Optimization of deep neural network architectures for the forest fire detection

Abstract

Forests are critical for all living creatures in vital issues such as biodiversity conservation, the sustainability of the ecosystem, and the continuation of the water and nutrient cycle. Forest fires are large uncontrolled incidents that burn and quickly spread through wild landscapes. According to research, average temperatures are on the rise forest fires seem to appear more often and more destructive. Therefore, detecting forest fires as early as possible and responding quickly have become essential to toward potentially deadly disasters. Various studies are conducted in the literature on the detection of forest fires. Besides more traditional image processing and machine learning methods, deep learning-based methods have been frequently employed in the recent literature. Despite their exceptional performances, it has been a challenge to utilize resulting prediction models in edge and low-power computational devices due to their massive sizes and computational complexity. However, there are no studies in the literature addressing these issues for detecting forest fires. On the other hand, there are well-established optimization techniques to make deep neural networks less power-hungry and more computationally efficient. This thesis aims to bridge this gap by utilizing such optimization techniques for their use in deep learning-based forest fire detection models. In this thesis, we aim to develop efficient as well as accurate deep learning models for detecting forest fires. Firstly, a dataset is created by using the public data in the literature. Then, ResNet101, RegNet-32X-GF, DenseNet-169 and Inceptionv3 networks were trained to detect forest fires. All methods had comparable and satisfactory performance reaching over 95% test accuracy. Then, the resulting models are optimized and compared using various quantization and pruning methods. DenseNet showed better resiliency against accuracy drop and size reduction compared to the other methods. Furthermore, the classification problem extended to object detection to explore further the potential of the proposed approach. Faster R-CNN was employed with the DenseNet backbone and optimized with previously determined optimal hyperparameters. Then, it was compared to the state-of-the-art optimized YOLOv5 and showed not only higher accuracy but also lower size. Therefore, this study shows that the employed methods and followed methodology has a promising potential to bring the power of deep learning-based techniques to the edge and low-power devices for the forest fire detection problem.