MAFF-YOLO:面向造林验收的明穴目标检测模型  

MAFF-YOLO:a target detection model for planting holes in afforestation acceptance

作　　者：石燕妮 王武魁[1] 吴明晶 张大兴 廉瑞峰 谷亚宇 Shi Yanni;Wang Wukui;Wu Mingjing;Zhang Daxing;Lian Ruifeng;Gu Yayu(School of Economics and Management,Beijing Forestry University,Beijing 100083,China;Fujian Jiangle State Owned Forest Farm,Sanming 353300,Fujian,China)

机构地区：[1]北京林业大学经济管理学院,北京100083 [2]福建省将乐国有林场,福建三明353300

出　　处：《北京林业大学学报》2025年第4期142-154,共13页Journal of Beijing Forestry University

基　　金：城市科技与精细化管理项目(Z221100005222108);福建省将乐国有林场基于无人机影像的造林监管自动化系统研建(20220517)。

摘　　要：【目的】为解决传统林场造林验收过程中存在的主观性强、缺乏科学性以及管理人员不足等问题,本研究提出一种基于单阶段目标检测框架的造林明穴检测模型MAFF-YOLO,旨在自动识别并统计造林明穴的数量和位置,推动造林验收的数字化转型,提高验收效率和科学性。【方法】基于YOLOv8模型,通过多方面改进获得了MAFF-YOLO。首先,采用MobileNetV4作为主干网络,引入更多参数和层次结构提高检测精度;其次,添加基于归一化的注意力模块(NAM),增强对明穴特征的捕捉能力,减少误检和漏检;然后,将特征融合模块替换为跨尺度特征融合模块(CCFM),在降低特征图拼接计算量的同时整合不同尺度特征,提升对小尺度明穴的检测能力;接着,将检测头替换为RFAHead,根据数据的复杂性和重要性动态调整感受野,增强网络对不同输入特征的适应性;最后,优化边界框损失函数为FocusCIoU,改善正负样本分布不平衡问题,增强对关键样本的学习能力。【结果】MAFF-YOLO在识别明穴数量和位置方面表现出较高的准确性。与基础YOLOv8模型相比,其精度提高了1个百分点,mAP50提高了0.7个百分点,F_(0.5)提高了0.6个百分点,且算法复杂度显著降低。【结论】在相同实验条件下,MAFF-YOLO相较于其他现有方法,在提升模型对造林明穴目标的检测效果方面表现出显著优势,并已成功集成至端到端的检测系统中,为造林验收的数字化提供了技术支持,进一步提升了造林验收的效率和科学性。[Objective]To solve the problems like strong subjectivity,lack of scientific basis,and insufficient personnel in traditional afforestation acceptance,this paper proposes an afforestation hole detection model of MAFF-YOLO.It aims to automatically identify and count the number and location of afforestation holes,promoting the digital transformation of afforestation acceptance and improving efficiency and scientific accuracy.[Method]Based on the YOLOv8 model,MAFF-YOLO was obtained through multiple improvements.First,it used MobileNetV4 as the backbone network to increase parameters and layers,enhancing detection accuracy.Second,it added a normalization-based attention module(NAM)to better capture hole features and reduce false detections.Third,it replaced the feature fusion module with a cross-scale feature fusion module(CCFM),which integrated features of different scales and reduced computational load,improving detection of small holes.Fourth,it replaced the detection head with an RFAHead,which dynamically adjusted the receptive field based on data complexity and importance,thereby enhancing adaptability to different input features.Finally,the bounding box loss function was optimized to FocusCIoU to address sample imbalance and improve learning capability for key samples.[Result]MAFFYOLO demonstrated high accuracy in identifying the number and location of planting holes.Compared with basic YOLOv8 model,its precision increased by 1 percentage point,mAP50 by 0.7 percentage points,and F_(0.5) by 0.6 percentage points.Moreover,the algorithm complexity was significantly reduced.[Conclusion]Under the same experimental conditions,MAFF-YOLO shows significant advantages over other existing methods in improving the detection performance of afforestation holes.It has been successfully integrated into an end-to-end detection system,providing effective technical support for the digitalization of afforestation acceptance and further enhancing the efficiency and scientific nature of the acceptance process.

关 键 词：小目标检测 YOLOv8 算法 数字化造林验收 无人机 MobileNetV4 NAM CCFM 

分 类 号：S757[农业科学—森林经理学] TP391[农业科学—林学] TP79[自动化与计算机技术—计算机应用技术]