关键词: 高分二号, 特征选择, XGBoost, 树种分类, 遥感

Abstract:

Tree species distribution is an important indicator of forest resources monitoring,and it is also one of the difficulties in the application of remote sensing images in forest resources monitoring.Based on domestic GF-2 satellite imagery data,forest resources survey data and DEM data,combined with spectral,texture,index and topographic factors and other characteristics.Three classification algorithms,support vector machine,random forest and XGBoost,are compared,and the optimal algorithm (XGBoost) is selected according to classification accuracy.Four dominant tree species,broad-leaved trees,Pinus massoniana,Chinese fir and Phyllostachys pubescens,were classified by row feature selection.The results show that the total classification accuracy of XGBoost classification model is 83.58% and the Kappa coefficient is 0.77,which is significantly higher than that of support vector machine and random forest classification method.After feature selection,the classification accuracy of tree species has not been improved obviously,but the redundancy of features can be reduced,which provides a reference for feature selection and dimensionality reduction in small sample data.

Key words: GF-2, feature selection, XGBoost, tree species identification, remote sensing