关键词: 植被类型识别;机器学习;Sentinel-2A;特征选择;递归特征消除法

Abstract: Using Sentinel-2A image data and real vegetation sample data from Leizhou Peninsula as the research area,this paper comprehensively discusses the classification effects of random forest and support vector machine in machine learning,and compares them with the traditional maximum likelihood method.Firstly,9 bands,7 vegetation indexes and 72 texture features of Sentinel-2A image are extracted successfully,then the feature combination of 10 features is selected by recursive feature elimination method and applies to three classification methods,and the classification effect is compared.The results show that:①Effectively using a variety of characteristic variables is the key to improve the vegetation type recognition accuracy,in terms of the importance of the different characteristics of vegetation type recognition,the spectral features are the same to the texture features and greater than vegetation index,three importance are similar.②Random forest classification has the best effect,which can not only select features effectively,but also ensure the precision of vegetation type extraction and improve the operation efficiency.③The feature combination based on the recursive feature elimination method of random forest feature selection can not optimize the performance of other classifiers,and the optimization effect of the random forest model itself is limited.

Key words: vegetation type identification;machine learning;Sentinel-2A;feature selection;recursive feature elimination