DRONES CONVOLUTIONAL NEURAL NETWORKS AND VISION TRANSFORMER OF EARLY DETECTION OF WILDFIRES

Abstract

Wildfire detection is critical in reducing environmental and economic losses resulting from uncontrolled forest fires. The proposed paper develops a UAV-based wildfire detection system incorporating CNN, ViT, and FL to enable distributed, privacy-preserving, and efficient real-time monitoring. In our proposed system, deep learning models are trained locally on UAVs, with updates aggregated through FL, hence saving data transmission while guaranteeing confidentiality. Extensive experiments on the Kaggle wildfire dataset have yielded considerable improvements in detection accuracy and generalization. Accordingly, the model obtained a training accuracy of 98.5% and a validation accuracy of 96.5%, with robust performance at low loss and very minimal overfitting. The comparative study against previous work confirms that the proposed system outperforms the traditional centralized and non-federated models in terms of both accuracy and scalability, opening new prospects toward smart real-time wildland fire surveillance with UAV networks.