Forecasting dead fuel moisture content below forest canopies – A seven-day forecasting system

Abstract

Accurate forecasting of forest fuel moisture is critical for decision making for bushfire risk and prescribed burning. In-situ dead fuel moisture content (DFMC) monitoring (fuelsticks) has improved significantly, along with improvements in weather forecasting and spatial representation of forest density. Machine learning (ML) models have also out-performed traditional fuel moisture estimation approaches on open sites, however, these models are yet to be tested on a diverse range of below-canopy conditions using above-canopy weather observations. Even with significant advancements, forecasting DFMC has shown little improvement, as there are notable spatial and temporal problems associated with DFMC forecasting below forest. This research develops and validates a below canopy, 7-day-ahead forecasting system of daily minimum forest fuel dryness (10-h DFMC) that integrates an automated fuel sensor network, gridded weather forecasts, landscape attributes and a ML model (Gradient boosting algorithm; LightGBM). The study area was established across a diverse range of 28 sites in south-eastern Australia, producing the largest below canopy validation of its kind. Fuel moisture was measured half-hourly using 10-hour automated fuelsticks, with five years of observations. The model performance was evaluated on its capacity to predict minimum daily DFMC, and when DFMC conditions were within the burnable (9% – 16% DFMC) and high risk (<9% DFMC) ranges. Long-term sites were validated on a years’ worth of observations, assessing seasonal variability. The complete network of sites showing best performance in the first day of forecast (for both datasets mean R² of 0.88 and 0.87; RMSE of 6.06% and 6.07%), with degraded performance to day seven (mean R² of 0.63 and 0.52; RMSE of 11.84% and 13.33%). The results demonstrate that accurate DFMC forecasts can be achieved by the newly developed forecasting framework. The proposed system has the potential to be applied in any wildland fire setting where weather forecasts are available.