Hi all, lead author here. Thanks for the discussion. I'm more of an ecologist than a machine learning specialist (though we do have a couple computer scientists and statisticians as co-authors), so it's interesting to see your perceptions
Yes, basically we turned a 1-in-3-odds problem into the odds of a coin flip, on average. What failed to be included in the article from our email interview was that we still find this encouraging and worthy of sharing for a couple reasons. The performance is best for the largest fire group, which cause about 90% of the destruction. The model can "catch" 65% of the fires that do become large (recall), with a precision of 53%. This allows us to do good enough to identify a small subset of fires that will account for a majority of the burned area (for example, 40% of fires accounting for almost 80% of the burned area, etc by Figure 7). That could absolutely be useful information for triaging efforts. In some of these cases, dozens of fires are breaking out in a day or two, and managers need to make decisions faster than is possible by running full fire spread models on every ignition.
We certainly would encourage future research to use more complex input information to increase this accuracy (our learning curve analysis suggests that the number of fires in the dataset isn't what's limiting accuracy), in order to increase its usability for managers on the ground. In essence, we're setting up a new framework that is especially relevant for areas where fire frequency is increasing due to climate change, and hopefully others can build off it.
Fire prediction is inherently a really difficult and chaotic problem, so it's interesting that just a couple simple variables can explain even this much of the variance from the time of ignition, especially when many of these big fires burn for weeks or months.
TL;DR: Far from perfect, but useful (and scientifically interesting) nonetheless