基于多源遥感数据的输电线走廊树种分类

Tree Species Classification of Power Line Corridor Based on Multi-source Remote Sensing Data

针对目前树冠提取中受背景影响和易出现过度分割的问题,首先,采用可见光差异植被指数和双边滤波对传统的单木树冠分割方法进行了改进;然后,以单木树冠为对象提取多维特征,并利用XGBoost算法进行特征重要性排序和特征选择;最后,使用随机森林、支持向量机、人工神经网络3种非参数分类器,设计了12种分类方案,进行了单木树种分类和精度评价。结果表明,改进的单木分割方法可以有效提高树冠提取精度,得到的树冠分割精度在80%以上;将Li DAR数据和航空正射影像相结合,采用XGBoost算法进行特征选择后,使用ANN分类器的分类方案精度最高,总体精度为86.19%,说明多源数据协同作用和特征选择可以提高树种分类精度,在单木尺度上ANN分类器对现有树种类型的分类能力最强。

The effectiveness of airborne LiDAR point cloud and aerial imagery on tree species classification and the effect of XGBoost algorithm for feature selection on tree species classification accuracy were researched,and the ability three non-parametric classifiers of random forest,support vector machine and artificial neural network to classify tree species on a single-wood scale were evaluated.Aiming at the current background effect of canopy extraction and the problem of over-segmentation,the traditional single tree canopy segmentation method was improved by using the visible light difference vegetation index and bilateral filtering;and then the single tree canopy was used as an object to extract multi-dimensional features by using the XGBoost algorithm to perform feature importance ranking and feature selection.Finally,three non-parameter classifiers of random forest,support vector machine and artificial neural network were used to design 12 classification schemes to classify single tree species and do accuracy evaluation.The results showed that the improved single tree segmentation method can effectively improve the accuracy of tree crown extraction,and the accuracy of the obtained tree canopy segmentation results was more than 80%;the LiDAR data and aerial orthophotos were combined,and the ANN classifier was used for feature selection after XGBoost algorithm for feature selection.The scheme had the highest accuracy,with an overall accuracy of 86.19%,indicating that multi-source data synergy and feature selection can improve the accuracy of tree species classification.The ANN classifier had the strongest ability to classify existing tree species on a single tree scale.