Although this may have been posted in jest, it would probably work quite well or at least be an interesting experiment. There's a large corpus of images in bot

Although this may have been posted in jest, it would probably work quite well or at least be an interesting experiment. There's a large corpus of images in both categories, and it's a binary classifier which keeps things simple. There are numerous image cues that people have identified such as the flap angle relative to the wings, absence of smoke from the tyres, absence of heat haze from the jet wash, wheel-spin and so on. Depending on angle you could also get the landing markers on the runway. It seems reasonable that a convnet would identify these features. My only concern is that features like heat haze and wheel-spin might require very high resolution inputs to the net. Also it's quite hard to search explicitly for "planes taking off" as you get images from aircraft about to take off, aircraft taxiing, etc. Finally it's quite hard to find publicly available images, e.g. most photos on Aviation.net are copyrighted.