Exploring State-of-the-art models for Early Detection of Forest Fires

TL;DR

The paper tackles early forest-fire detection from visual data, addressing data scarcity by generating a synthetic dataset of smoke plumes using a game engine and augmenting with real images. It benchmarks four state-of-the-art detectors (YOLOv7, YOLOv7-tiny, DETR, Deformable DETR) on the combined dataset for both classification and localization tasks. The results reveal a trade-off: Deformable DETR yields higher precision and mAP but at higher computational cost and lower recall, while YOLOv7-tiny offers the best balance of accuracy and edge-friendly inference speed. The work demonstrates a practical pathway to deploy continuous 24/7 forest-fire warnings on embedded hardware, by prioritizing lightweight models like YOLOv7-tiny and leveraging synthetic data for early detection.

Abstract

There have been many recent developments in the use of Deep Learning Neural Networks for fire detection. In this paper, we explore an early warning system for detection of forest fires. Due to the lack of sizeable datasets and models tuned for this task, existing methods suffer from missed detection. In this work, we first propose a dataset for early identification of forest fires through visual analysis. Unlike existing image corpuses that contain images of wide-spread fire, our dataset consists of multiple instances of smoke plumes and fire that indicates the initiation of fire. We obtained this dataset synthetically by utilising game simulators such as Red Dead Redemption 2. We also combined our dataset with already published images to obtain a more comprehensive set. Finally, we compared image classification and localisation methods on the proposed dataset. More specifically we used YOLOv7 (You Only Look Once) and different models of detection transformer.

Figures (5)

• Figure 1: Challenges in forest fire detection. (a) Orange colorization of the sky at sunset. (b) Ambiguity between smoke, clouds and fog. (c) Small smoke plume at a very distant location. (d) Identifying smoke at night without thermal imaging.
• Figure 2: Sample images from dataset including samples of both synthetic, augmented and real images.
• Figure 3: Confusion matrices for the four models show us the number of True and False Positives and True and False Negatives.
• Figure 4: Inference results for classification. (GT: Ground truth, S: Smoke, F: Fire, N: Normal).
• Figure 5: Inference results for localization.