A Deep Learning Framework for UAV-Based Wildfire Detection Using Transfer Learning on the FLAME Dataset

Abstract

Wildfires are one of the most devastating natural catastrophes on an international level with a level of ecological, economic, and other humanitarian impacts that have never been experienced before. The traditional fire detection systems, which are based on ground-based sensor networks and satellite images, have severe drawbacks such as limited spatial resolution, atmospheric scattering, and intolerable response time. The given paper includes a detailed deep learning architecture of aerial wildfire detection with the use of Unmanned Aerial Vehicles (UAVs). The suggested system combines a Deep Neural Network (DNN) to extract hierarchies of features with a fine-tuned Convolutional Neural Network (CNN) to classify binary images (fire vs. non-fire) with the help of transfer learning with ImageNet-pretrained networks. The experiments are performed on the FLAME (Fire Luminosity Airborne-based Machine learning Evaluation) dataset that consists of paired RGB and infrared (IR) aerial images of 48,010 labeled frames. To deal with the natural imbalance in the number of classes and the variability of the environment, a strict pre-processing pipeline has been implemented, such as image normalization, resizing, contrast-limited adaptive histogram equalization (CLAHE) and data augmentation. Various backbone architectures (VGG16, ResNet50, InceptionV3, MobileNetV2, and DenseNet121) are trained and tested on the transfer learning paradigm, and ensemble strategies are tested to maximize performance. The suggested combination of ResNet50 and InceptionV3 attains 99.12% accuracy, 98.97% precision, 99.34% recall, and an F1-score of 99.15% on the FLAME test set, which is significantly better than single-model baselines and previous art. A critical analysis reveals such inexorable issues as domain shift, constraints on real-time edge deployment, and representativeness of datasets, as well as tangible future research directions. The findings prove the suggested framework as a feasible, deployable solution to early wildfire detection with the help of UAVs.