Study of high-dimensional fuzzy classification based on random forest algorithm

Abstract

The spatial resolution of hyperspectral data is generally very low,the mixed pixels are extensively distributed,and hence fuzzy classification is commonly used in the mixed pixel analysis. As the accuracy of fuzzy classification is often limited by the feature dimensions and fuzzy samples selection,the random forest( RF)algorithm is put forward in this paper to select features and obtain fuzzy samples; in the low- dimensional feature space,fuzzy samples are used to make fuzzy classification. Fuzzy classification and RF are merged by using two-step classification,following the principle of unanimity assumption. Using different samples,different experimental areas and different partition optimization situations,the authors conducted three comparative experiments,and the results show that the method proposed in this paper solves the limitation of fuzzy classification and improves its accuracy. It is also proved that the classification accuracy of the method is robust for the original sample.