The model implements the FireCast architecture as described in this paper by [Radke et al., 2019].
The model generates only one prediction for a given input as opposed to perhaps a more generative approach: as a result, it requires many predictions for each fire.
To train your own copy of the model as described in firecast.py, you should configure the data loader to your own specifications. The dataset and visualization notebooks (availability pending) should give you a clearer picture of the data used, and how to select features for your own model. In particular, this line governs which LANDFIRE attributes are used:
landfire_attrs = ('SLP', 'Sparse', 'Tree', 'Shrub', 'Herb')The training loop and model architecture should account for any changes made in the data loader, but I didn't test this extensively.