Machine Learning for Fuel Moisture in Time and Space, Dead or Live

The WRF-SFIRE modeling system integrates a high-resolution atmospheric model with a fire spread model using the level set method and a dead fuel moisture model. Recently, we developed a method that pre-trains a redundant Recurrent Neural Network (RNN) to serve as a solver for the differential equations of a base physical model. The RNN is then trained using standard methods to translate input time series of environmental data, such as temperature, relative humidity, solar radiation, and wind, into time series of fuel moisture data. The input data can be augmented with static information, such as geography and land use, as well as satellite data streams. In prediction mode, the RNN uses a time series of weather forecasts, along with static data and location, to produce an estimate of the time series of fuel moisture values.

The method was previously applied to dead fuel moisture, but can also be applied to live fuel moisture, which evolves on a time scale of weeks to seasons, rather than hours to hundreds of hours like dead fuel moisture. The inclusion of satellite data streams is crucial in capturing the spatial variability in geography and plant cover species. The complexity and redundancy of the initial RNN must increase to enable capturing a larger hidden state with a more complicated dynamic. This method can be considered as adding a time component to classical machine learning methods in remote sensing, which then play the role of feature engineering.

We will update on the progress of using this method to create surrogates for existing physical models and to create machine learning models from observational data.