Modelling chamise fuel moisture content across California: a machine learning approach

Live fuel moisture content plays a significant and complex role in wildfire propagation. However, in situ historical and near real-time live fuel moisture measurements are temporally and spatially sparse within wildfire-prone regions. Routine bi-weekly sampling intervals are sometimes exceeded if the weather is unfavourable and/or field personnel are unavailable. To fill these spatial and temporal gaps, we have developed a daily gridded chamise (Adenostoma fasciculatum) live fuel moisture product that can be used, in conjunction with other predictors, to assess current and historical wildfire danger/behaviour. Chamise observations for 52 new- and 41 old-growth California sites from the National Fuel Moisture Database were statistically related to dynamically downscaled high-resolution weather predictors using a random forest machine learning model. This model captures reasonably well the temporal and spatial variability of chamise live fuel moisture content within California. Compared with observations, model-predicted live fuel moisture values have an overall R-2, root mean squared error (RMSE) and bias of 0.79, 15.34% and 0.26%, respectively, for new growth and 0.63, 8.81% and 0.11% for old growth. Given the success of the model, we have begun to use it to produce daily forecasts of chamise live fuel moisture content for California utilities.