Answers are grounded in this paper and cite their sources.
Wildfires cause severe ecological and economic damage, and limiting it requires rapid and accurate detection. This study presents an image classification framework for wildfire detection in RGB images from the FLAME 3 dataset, captured by an unmanned aerial vehicle. The preprocessing pipeline applies gamma correction, sharpening, contrast-limited adaptive histogram equalisation (CLAHE) and class-balanced data augmentation. The proposed model combines a vision transformer (ViT) and MobileNetV2 in parallel branches, fusing global contextual features with local spatial features. Explainable artificial intelligence techniques (Grad-CAM, Grad-CAM++ and LIME) are used to visualise the image regions that influence the predictions. The model is compared with nine convolutional and five transformer-based baselines and evaluated with accuracy, Cohen’s κ, the Matthews correlation coefficient (MCC), the critical success index (CSI) and related metrics. Trained with the AdamW optimiser, the ViT–MobileNetV2 model achieves a test accuracy of 98.44%, a macro-average F1-score of 0.984 and a recall of 1.000 for the fire class on the FLAME 3 RGB classification subset. Inference takes 10–15 ms per image on an NVIDIA T4 GPU in half precision and 50–70 ms on an x86 CPU with INT8 quantisation. The results apply to the evaluated dataset and experimental protocol.
No takes yet. Share an insight, caveat, or question.
Rohan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: