Enhancing forest fire detection in satellite imagery using vision transformers techniques

Forest fires pose significant threats to ecosystems and human life. Early detection and effective intervention of fires are crucial for minimizing damage and preventing loss of life. Due to limitations of traditional fire detection methods, deep learning-based approaches offer promising solutions for automatically detecting fires from satellite imagery. This thesis investigates the implementation and performance of Vision Transformer (ViT) models for forest fire detection. Using the Canada Forest Fires Dataset, a custom CNN model, ResNet50, EfficientNetB4, EfficientNetB7, a ViT model trained from scratch, and a pretrained ViT model ("vit base-patch16-224-in21k") were trained and tested. The experimental results demonstrate that the pretrained ViT model achieved the highest performance among all models, with an accuracy of 99.22% and an F1 score of 0.99, in detecting forest fires. ViT models' ability to model long-range dependencies and their lower inductive bias provides a significant advantage in fire detection. EfficientNetB4 and EfficientNetB7 models also achieved high accuracy and F1 scores, but their training times were longer compared to ViT models. The custom CNN and ResNet50 models exhibited lower performance compared to ViT and EfficientNet models. This study shows that ViT technology holds great potential for the development of forest fire detection and monitoring systems. ViT models can contribute to more effective fire intervention and damage minimization by providing high accuracy and fast detection. Future studies can focus on exploring different ViT architectures, data augmentation techniques, and hybrid models to achieve even higher performance in forest fire detection.