Research On Features Selection Method In Time Series Of Remote Sensing Data In Vegetation Classification

When different vegetations are affected by the change of the climate cycle,different rhythm changes will appear.And these rhythms can be found in the observation results of time series of remote sensing data.The vegetation classification features extracted from time series data can reflect the phenological between different vegetations accurately.A variety of features extracted from different time series data form the feature set that has characteristics of high dimension and heterogeneity,and the spatial distribution of this kind of feature sets are not consistent with normal distribution,so traditional dimensionality reduction algorithm based on the assumption of normal distribution can not be used to make a feature selection from this kind of feature sets.In order to reduce the dimension of the feature set that has the characteristics of high dimension,heterogeneity and non normal distribution,and to solve the limitation of different kinds of data in the application of vegetation classification,this passage takes Zhangye City of Gansu Province and Tengchong City of Yunnan Province as the research area,extracts the vegetation classification features from several time series of remote sensing data by two ways whose names are classification feature extraction based on time series curve and classification feature extraction based on data dimension reduction method.And then we caculate the Gini coefficient of the classification features that are extracted based on time series curve and selecte the feature according to Gini coefficients.At the same time,we selected the bands extracted by the data dimension reduction method according to information content.Finally,the feature subsets selected by two methods above are used to make a classification based on Random Forest Classifier and we analyze the influence of different feature selection methods on classification accuracy.We can make conclusions as following:(1)This paper made a selection for classification features based on the information entropy principle and finished reducing the dimension of multi-source heterogeneous and non normal distribution feature sets,and thus the limitations of different types of data in the application of vegetation classification can be solved.(2)We can extract the phenological features from different time series curves of vegetations and form their own phenological feature sets.These feature sets can reflect the differences of vegetation phenology between different vegetations accurately.(3)We can get the best classification feature sets based on Gini coefficients and the feature sets have the characteristics of low dimension and low redundancy.We can raise the classification accuracy by this kind of feature sets.The result shows that ZhangYe district obtain the accuracy of 77.31% and TengChong district obtain 75.85% by using features selected by Gini coefficients.And these two accuracies are higher than the classification result by using all features.