Wildfire Detection Using Vision Transformers with The wildfire detection

Abstract

Wildfire in the U.S., especially in California, are increasing in frequency and intensity, with 130 deaths recorded in 2023, the highest since 1990. The January 2025 Los Angeles wildfires burned 40,000 acres and 12,000 buildings, highlighting the urgent need for advanced detection systems. Vision Transformers (ViTs) offer a powerful deep learning approach for early wildfire detection by analyzing complex image data with high accuracy. However, challenges such as real-time data availability, sensor limitations, and environmental factors like smoke hinder detection. Additionally, training deep learning models is computationally expensive, with concerns over false positives and scalability. In this study, a pre-trained ViT model (vit_base_patch16_224) was fine-tuned on a 10.74 GB wildfire dataset for binary classification. Using PyTorch, images were resized, converted into tensors, and normalized. The model, trained with Adam optimizer and CrossEntropyLoss, achieved 96.10% accuracy, demonstrating ViT’s effectiveness in real-time wildfire detection and postevent analysis.