Deep Neural Networks for Wild Fire Detection and Monitoring with UAV

Vishal Gewali, Sanjeeb Prasad Panday

Abstract: Unmanned Aerial Vehicles (UAVs) which are equipped with camera sensors are easy means for making observations in remote, inconvenient and inaccessible areas and assist to apprehend the situation for many emergencies and disaster management applications. With the increase in the frequency and the severeness of the forest fire in the recent times, UAVs become the cost-effective means to provide high resolution images in wildfire detection in comparison to other techniques such as satellite and CCTV Cameras. This chapter is focused in the use of two different variations of CNN architectures models of VGG (VGG16, VGG19) and GoogleNet (InceptionV3, Xception) in developing models to correctly classify the forest fire and evaluating their performance. The models are analyzed in different UAV video footages using the Grad-CAM algorithm over the heat-maps to determine how well the model is working in the differentiating the given sample. Using the Adaboost optimizer, all models have shown the accuracy of over 96% and InceptionV3 model is found to have the better performance in comparison to the other VGG models, and also has shown slightly better outcomes against its latter version, Xception.