Privileged knowledge distillation for efficient fire classification in resource-constrained wildfire monitoring

Abstract

Wildfire is a major threat in today’s world which is increasing day by day because of climate change. Thus, efficient wildfire detection is crucial to mitigate the economic, social and environmental loss. In order to contribute to studies related to wildfire, we suggest Unmanned Aerial Vehicles (UAV) with multimodal sensors like both RGB and IR cameras. However, the cost of IR cameras and thermal sensors are high but it is needed for night vision. To mitigate this problem, we proposed a privileged knowledge distillation method. The novelty of this technique is that it is trained with a heavy CNN teacher model and all the information of both RGB and IR can be mimicked by the RGB only student model. Thus, it is both cost effective and deployable on UAVs. The experimental results are satisfactory with 94% accuracy for the teacher model and an f1 score of 0.91 for student mode. This result is possible for accurate pre-processing techniques that include contrast stretching, CLAHE for increasing the intensity of the image and median filtering to remove noise from the image collected from FLAME-2 dataset. Thus, the work establishes a promising technique for detecting wildfire.