Wildfires pose a threat to naure and human health especially in the age of global warming. Their early detection is key to an effective fighting, because once a wildfire reaches a certain size it can be hardly controlled.

Consequently, automated wildfire detection systems have been developed to aid rangers in their work to prevent wildfire hazards. One of the most successful sysems was developed by the German aerospace center DLR (Deutsches Zentrum für Luft- und Raumfahrt) and relies on sophisticated hand-crafted feature. The algorithm behind this system is called fshell.

Convolutional neural networks (CNN) are specialized artificial neural networks that are the state-of-the-art for image recognition tasks. It is therefore investigated if these trainable models can be applied to the wildfire dectection problem to improve the performance of existing systems. Besides CNNs that use spatial features only, C3D networks that can also extract temporal features are considered. To tackle the detection of smoke in a larger image, a selective search variant was developed, that preselects regions of observable motion, which are then classified by the actual models.

The results clearly show that the trained networks learned to distinguish wildfire smoke from other objects, although they do not reach the fshell's performance. However, fshell makes use of additional prior knowledge whereas the CNNs rely on image data only. The results also show that C3D networks are the best performing single models, suggesting that the use of temporal features is important for accurate detection.