Forest Fire Detection Using Classifiers and Transfer Learning

Application of fire detection is gaining a lot of attention due to the increasing threat from global warming that causes a lot of economic distress and threat to public safety. There are over 200,000 forest fires each year which destroys a total area of about 3.5-4.5 million sq. km [1]. The current detection methods involve the use of sensors whose data usually depends on pressure and temperature of the environment. Often these conditions are met when the fire gets out of control. This paper demonstrates various methods to detect fire from photos of the site using transfer learning and image processing. The methods demonstrated in the paper uses about 3500 images of forest fires and 1900 images of forests with no fire to train the models. These methods can detect fire at initial stages compared to existing methods of fire detection. Using only cameras instead of multiple sensors reduces the cost and makes this system more efficient. In this paper we show the comparison of different convolution neural network (CNN) based pre-trained models used for the feature extraction of the images via transfer learning.