I did a quick search and found these two articles:
“From public fire dataset and Internet, we collect 7000 images for training and 4494 images for test, and then run experiments with the comparison of four baseline methods including deep neural network, support vector machine based on scale-invariant feature transform feature, stack auto-encoder and deep belief network.”
You’re going to need academic access to get at the original article.
This article directly links to eleven (11) datasets.
The references on these articles should get you into the various academic work that has already been done.
My first thought for ‘original’ content is YouTube (which you would have to scrape, but would get you a lot of aerial photography) and the Alert Wildfire camera network: http://www.alertwildfire.org/ (which would get you a lot of ‘non fire’ images for comparison). Then Instagram, with an honorable mention to https://www.instagram.com/slvsteve/