Wed, Aug 04, 2021:On Demand
Background/Question/Methods
Land Management agencies in the Intermountain West use information about regional variation in fuel loads to optimally allocate firefighting resources at the start of the fire season. While static maps work well for large fuel classes, fine fuels, the driver of rangeland fires, vary dramatically from year-to-year. Our goal was to develop a fine fuels forecast to help fire managers anticipate spatial variation in fuel loads before the start of the fire season. We compiled a historical record of fine fuel loads reported to the Great Basin Coordination Center and combined it with the Rangeland Analysis Platform (RAP) herbaceous productivity dataset. Using a Bayesian State Space approach, we built a process model where the latent fuel at a location depends on fuel in the previous year and the current year’s productivity. The state-space model structure allowed us to estimate latent fuel loads across the Intermountain West with separate process and observation error terms. We then forecasted RAP productivity using remotely sensed data available on Google Earth Engine in early spring with a linear regression model to create productivity forecasts. Finally, we forecast fuel loads based on forecasted productivity and the outputs of our state-space process model, and quantified the associated uncertainty.
Results/Conclusions Estimating latent fine fuel loads using these two datasets had high process uncertainty (13% of mean) and the fuel carryover term’s posterior distributions indicates the previous year’s fuel load contributes substantially (78%+-6%) to the following year’s fuel load compared to contribution of the current year’s productivity (22% +-6% ). Our forecasts of RAP productivity have low mean absolute predictive error (0.20) based on hindcasts from 1987-2019, and outperform (6%) a null model built using only spatial data and autoregressive productivity data. While process uncertainty remains a significant limitation of our fuels forecast, our attempt illustrates challenges in ecological forecasting, mismatches between field and remotely sensed data, and lessons for future efforts to use widely available remotely sensed data products to answer ecological management questions.
Results/Conclusions Estimating latent fine fuel loads using these two datasets had high process uncertainty (13% of mean) and the fuel carryover term’s posterior distributions indicates the previous year’s fuel load contributes substantially (78%+-6%) to the following year’s fuel load compared to contribution of the current year’s productivity (22% +-6% ). Our forecasts of RAP productivity have low mean absolute predictive error (0.20) based on hindcasts from 1987-2019, and outperform (6%) a null model built using only spatial data and autoregressive productivity data. While process uncertainty remains a significant limitation of our fuels forecast, our attempt illustrates challenges in ecological forecasting, mismatches between field and remotely sensed data, and lessons for future efforts to use widely available remotely sensed data products to answer ecological management questions.