Remote Sensing of Live Fuel Moisture Content Using Machine Learning

Fuel moisture is one of the most uncertain inputs into wildfire rate of spread models and fire danger products. The dead fuel moisture has been studied extensively and mainly depends on environmental conditions. In contrast, live fuel moisture, apart from environmental conditions, depends on seasonal changes (green-up, dormancy, curing, etc.), evapotranspiration, and other long-term weather conditions which are difficult to model. The live fuel moisture is critical in the fire danger assessment as well as the fire spread modeling. For that reason, live fuel moisture observations remain critical to assess the spatial and temporal variability in the fuel moisture of live fuels. Direct observations of live fuel moisture are labor-intensive and time-consuming. The live fuel data collection requires destructive plant sampling (clipping a part of the plant), weighing the sample, oven drying, and then weighing the dry sample again. The difference between the mass of the fresh sample and the dried one is used to estimate the water content, which normalized by the dry mass gives the live fuel moisture content. For that reason, the direct live fuel moisture observations are infrequent (typically once or twice a month) and with poor spatial coverage. The live fuel moisture is commonly estimated using satellite data to alleviate these limitations. There are two main strategies when estimating live fuel moisture from satellite data. Either using a physical model based on a Radiative Transfer Model (RTM) or an empirical model based on statistical fitting of live fuel moisture observations. The first approach uses satellite data to resolve the physics associated with the water content in the fuel requiring detailed characteristics of the plant, which makes it difficult to generalize for different types of plants. On the other hand, the second approach is much more simplistic. Although it does need to handle the problem of sparse observations, it does not require complex plant characteristics. In this work, a new live fuel moisture model leveraging a supervised machine-learning method based on the live fuel moisture observations from the National Fuel Moisture Database (NFMDB), satellite, weather, and ancillary data is presented. The satellite data include all Landsat-8 surface reflectance, as well as temperature bands, and microwave wavelengths from all Sentinel-1 SAR bands. The weather data is obtained from WRF-based simulations, and the ancillary input contains multiple static layers from the LANDFIRE database. We test the presented model based on the ground live fuel moisture observations. Our preliminary results show the potential of such data-driven live fuel moisture models leveraging machine learning.