AI-Enhanced Autonomous Drones for Early Detection of Forest Fires in Finnish Boreal Regions

Abstract

Ecosystems, infrastructure, and human safety are all at risk, as wildfires can cause extensive damage to them. Early detection is therefore recognized as a crucial component of effective disaster management. In this dissertation, the application of deep learning is explored for classifying wildfire images, with the objective of developing an autonomous long-term solution for early detection and monitoring using AI-controlled drones. A total of five models were trained and tested on the FlameVision dataset: DenseNet121, InceptionV3, EfficientNetB3, ResNet152V2, and a modified Convolutional Neural Network (CNN). Data preprocessing and analysis procedures were employed, and an offline augmentation approach was maintained to increase the dataset size threefold. Model performance was measured in terms of accuracy, loss, and area under the curve (AUC) across training, validation, and test sets. Findings indicated that the pre-trained networks consistently outperformed the custom CNN. DenseNet121 and ResNet152V2 achieved perfect classification performance, while InceptionV3 and EfficientNetB3 exceeded 99% accuracy. Despite high training accuracy, the custom CNN exhibited limited robustness during validation, highlighting the critical role of transfer learning in achieving stable and accurate detection. Generalization across models was further improved through data augmentation, confirming its importance in compensating for dataset limitations and accelerating model convergence. Overall, the results demonstrate that state-of-the-art deep learning architectures are capable of achieving highly accurate wildfire detection. This research underscores the potential for integrating such models into UAV-based surveillance systems in real time, thereby enhancing early warning systems, supporting firefighting operations, and mitigating the growing impacts of global wildfires.