A 32-year species-specific live fuel moisture content dataset for southern California chaparral

Live fuel moisture content (LFMC) strongly affects the behavior of wildland fire, resulting in its incorporation into wildfire spread models and danger ratings. In this dataset, over ten thousand LFMC observations were combined with predictor variables from Landsat imagery and the Weather Research and Forecasting model to train species-specific random forest models that predict the LFMC of four fuel types—chamise, old growth chamise, black sage, and bigpod ceanothus. These models are then utilized to create a historical, 32-year long, LFMC dataset in southern California chaparral. Additionally, the high spatial and temporal sampling frequency of chamise allowed for quantile mapping bias correction to be applied. The final chamise output, which is the most robust, has a mean absolute error of 9.68% and an R2 value of 0.76. The LFMC dataset successfully captures the variability in the annual cycle, the spatial heterogeneity, and the interspecies differences, which makes it applicable for better understanding varying fire season characteristics and landscape level flammability.

Methods
LFMC observations were acquired as the model predictands. Then, predictors were calculated, including long-term lag variables, such as 90-day precipitation, 90-day mean temperature, and 150-day mean insolation, as well as short-term lag or instantaneous predictors, such as day length, 7-day mean soil moisture, and near-infrared reflectance of vegetation (NIRv). Species-specific random forest models were then trained and tested with two techniques: 5-fold cross validation using all observations and observation site specific cross validation, where each site was individually withheld as the test data. After testing, models were trained on all predictors/predictands and used to create the species-specific LFMC datasets. Lastly, quantile mapping bias correction was applied to the modeled chamise LFMC dataset, which was possible due to the high spatial frequency of observation sites.

活燃料含水率（Live Fuel Moisture Content, LFMC）对野外火灾行为具有显著调控作用，因此被纳入野火蔓延模型与火灾危险等级评估体系。本数据集整合了超一万条LFMC观测数据与来自Landsat影像及天气研究与预报模型的预测变量，用于训练针对四类燃料类型的专属随机森林模型，以预测查帕拉尔灌丛（chamise）、老龄查帕拉尔灌丛、黑鼠尾草（black sage）以及大果美洲茶（bigpod ceanothus）的LFMC。随后基于这些模型生成了南加州查帕拉尔灌丛地区长达32年的历史LFMC数据集。此外，得益于查帕拉尔灌丛样本的高时空采样密度，可对其实施分位数映射偏差校正。其中表现最稳健的查帕拉尔灌丛LFMC最终产品的平均绝对误差为9.68%，决定系数（R²）为0.76。该LFMC数据集可有效捕捉年周期变化、空间异质性与物种间差异，能够为深入解析不同火灾季节特征及景观尺度可燃性提供支撑。

方法
本研究以LFMC观测数据作为模型预测目标。随后计算得到各类预测变量，涵盖长期滞后变量（如90天累积降水量、90天平均气温及150天平均日照辐射）与短期滞后或瞬时变量（如日照时长、7天平均土壤湿度及植被近红外反射率（NIRv））。随后采用两种技术训练并测试物种专属随机森林模型：一是基于全部观测数据的5折交叉验证，二是观测站点专属交叉验证，即单独将每个站点的观测数据作为测试集留出。模型测试完成后，基于全部预测变量与预测目标重新训练模型，并用于生成各物种专属的LFMC数据集。最后，由于观测站点具备较高的空间采样密度，可对建模得到的查帕拉尔灌丛LFMC数据集应用分位数映射偏差校正方法。