Objective In response to the increasingly frequent forest fires that threaten ecological security, developing a lightweight real-time intelligent monitoring system is of crucial importance for enhancing ecological risk prevention and control. However, existing fire-detection methods tend to be easily affected by environmental disturbances, and it remains difficult to balance model complexity with real-time performance. Therefore, this study aims to develop an efficient lightweight detection model that can be trained from scratch without relying on external pre-trained weights.

Methods We first constructed and publicly released a high-quality forest wildfire monitoring dataset containing more than 10 000 images. On this basis, we propose FireLight-YOLO, a lightweight architecture based on YOLOv8. The model incorporates Ghost Convolution to compress redundant computation and introduces the FasterC2fBlock, which integrates partial convolution and point-wise convolution to form a T-shaped receptive field, thereby strengthening key region perception. Additionally, we optimize the SPPF module and propose a Feature Pyramid Shared Convolution (FPSC) mechanism to achieve efficient cross-scale feature fusion. Comprehensive evaluations, including cross-validation, independent testing, ablation studies, and multi-noise robustness experiments, were conducted to assess performance.

Results Without using any pre-trained weights, FireLight-YOLO achieves an mAP@0.5 of 0.491 with only 2.26 × 10⁶ parameters and 5.9 GFLOPs. The model effectively balances accuracy, lightweight design, and real-time performance. Compared with the original YOLOv8, computational cost is reduced by 2.2 GFLOPs, parameter quantity decreases by 25%, inference speed improves by 15%, and the model demonstrates strong robustness in complex interference scenes.

Conclusion FireLight-YOLO achieves accurate wildfire detection under lightweight constraints. This work provides a low-cost and efficient solution for intelligent wildfire monitoring, and its lightweight design significantly enhances deployment adaptability on mobile edge devices. The proposed approach offers promising support for forest ecosystem protection and restoration.