基于深度学习的森林火灾隐患识别定位方法研究

Research on Forest Fire Hazard Detection and Localization Method Based on Deep Learning

摘要: 针对森林场景光照变化剧烈、目标遮挡严重、早期火情目标尺度微小、识别难度大等问题，以轻量化YOLOv8n模型为基础，引入MLPA多层级感知注意力机制、C2f_D轻量化特征融合模块及WIoU边界框回归损失函数，构建改进YOLOv8n森林火灾隐患识别定位模型。以某森林保护区多源影像为数据源，通过制备烟雾、火焰两类目标样本集，设计消融实验，同步构建单一模块改进模型及多模块协同改进模型，验证了各单一改进模块的独立贡献及多模块协同优化的增益效应；同时利用多类型复杂场景开展对比实验，验证了改进YOLOv8n模型在小目标火情、目标遮挡、白云干扰、复杂背景、微弱烟雾及光照干扰等复杂场景下，均可准确识别烟雾及火焰目标，模型抗干扰能力与泛化能力显著提升，为森林火灾早期智能预警与自动化防控提供了可靠技术方案。

Abstract: To address challenges such as drastic lighting variations, severe target occlusion, small-scale early fire targets, and high recognition difficulty in forest environments, this study proposes an improved YOLOv8n model for forest fire hazard detection and localization. Building upon the lightweight YOLOv8n architecture, the model incorporates the MLPA multi-level perception attention mechanism, the C2f_D lightweight feature fusion module, and the WIoU bounding box regression loss function. Using multi-source imagery from a forest conservation area, we construct datasets of smoke and flame targets, design ablation experiments, and simultaneously develop single-module enhanced models and multi-module collaborative optimization models to validate the individual contributions of each improvement and the synergistic benefits of combined modules. Furthermore, comparative experiments conducted across diverse and complex scenarios demonstrate that the improved YOLOv8n model can accurately detect smoke and flames under challenging conditions including small targets, object occlusion, cloud interference, complex backgrounds, faint smoke, and lighting disturbances. The model exhibits significantly enhanced robustness and generalization capability, providing a reliable technical solution for intelligent early warning and automated prevention of forest fires.