基于YOLOv8的陕西黄土高原刺槐林枯立木识别  

Identification of standing dead trees in typical Robinia pseudoacacia plantation in the Loess Plateau of Shaanxi Province based on YOLOv8

作　　者：张利 高晓东[2] 张志博 张旭 赵西宁[2] ZHANG Li;GAO Xiaodong;ZHANG Zhibo;ZHANG Xu;ZHAO Xining(School of Water Conservancy and Architectural Engineering,Northwest Agricultural and Forestry University,Yangling 712100,China;Institute of Soil and Water Conservation,Northwest Agricultural and Forestry University,Yangling 712100,China)

机构地区：[1]西北农林科技大学水利与建筑工程学院,杨凌712100 [2]西北农林科技大学水土保持研究所,杨凌712100

出　　处：《林业工程学报》2024年第4期112-121,共10页Journal of Forestry Engineering

基　　金：国家重点研发计划(2021YFD1900700);国家杰出青年科学基金(42125705);陕西省杰出青年科学基金(2021JC-19)。

摘　　要：黄土高原在生态恢复的过程中由于不适宜造林加之极端干旱频发,导致人工刺槐林出现退化甚至死亡,从而产生枯立木。为准确识别陕西黄土高原刺槐林枯立木,进而分析其空间分布格局,本研究以典型人工林刺槐为对象,沿陕西黄土高原降水梯度设置6个样地,利用无人机获取刺槐林高分辨率RGB影像,结合多个深度学习目标检测模型识别刺槐林枯立木,通过对比与分析证明了YOLOv8模型在检测效率以及检测精度方面的优越性。通过对刺槐林数据集训练与测试以及刺槐林RGB影像的拼接,从而分析刺槐林枯立木的分布格局,用枯立木指数(SDTI)反映刺槐林退化程度,同时分析了坡向对其影响。YOLOv8模型训练后平均帧率FPS达到了68.7帧/s,且其验证集平均精度均值(mAP)达到了94.6%,F1得分为0.92,均好于其他模型。综合不同目标检测模型的各项主要评价指标及泛化能力,选择YOLOv8模型分析刺槐林枯立木分布格局,表明陕西黄土高原刺槐林SDTI随降水梯度递减呈递增趋势;不同坡向SDTI呈阳坡>半阳坡>半阴坡>阴坡的变化规律。本研究可为黄土高原刺槐林退化程度评估及其合理的经营管理提供科学依据。During ecological restoration on the Loess Plateau,unsuitable afforestation efforts combined with frequent occurrences of extreme drought have precipitated the degradation and occasionally the demise of Robinia pseudoacacia plantation,resulting in the presence of standing dead trees.This study aimed at precisely identifying and analyzing the spatial distribution patterns of standing dead trees in the R.pseudoacacia plantation of Shaanxi Province.To achieve this,six sample plots were established across a precipitation gradient on the Loess Plateau,and high-resolution RGB imagery of these R.pseudoacacia plantation was captured using drones.This research utilized a variety of deep lear-ning object detection models to identify standing dead trees effectively.Among these models,YOLOv8 stood out for its detection efficiency and accuracy.This study involved training and testing the R.pseudoacacia plantation's dataset and integrating RGB images to examine the distribution patterns of standing dead trees.The distribution pattern of standing dead trees in R.pseudoacacia plantation was analyzed,and the degradation degree of R.pseudoacacia plantation was reflected by standing dead trees'index(SDTI),which also facilitated the analysis of the impact of slope orientation.After training,the YOLOv8 model achieved an average frames per second(FPS)of 68.7 frames/s,with a validation set average precision mean(mAP)of 94.6%and an F 1 score of 0.92,which were all superior to other models exa-mined.A thorough assessment of key performance metrics and the generalizability of various object detection models guided the choice of YOLOv8 for an in-depth examination of standing dead trees distribution patterns in R.pseudoacacia plantation.The analysis indicated a negative correlation between the SDTI and the precipitation gradient in the R.pseudoacacia plantation of the Shaanxi Loess Plateau.Additionally,SDTI differences were observed in relation to slope aspect,displaying a sequence from sunny to semi-sunny,semi-shady,and shady slopes.This s

关 键 词：黄土高原 刺槐林 深度学习 无人机遥感 枯立木指数(SDTI) YOLOv8 

分 类 号：S792.27[农业科学—林木遗传育种]