Quantifying fuel load over large areas is essential to support integrated fire management initiatives in fire-prone regions to preserve carbon stock, biodiversity and ecosystem functioning. It also allows a better understanding of global climate regulation as a potential carbon sink or source. Large area assessments usually require data from spaceborne remote sensors, but most of them cannot measure the vertical variability of vegetation structure, which is required for accurately measuring fuel loads and defining management interventions. The recently launched NASA's Global Ecosystem Dynamics Investigation (GEDI) full-waveform lidar sensor holds potential to meet this demand. However, its capability for estimating fuel load has yet not been evaluated. In this study, we developed a novel framework and tested machine learning models for predicting multi-layer fuel load in the Brazilian tropical savanna (i.e., Cerrado biome) using GEDI data. First, lidar data were collected using an unnamed aerial vehicle (UAV). The flights were conducted over selected sample plots in distinct Cerrado vegetation formations (i.e., grassland, savanna, forest) where field measurements were conducted to determine the load of surface, herbaceous, shrubs and small trees, woody fuels and the total fuel load. Subsequently, GEDI-like full-waveforms were simulated from the high-density UAV-lidar 3-D point clouds from which vegetation structure metrics were calculated and correlated to field-derived fuel load components using Random Forest models. From these models, we generate fuel load maps for the entire Cerrado using all on-orbit available GEDI data. Overall, the models had better performance for woody fuels and total fuel loads (R² = 0.88 and 0.71, respectively). For components at the lower stratum, models had moderate to low performance (R² between 0.15 and 0.46) but still showed reliable results. The presented framework can be extended to other fire-prone regions where accurate measurements of fuel components are needed. We hope this study will contribute to the expansion of spaceborne lidar applications for integrated fire management activities and supporting carbon monitoring initiatives in tropical savannas worldwide.

量化大面积的燃料负荷对于支持火灾易发地区的综合火灾管理举措以保护碳储量、生物多样性和生态系统功能至关重要。它还可以更好地了解作为潜在碳汇或碳源的全球气候调节。大面积评估通常需要来自星载遥感器的数据，但其中大多数无法测量植被结构的垂直变化，而这是准确测量燃料负荷和定义管理干预所必需的。最近推出的 NASA 全球生态系统动力学调查 (GEDI) 全波形激光雷达传感器具有满足这一需求的潜力。然而，其估计燃料负荷的能力尚未得到评估。在这项研究中，我们开发了一个新颖的框架并测试了机器学习模型，用于使用 GEDI 数据预测巴西热带稀树草原（即塞拉多生物群落）中的多层燃料负荷。首先，使用未命名的飞行器 (UAV) 收集激光雷达数据。飞行是在不同的塞拉多植被地层（即草原、热带稀树草原、森林）中的选定样本地块上进行的，在那里进行了实地测量以确定地表、草本、灌木和小乔木、木本燃料和总燃料负荷的负荷。随后，从高密度无人机激光雷达 3-D 点云模拟 GEDI 类全波形，从中计算植被结构指标，并使用随机森林模型将其与现场衍生的燃料负载分量相关联。从这些模型中，我们使用所有在轨可用 GEDI 数据生成整个塞拉多的燃料负荷图。总体而言，这些模型在木质燃料和总燃料负荷（R²  = 0.88 和 0.71，分别）。对于较低层的组件，模型具有中到低的性能（R ²介于 0.15 和 0.46 之间），但仍显示出可靠的结果。所提出的框架可以扩展到其他需要准确测量燃料成分的火灾易发地区。我们希望这项研究将有助于扩展星载激光雷达在综合火灾管理活动中的应用，并支持全球热带稀树草原的碳监测计划。