Remotely sensed live fuel moisture content across the Western USA and its relationship with climate aridity
Abstract
Forest wildfires are increasing in size and frequency in the Western USA. Our ability to accurately forecast wildfire risk is diminished by the lack of objective estimates of live fuel moisture content (LFMC) - the amount of moisture per unit biomass in vegetation. Current subjective estimates of LFMC derived from climate and weather may introduce significant uncertainties in the quantification of ignition probability and fire spread. In this study, we combine a recurrent neural network with remotely sensed synthetic aperture radar backscatter (from Sentinel-1) and optical frequency reflectances (from Sentinel-2) to create wall-to-wall maps of LFMC at 500 m resolution across the Western United States. Measurements from the National Fuel Moisture Database (NFMD) are used as training data. Our model showed a strong ability to predict LFMC (R2 = 0.73, RMSE = 21.0) when compared against NFMD sites not used in training (cross-validation). This performance is on par with or exceeds previous models (R2 ∈ [0.60, 0.72], RMSE ∈ [17.1, 27.1]), despite the fact that our study covered a much larger extent and diversity of vegetation types. Our model predicted spatial variability of LFMC with very high accuracy (R2 = 0.85, RMSE = 9.9), but temporal deviance from the mean with lower (but still highly significant) accuracy (R2 = 0.54, RMSE = 21.0).
Using the resultant wall-to-wall maps of LFMC (Fig. 1), we analyze spatio-temporal contrasts in LFMC and climate-based aridity indices. In particular, we test the hypothesis that climate-based metrics alone do not fully capture vegetation wetness characteristics. We then investigate the reasons behind the disagreement between vegetation wetness and climate metrics, with an eye towards sources of variation in plant drought strategies that lead to differing rates of water uptake. Understanding the spatio-temporal characteristics of LFMC can lead to better fuel availability estimates and thus improved wildfire risk forecasts.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMNH52A..03R
- Keywords:
-
- 3390 Wildland fire model;
- ATMOSPHERIC PROCESSES;
- 1817 Extreme events;
- HYDROLOGY;
- 4315 Monitoring;
- forecasting;
- prediction;
- NATURAL HAZARDS;
- 4323 Human impact;
- NATURAL HAZARDS