基于无人机影像与OBIA-CFS算法的荒漠草原鼠洞斑块识别  

Identification of rodent hole patches in desert grasslands using UAV imagery and OBIA-CFS algorithms

作　　者：戚桂美 郁志宏[2] 单艳敏 田彦军 QI Guimei;YU Zhihong;SHAN Yanmin;TIAN Yanjun(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010022,Inner Mongolia,China;College of Mechanical and Electrical Engineering,Inner Mongolia Agricultural University,Hohhot 010010,Inner Mongolia,China;Forestry and Grassland Bureau of Inner Mongolia,Hohhot 010050,Inner Mongolia,China;Forestry and Grassland Bureau of Otog Banner,Ordos 016199,Inner Mongolia,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,内蒙古呼和浩特010022 [2]内蒙古农业大学机电工程学院,内蒙古呼和浩特010010 [3]内蒙古自治区林业和草原局,内蒙古呼和浩特010050 [4]鄂尔多斯市鄂托克旗林业和草原局,内蒙古鄂尔多斯016199

出　　处：《草业科学》2024年第10期2275-2283,共9页Pratacultural Science

基　　金：国家自然科学基金项目(52265035、61806103);内蒙古自然科学基金项目(2024LHMS03036)。

摘　　要：鼠洞斑块的快速确定在生态上非常重要,在技术上也具有挑战性。无人机影像与面向对象分析技术(OBIA)相结合为鼠洞斑块识别提供了新的技术手段,也为理解鼠洞斑块与植被盖度之间的空间格局提供了可能。然而OBIA的扩展特征空间提供了大量冗余信息,影响了鼠洞斑块提取的精度和效率。为此提出一种OBIA耦合特征选择的荒漠草原鼠洞斑块识别框架,研究支持向量机、随机森林、K-最近邻在鼠洞斑块识别上的性能,探讨鼠洞斑块面积与植被盖度之间的关系。结果表明:特征选择与随机森林相结合的算法总分类精度高达91.74%,Kappa系数为0.89,优于支持向量机和K-最近邻,表明特征选择在降低特征维度的同时可以提升随机森林算法的性能。基于最优特征集的支持向量机在处理无人机影像上的时间成本最低,样本的平均处理时间为11.48 s,特征选择可以有效提高影像处理的速度。本文还证明了鼠洞斑块面积与植被盖度之间满足二次函数关系。研究结果为基于无人机影像的荒漠草原鼠害监测提供了一种新的方法,也为鼠害防治和草原治理提供了理论指导。Identifying rodent hole patches is ecologically crucial,yet it presents technical challenges.The combination of Unmanned Aerial Vehicle(UAV)imagery and object-oriented image analysis(OBIA)offers a promising approach to detect these patches,and explore the spatial relationships between gerbils and grass cover.However,the extensive feature space data obtained in OBIA often contains redundant information,which can reduce the efficiency and accuracy of rodent hole patch identification.This study proposes a framework integrating feature selection with OBIA to enhance rodent hole patch detection.The performance of three machine learning algorithms,support vector machine,random forest,and K-nearest neighbor,was evaluated.Our results demonstrate that the combination of feature selection and random forest algorithm achieved an overall accuracy of 91.74%,outperforming the support vector machine and K-nearest neighbor algorithms by 9.53%and 20.62%,respectively.These results highlight the effectiveness of feature selection in improving random forest’s performance while reducing the feature dimensionality.Furthermore,the use of support vector machine algorithm along with the optimal feature set exhibited the shortest processing time with an average runtime of 11.48 s per image.Additionally,our study revealed a quadratic relationship between the area of rodent hole patches and grass cover.These findings provide valuable insights for the development of data-driven rodent censuses in desert grassland.

关 键 词：无人机影像 面向对象分析 特征选择 鼠洞斑块 

分 类 号：S812.6[农业科学—草业科学] TP391.41[农业科学—畜牧学]