DataSheet1_Forest emissions reduction assessment from airborne LiDAR data using multiple machine learning approaches.docx

Objective: This study aims to evaluate the accuracy of different modeling methods and tree structural parameters extracted from airborne LiDAR for estimating carbon emissions reduction and assess their reliability as Certified Emission Reduction (CER) assessment techniques.
Methods: LiDAR data was collected from an afforestation project in Beijing, China. Various modeling methods, including statistical regression and machine learning algorithms, were used to estimate biomass and carbon emissions reduction. The models were evaluated under two schemes: tree-species-specific modeling scheme (Scheme 1) and all-sample modeling scheme (Scheme 2) using cross-validation and compared with ground-based estimations and pre-estimated emission reductions.
Results: Totally, the biomass estimation models in scheme 1 showed better accuracy than scheme 2. In scheme 1, The Random Forest (RF) and Cubist models achieved the highest prediction accuracy (R2 = 0.89, RMSE = 22.87 kg, CV RMSE = 52.00 kg), followed by GDBT and Cubist, with SVR and GAM performing the weakest. In scheme 2, Cubist model had the highest accuracy (R2 = 0.75, RMSE = 33.95 kg, CV RMSE = 36.05 kg), followed by RF and GBDT, with SVR and GAM performing the weakest. LiDAR-based estimates of carbon emissions reduction were closer to ground-based estimations and higher than pre-estimated values.
Conclusion: This study demonstrates that LiDAR-based models using tree structural parameters can accurately assess carbon emissions reduction. The models outperformed traditional methods in terms of cost and accuracy. Considering tree species in the modeling process improved the accuracy of the models. LiDAR technology has the potential to be a reliable assessment technique for carbon emissions reduction in forestry projects. The pre-trained models can be used for multiple predictions, reducing the cost of carbon sink surveys. Overall, LiDAR-based models provide a promising approach for assessing carbon emissions reduction and can contribute to mitigating climate change.

研究目标：本研究旨在评估不同建模方法，以及从机载激光雷达（airborne LiDAR）中提取的树木结构参数，在估算碳排放减少量时的准确性，并评估其作为核证减排量（Certified Emission Reduction, CER）评估技术的可靠性。
研究方法：本研究采集了中国北京某造林项目的机载激光雷达数据。采用包括统计回归与机器学习算法在内的多种建模方法，对生物量与碳排放减少量进行估算。通过交叉验证，分别基于树种专属建模方案（方案1）与全样本建模方案（方案2）对模型进行评估，并将模型结果与地面实测估算值及预估算减排量进行对比。
研究结果：整体而言，方案1下的生物量估算模型精度优于方案2。在方案1中，随机森林（Random Forest, RF）与Cubist模型取得了最高的预测精度（决定系数R²=0.89，均方根误差RMSE=22.87 kg，交叉验证均方根误差CV RMSE=52.00 kg），紧随其后的为梯度提升决策树（Gradient Boosting Decision Tree, GDBT），支持向量回归（Support Vector Regression, SVR）与广义可加模型（Generalized Additive Model, GAM）表现最差。在方案2中，Cubist模型精度最高（R²=0.75，RMSE=33.95 kg，CV RMSE=36.05 kg），紧随其后的为RF与GDBT，SVR与GAM表现最差。基于机载激光雷达的碳排放减少量估算结果更贴近地面实测估算值，且高于预估算减排量。
研究结论：本研究证实，利用树木结构参数构建的机载激光雷达模型可精准评估碳排放减少量。该模型在成本与精度两方面均优于传统方法。在建模过程中纳入树种因素可有效提升模型精度。机载激光雷达技术有望成为林业项目碳排放减少量评估的可靠技术手段。预训练模型可用于多场景预测，降低碳汇调查的成本。总体而言，基于机载激光雷达的模型为碳排放减少量评估提供了极具前景的技术路径，可为气候变化减缓工作提供重要支撑。