Realization And Application Of Random Forest Algorithm In Land Cover Classification Of High Resolution Remote Sensing Imagery

Study on change of land cover has become an important topic of global change research and modern geo-scientific research.To the research of land cover change,land cover classification is the basis and important part.Robust classification method is required to land cover classification,random forest is a powerful machine learning classifier,which is an ensemble learning algorithm based on nonparametric regression compared to the traditional decision tree algorithm.In the recent years,number of high resolution satellite has been launched,high-resolution images with amount of information,however,the accuracy of classification did not increase with the improvement of spatial resolution.The accuracy obtained by the traditional classification methods based on spectrum applied in high resolution imagery classification is far behind requirements of production.Increasing of features' information reflect in texture information become more abundant in the imagery.Texture structure getting much clear with the increasing of spectral resolution.Extraction of texture is a key step in image classification processing.To extract the stability texture feature with strong ability of identify are meaningful to improve the accuracy of imagery classification.In this paper two townships of Shitai County are taken as study area,using Rapideye imagery with high resolution.The landscape of study area with high spatial heterogeneity and large shadow area.The field survey found out that some categories in the imagery display are hard to be separated by color.Therefore,geo-statistic method which has been widely used in remote sensing image texture extraction is used in the study to extract texture and using Band Math to extract vegetation index.Using random forest to selecting features to compute feature importance.To see the variation trend of accuracy of classification how to change by change the number of trees,one of parameters of random forest and features combination.Compared to maximum likelihood method,to see which one can get higher accuracy.The results show that:(1)The generalized error of random forest tent to stable when the number of random forest trees(N)over a fixed value.In other words it will not increase with the increase of N while the efficiency of computation will decrease.It can meet the accuracy requirements when N=100 according to other reference.But in this study,N=500 can not only meet the classification accuracy will also ensure operational efficiency;(2)Applied random forest and maximum likelihood classification to evaluate adding texture features yield Kappa coefficients are 0.7134 and 0.6315,higher than adding vegetation indices and the combination of texture and vegetation indices.Applied random forest can yield higher accuracy than applying maximum likelihood classification.In addition,adding texture textures can improve the of accuracy classification in a certain extent: it can help improving accuracy of cultivated land and building which have obviously geometry and strongly regularity significantly.It also can provide an effective basis for the distinction different categories with similar spectral values.The comprehensive performance of the random forest is better,it not only can ensure the precision of the classification,but also to ensure the efficiency of operations.It is more suitable for actual production application and easy to operate.It can filter the desired characteristics and give the number of features required by computer and meet the requirements of the classification accuracy by prediction of N through writing R language.The maximum likelihood procedures are not complicated,but the accuracy is relatively low.