmentation precision,and overall accuracy exceeding 0.800.(2)The carbon stock estimation models based on the AutoGluon-Tabular framework exhibited signi-ficant variations across age groups and feature groups.LightGBMLarge_BAG_L1(morphological feature group)achieved the highest R2(0.856)for young forests.LightGBMLarge_BAG_L1 based on the vegeta-tion morphological feature group,with an R2 of 0.856;for middle-aged forests,ExtraTreesMSE_BAG_L1(combined vegetation and point cloud features)reached an R2 0.913;for near-mature forests,Random-ForestMSE_BAG_L1(combined features)reached an R2 of 0.838;for mature forests and over-mature forests,CatBoost_BAG_L1(morphological features)attained R2 values of 0.673 and 0.822,respectively.Compared to the Random Forest and GBRT models,AutoGluon-based optimal models improved R2 val-ues by about 0.24,reduced RMSE by about 36.43%,and lowered MAE by about 37.68%.[Conclusion]The integration of UAV LiDAR data with AutoML technology simplifies the traditionally complex parameter-tuning process and significantly enhances the accuracy of forest carbon stock estimation.The proposed approach provides an efficient and reliable technical solution for regional forest carbon stock dynamic mon-itoring and management,providing valuable insights into the sustainable development of smart forestry.关键词
杉木/碳储量/UAV-LiDAR/单木分割/AutoGluonKey words
Chinese fir/carbon stock/UAV-LiDAR/individual tree segmentation/AutoGluon分类
农业科技引用本文复制引用
赵婧雯,陈科,周润发,刘金福,陈博,洪宇,康开权,朱建琴..基于AutoGluon-Tabular的不同龄组杉木人工林碳储量最优估算模型研究[J].林业科学研究,2025,38(6):68-79,12.基金项目
福建省林业科技项目"武夷山国家公园碳储量高光谱遥感长期动态监测技术研究与示范"(2022FKJ26) （2022FKJ26）
福建省林业科技项目"双碳背景下福建省高碳汇林业实现路径及维持技术"(2022FKJ02) （2022FKJ02）
福建省科技创新重点项目"湿地碳库生态安全动态评价关键技术研发及应用"(2021G02007) （2021G02007）
