基于XGBoost的机载激光雷达与高光谱影像结合的特征选择算法

Feature Selection Algorithms of Airborne LiDAR Combined with Hyperspectral Images Based on XGBoost

为了解决地物分类的机载激光雷达(LiDAR)与高光谱特征构造中存在的特征维数过高的问题,提出了一种基于XGBoost与皮尔逊相关系数相结合的特征选择算法——XGB-PCCS,同时设计了XGBoost与序列后向选择相结合的特征选择算法——XGB-SBS与之对比。采用真实数据验证所设计的两种算法,结果表明:两种算法均可在保证分类结果准确率的基础上有效地减小特征集维数;XGB-SBS算法保留的特征维度为33,得到的总体分类精度为95.63%,Kappa系数为0.943;XGB-PCCS算法保留的特征维度为25,总体分类精度为95.55%,Kappa系数为0.942。XGB-PCCS算法的人为干预程度较低,运行时间较短,保留的特征集更精简。此外,对比了两种算法得到的特征子集,并总结了LiDAR点云与高光谱影像多模态特征构造中重要程度较高的24种特征。

In order to solve the problem of high feature dimension in the feature construction of airborne light detection and ranging(LiDAR) and hyperspectral images for the classification of ground objects,we propose a feature selection algorithm based on extreme gradient boosting(XGBoost) combined with Pearson correlation coefficients(PCCS),named XGB-PCCS.Meanwhile,another feature selection algorithm based on XGBoost combined with sequential backward selection(SBS),named XGB-SBS,is designed to compare with XGB-PCCS.The real data is used to verify the two algorithms designed above.The results show that both algorithms can effectively reduce the dimension of feature sets on the basis that the accuracy of classification results is ensured.As for the XGB-SBS algorithm,the retained feature dimension is 33,the overall classification accuracy is 95.63%,and the Kappa coefficient is 0.943.In contrast,as for the XGB-PCCS algorithm,the retained feature dimension is 25,the overall classification accuracy is 95.55%,and the Kappa coefficient is 0.942.The XGB-PCCS algorithm has low degree of human intervention and short running time,and the retained feature set is compact.In addition,the feature subsets obtained by the two algorithms are compared,and 24 kinds of features with high importance in the multi-modal feature construction of LiDAR point cloud and hyperspectral images are summarized.