1.School of Remote Sensing and Geomatics Engineering,Nanjing University of Information Science and Technology,Nanjing,Jiangsu 210044,China
2.Meteorological Disaster Prevention and Environmental Meteorology Center of Hebei Province,Shijiazhuang,Hebei 050021,China
3.Shijiazhuang Meteorological Bureau of Hebei Province,Shijiazhuang,Hebei 050081,China
Show less
文章历史+
Received
Published
2025-07-04
2026-09-10
Issue Date
2026-06-11
PDF (4935K)
摘要
目的 基于传统森林火灾深度学习监测模型对小型火源实时精确监测的局限性,提出一种改进YOLOv8n模型的森林火灾监测模型YOLOv8n_MB。 方法 在YOLOv8n的基础上,引入混合局部通道注意力(MLCA)模块,增强模型对小火源关键特征的敏感度;通过加权双向特征金字塔(BiFPN)实现跨尺度高效特征融合,以提升小目标监测能力,最终建立森林火灾监测模型YOLOv8n_MB。设计消融试验验证各改进模块的有效性,并结合CAM可视化技术及热力图重合度(intersection over Union of CAM,IoU_CAM)、显著性能量比(energy ratio,ER)、最大响应点命中率(pointing-game,PG)等量化指标分析模型,验证“模型关注区域与真实火点”的一致性。 结果 与YOLOv8n相比,YOLOv8n_MB模型的识别精确率(Precision)及平均精确度均值mA@0.5和mA@0.5:0.95分别提高2.8,3.7和2.1个百分点,召回率(Recall)下降了0.2个百分点,火点识别能力优于YOLOv8其他系列模型。改进后的YOLOv8n_MB模型在保持计算效率的同时显著提升了监测精度,能够快速、准确监测地面火源。消融试验表明,与原始YOLOv8n模型相比,YOLOv8n_MB模型的参数量减少0.19 MB,平均精确度(average precision,AP)提升2.4个百分点,具备更优的森林火灾监测性能和较低部署成本。CAM可视化与量化分析显示,YOLOv8n_MB在大范围、小目标及复杂场景下的林火识别能力均优于其他模型。 结论 YOLOv8n_MB模型实现了精度与轻量化的平衡,具备良好的小目标森林火灾监测能力。
Abstract
Objective To address the limitations of traditional deep learning models in the real-time and accurate monitoring of small forest fire sources,this study proposes an improved YOLOv8n-based forest fire detection model,named YOLOv8n_MB. Method Based on the YOLOv8n architecture,a Mixed Local Channel Attention (MLCA) module was introduced to enhance the model’s sensitivity to key features of small fire sources.In addition,a Weighted Bidirectional Feature Pyramid Network (BiFPN) was employed to achieve efficient cross-scale feature fusion,thereby improving the small-target detection capability.The resulting forest fire monitoring model,named YOLOv8n_MB,was further evaluated through ablation experiments to verify the effectiveness of each improved module.Moreover,Class Activation Map (CAM) visualization,together with quantitative metrics such as the intersection over union of CAM (IoU_CAM),Energy Ratio (ER),and Pointing Game (PG),was used to analyze the consistency between the model’s attention regions and actual fire locations. Result Compared with YOLOv8,the proposed YOLOv8n_MB model achieved improvements of 2.8,3.7,and 2.1 percentage points in detection precision,mAP@0.5, and mAP@0.5:0.95,respectively,with a decrease of 0.2 percentage points in Recall.This demonstrated its superior accuracy in fire spot detection over other YOLOv8 variants.The improved YOLOv8n_MB model significantly enhanced detection precision while maintaining computational efficiency,enabling fast and accurate monitoring of ground fire sources.Ablation experiments showed that,compared with the original YOLOv8n,YOLOv8n_MB reduced the parameters by 0.19 MB and increased the Average Precision (AP) by 2.4 percentage points,demonstrating superior forest fire detection performance and lower deployment cost.Furthermore,CAM visualization and quantitative analysis revealed that YOLOv8n_MB exhibited stronger forest fire recognition capability across large-scale,small-target,and complex scenarios than other models. Conclusion The YOLOv8n_MB model achieves a balance between detection accuracy and lightweight design,demonstrating strong capability on small-target forest fire monitoring.
