无人机点云与图像跨模态混合融合的乔木林单木尺度树种分类研究
Study on Classification of Arbor Tree Species at Single Tree Scale based on Cross-modal Hybrid Fusion of UAV Point Cloud and Image
鄢敏 1夏永华 2王冲 3孔夏丽 1太浩宇 1李晨1
作者信息
- 1. 昆明理工大学 国土资源工程学院,云南 昆明 650093
- 2. 昆明理工大学 国土资源工程学院,云南 昆明 650093;昆明理工大学城市学院,云南 昆明 650051
- 3. 中国电建集团 昆明勘测设计研究院有限公司,云南 昆明 650200
- 折叠
摘要
为探索机载点云与无人机可见光影像在乔木树种识别与分类领域的应用潜力,提出了一种多模态特征与决策混合融合的无人机单木尺度树种分类识别方法.首先使用Kendall Rank相关系数法与排列重要性分析(Permutation Importance,PI)进行特征选择,采用高效低秩多模态融合算法(Low-rank Multimodal Fusion,LMF)融合点云与影像特征.再引入集成学习,将点云、影像及融合特征分别输入Stacking集成的极限梯度提升机(eXtreme Gradient Boosting,XGBoost)、轻型梯度提升机(Light Gradient Boosting Machine,LightGBM)与随机森林(Random Forest,RF)3 个基分类器,最后采用元分类器—朴素贝叶斯进行决策融合.实验数据表明:所提方法独立测试精度达99.4％,较传统的特征串联融合随机森林分类器提升了 22.58％,Kappa系数提升了 0.285 4.与卷积神经网络(Convolutional Neural Network,CNN)对比实验证明:所提算法在小样本训练的优势明显,且具有更好的泛化能力.
Abstract
To explore the application potential of airborne point cloud and UAV visible light image in tree spe-cies identification and classification,a single-tree scale tree species classification and recognition method based on UAV hybrid fusion of multi-modal features and decision was proposed.Firstly,Kendall Rank correlation co-efficient method and Permutation Importance(PI)were used for feature selection,and Efficient Low-Rank Multi-Mode Fusion Algorithm(LMF)was used to fuse the selected point cloud and visible image features.En-semble learning was introduced to input point cloud,image,and fusion features into eXtreme Gradient Boosting(XGBoost),Light Gradient Boosting Machine(LightGBM),and Random Forest(RF)base classifiers inte-grated by Stacking.Finally,the meta classifier,Naive Bayes,is used for decision fusion.The experimental da-ta show that the independent test accuracy of the proposed algorithm is 99.4％,which improves 22.58％com-pared with the Random Forest classifier by traditional feature concatenate fusion.In addition,the Kappa coeffi-cient also increased by 28.54％.The comparison experiment with Convolutional Neural Network(CNN)shows that the proposed algorithm has obvious advantages in small sample training and better generalization ability.
关键词
多模态融合/单木尺度/树种分类/集成学习Key words
Multimodal fusion/Single tree scale/Tree species classification/Ensemble learning引用本文复制引用
基金项目
国家自然科学基金(41861054)
国家自然科学基金(42161067)
技术服务项目(KKF0201956004)
出版年
2024
NETL
NSTL
维普
万方数据