Early Wildfire Detection Using YOLOv8 Nano Lightweight Classification

Abstract

The development of rapid, portable early detection technologies is necessary due to the increasing frequency and intensity of wildfires. In order to detect wildfires, this paper suggests a lightweight binary classification framework based on YOLOv8 microcls for edge devices with limited resources. The model achieves 96.02% overall accuracy on a test set of 402 images (156 fire, 246 no fire), with per-class F1 scores of 94.84% (fire) and 96.76% (no fire), after training on the Wildfire Dataset (2n version) for 30 epochs at 320x320 resolution. Three primary failure modes are identified by a thorough error analysis of the sixteen incorrectly classified samples: misreading of distant little fires, smoke-occluded flames (44.4% of false negatives), and sunset/sunrise chromatic confusion (42.9% of false positives). An analysis of public data for YOLOv5n, EfficientNet B0, ResNet 50, and MobileNetV2 assessed on data for ResNet 50, MobileNetV2, EfficientNet B0, and YOLOv5n, each assessed on its own original dataset, positions YOLOv8n cls favorably on the accuracy-efficiency tradeoff. With only 3.9 million parameters and inference speeds exceeding 100 FPS on GPU, the proposed system is architecturally suitable for real-time deployment on Raspberry Pi and IP camera networks, pending direct edge hardware validation. Targeted mitigation strategies, including dynamic thresholding, multi-frame temporal consistency, and seasonal domain adaptation, are proposed to reduce the false negative rate below 3% and the false positive rate below 2%.