YOLOv8是ultralytics公司于2023年1月推出的目标监测模型,是一个全新的SOTA(state of the art )模型[13],骨干网络采用梯度流丰富的C2f结构[14],头部网络(head)采用无锚点(Anchor-Free)模式[15],损失函数(loss)计算采用任务对齐分配器(Task-aligned assigner)样本分配策略,以有效地提升模型精度[16],并已在农林业中得到了探索应用[17-18]。按照模型规模的不同,YOLOv8分为n、s、l、m、x等5个类型,其中YOLOv8n模型规模小、参数数量少、计算量低,适合在资源受限的设备上运行。YOLOv8网络的结构如图1所示。
1.2 MLCA注意力机制
MLCA(mixed local channel attention)是一种轻量级混合局部通道注意力机制,能够实现性能和复杂性平衡[19],其结构如图2所示。输入的特征图分别被局部平均池化(LAP)[20]和全局平均池化(GAP)[21]进行局部和全局特征捕捉;池化后的特征通过卷积层(Conv1d)进行压缩,同时保持空间维度不变;局部池化的特征在重新排列后通过“乘法”操作与原始输入特征进行结合,强化对有用特征的关注;全局特征经过重新排列后,通过“加法”操作与局部池化特征进行结合,融合全局上下文信息;经过局部和全局注意力处理的特征图再次通过反池化(UNAP)[22]操作,恢复到原始的空间维度[23]。
使用精确率(precision,P)、召回率(recall,R)、平均精确度(average precision,AP)、平均精确度均值mAP@0.5和mAP@0.5:0.95、参数量(params)、浮点运算次数(floating-point operations per second,FLOPs)、模型大小(models-size)和帧速率(frames per second,FPS)等指标对深度学习模型进行评价。其中精确率指的是在所有被模型识别为目标的样本中,被正确地识别为目标的样本所占的比例,精确率越高代表模型的误检越少,反之代表误检越多。其计算公式如下:
参数量(params)用于衡量模型复杂度及其对内存的占用情况:浮点运算次数(FLOPs)为推理过程中,模型需要执行的浮点运算数量,其值越小,代表模型对计算资源需求越低。模型大小(models-size)指训练好的模型权重文件所占用的存储空间;帧速率(frames per second,FPS)指模型每秒能够处理并完成推理的图像数量。
MAOS H, WANGW D.Research review of YOLO series object detection algorithms based on deep learning[J].Journal of Yan’an University (Natural Science Edition),2024,43(2):88-95.
[8]
ALKHATIBA A A.A review on forest fire detection techniques[J].International Journal of Distributed Sensor Networks,2014,10(3):597368.
[9]
QIANJ J, LINH F.A forest fire identification system based on weighted fusion algorithm[J].Forests,2022,13(8):1301.
[10]
冷跻峰.基于改进YOLOv7的森林火灾检测算法[J].大众科技,2024,26(6):8-12.
[11]
LENGJ F.Forest fire detection algorithm based on improved YOLOv7[J].Popular Science & Technology,2024,26(6):8-12.
YEM L, ZHOUH Y, LIJ J.Forest fire detection algorithm based on an improved Swin Transformer[J].Journal of Central South University of Forestry & Technology,2022,42(8):101-110.
LINH F, MAY C, JIANGL,et al.Experimental design and implementation of small target forest fire detection based on deep learning[J].Journal of Nanjing Forestry University (Natural Sciences Edition),2025,49(5):235-241.
[16]
SULTANT, CHOWDHURYM S, SAFRANM,et al.Deep learning-based multistage fire detection system and emerging direction[J].Fire,2024,7(12):451.
ZHANGX W, ZHANGF Q.Visible forest fire detection using SBP-YOLOv7[J].Journal of Nanjing Forestry University (Natural Sciences Edition),2025,49(3):103-109.
YINJ Y, HER R, ZHAOF J,et al.Research on forest fire monitoring based on multi-source satellite remote sensing images[J].Spectroscopy and Spectral Analysis,2023,43(3):917-926.
QIY J, MAJ L, HANL P,et al.Lightweight maize seedling weed identification model based on improved YOLOv8s[J].Journal of Northwest A&F University(Natural Science Edition),2025,53(11):58-70.
LÜS C, MAB L, SONGL,et al.Nighttime detection me-thod of polymorphic Camellia oleifera fruits based on YOLON network[J].Journal of Northwest A&F University(Na-tural Science Edition),2023,51(8):141-154.
[32]
WAND H, LUR S, SHENS Y,et al.Mixed local channel attention for object detection[J].Engineering Applications of Artificial Intelligence,2023,123:106442.
[33]
PANDAM K, SHARMAA, BAJPAIV,et al.Encoder and decoder network with ResNet-50 and global average feature pooling for local change detection[J].Computer Vision and Image Understanding,2022,222:103501.
[34]
WANGQ L, XIEJ T, ZUOW M,et al.Deep CNNs meet global covariance pooling:better representation and generalization[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2021,43(8):2582-2597.
[35]
SHIX Y, LVF S, SENGDW,et al.Visualizing and understanding graph convolutional network[J].Multimedia Tools and Applications,2021,80(6):8355-8375.
[36]
IBRAHIME, SHERIDANT, MANDAVILLIS.Mesonephric like-carcinoma (MLCA) of the uterus with a predominant spindled morphology:a case report[J].American Journal of Clinical Pathology,2021,156(S1):76.
[37]
ZHENGX, CHENF, LOUL M,et al.Real-time detection of full-scale forest fire smoke based on deep convolution neural network[J].Remote Sensing,2022,14(3):536.
[38]
WANGS K, DONGQ, CHENX Q,et al.Measurement of asphalt pavement crack length using YOLO V5-BiFPN[J].Journal of Infrastructure Systems,2024,30(2):04024005.
[39]
SHAFIQM, GUZ Q.Deep residual learning for image recognition:a survey[J].Applied Sciences,2022,12(18):8972.
[40]
ZHANGZ Y, WANGZ,JOE I.CAM-NAS:an efficient and interpretable neural architecture search model based on class activation mapping[J].Applied Sciences,2023,13(17):9686.