无人机巡检是检测输电通道山火的重要方法, 但无人机拍摄的初期山火图像存在有效像素少和背景干扰等问题, 导致山火检测精度较低。该文研究了无人机视角下输电通道初期山火图像特征检测和分析方法。首先, 结合部分公开数据集, 建立了输电通道初期山火图像数据集, 分析了明火和烟雾2类典型初期山火图像特征; 其次, 针对YOLOv11n基线模型存在目标选择、特征提取和定位能力不足等问题, 将多尺度特征模块和注意力模块引入骨干网络, 并将卷积加法标记混合器嵌入具有2个卷积层的跨阶段局部瓶颈模块; 最后, 提出了适用于输电通道初期山火图像的检测算法模型YOLOv11n-MCC。结果表明: 初期山火图像具有目标尺寸小和分布广等特点; 相较于YOLOv11n基线模型, YOLOv11n-MCC模型的初期山火检测的精确度、召回率和mAP50的平均精度分别提高了9.0、3.8和5.7个百分点, 参数量和每秒十亿次浮点运算分别下降了0.149×10⁶和0.2。该文研究结果可为输电线路初期火灾检测提供参考。

Objective: Unmanned aerial vehicle (UAV) inspection technology has become an essential tool for monitoring wildfires in transmission line corridors due to its flexibility, cost-effectiveness, and adaptability to remote and complex terrains. However, current research lacks initial wildfire image datasets from the perspective of UAV inspections, and existing algorithms struggle to accurately identify early-stage wildfires. This study constructs an initial wildfire image dataset from a UAV perspective and proposes the YOLOv11n-MCC detection model for transmission line corridors. Methods: Through the field inspections of national transmission line corridors for 174 h, covering 82 000 km with UAVs, 1 308 transmission line corridor inspection images were obtained, and a transmission line corridor inspection image dataset from the perspective of the UAV was constructed. This dataset was analyzed to identify the characteristics of the initial wildfire images. The early wildfire detection model, YOLOv11n-MCC, based on an enhanced version of YOLOv11n, was then proposed. First, part of the traditional convolutional network in YOLOv11n was replaced with multi-scale feature convolution(MFConv) to reduce computational load while improving feature extraction. Second, the C2PSA module in the backbone network was replaced with the spatialand channel synergy attention(SCSA) mechanism to improve target localization. Finally, C3k2_ABlock with convolutional additive token mixer(CATM) at its core was embedded to improve target representation and selection in complex backgrounds. Results: MFConv, SCSA attention, and C3k2_ABlock with CATM sequentially improved the YOLOv11n-MCC model's ability to detect targets in complex scenes. Comparative experiments revealed that the YOLOv11n-MCC model significantly outperforms the YOLOv11n baseline model in terms of accuracy, mAP50, parameter count, and giga floating-point operations per second(GFLOPS) for early mountain fire small-target detection, making it portable but still computationally efficient. Specifically, precision increased by 9.0 percentage point, recall by 3.8 percentage point, and mAP50 by 5.7 percentage point. In addition, the number of parameters and GFLOPS decreased by 0.149×10⁶ and 0.2, respectively. The YOLOv11n-MCC architecture achieves enhanced multiscale feature representation while maintaining reduced computational complexity, thereby improving operational efficiency without compromising detection performance. Conclusions: The image dataset for transmission line corridor inspection constructed in this study can effectively support the training and testing of the improved fire detection algorithm. The proposed YOLOv11n-MCC model demonstrates stable performance in detecting small initial mountain fire targets and can be effectively applied to real-time wildfire detection in transmission line corridors using UAVs, thereby providing essential technical support for early wildfire detection. Future work will focus on examining the influence of varying smoke-to-fire ratios on model training to further enhance wildfire detection accuracy.