Wildforest Fire Detection with ShRe-Xception Network on Aerial Optical and Infrared Images.

It is commonly accepted that wildforest fires are one of the major natural disasters worldwide. This destruction has a crucial impact to fire fighters and civilian lives, animal extinction, and forest degradation. This work focuses on preventing potentially extensive disasters by developing a detection system that can classify images of forests in "Fire" and "non-Fire" cases with input from both RGB and Infrared (IR) cameras. Our design uses both an existing small – architecture of Xception network (small-arch Xception) and the proposed Short Recursive (ShRe)-Xception. We first use transfer learning on the small-arch Xception network with input from either RGB or IR images. We also train the proposed ShRe-Xception network by using all images from the very beginning. Datasets used for training and testing purposes contain "Fire" and "non-Fire" images of forests that are captured from UAVs and uploaded in IEEE portal. When testing the above models, the estimated accuracy was RGB: 77.72% / IR: 96.83% after transfer learning on small-arch Xception, RGB: 84.86% / IR: 29.18% when initially training the small-arch Xception, and RGB: 90.10% / IR: 99.31% after initially training the ShRe-Xception model. This makes the latter model the most accurate one for images of both spectrums in our experiments. The ShRe-Xception can potentially play a vital and efficient role on the real-time fire detection during aerial surveillance on wild-forests.