Construction of cloud-shadow-water mask based on Random Forests algorithm

Clouds and their shadows in the remote sensing images are the key factors that influence the application of the data in many fields .Several methods , such as constructing cloud mask , replacement of the pixels , linear mixture spectral analysis , and principal component analysis , have been proposed in the past decades to solve this problem.In this research, based on the analysis of spectral curve , the authors utilized Decision Tree (DT)classifier and Random Forest ( RF) classifier to obtain the cloud -shadow-water mask.There was little difference between the spectral curve of shadow and water due to the mixture of shadow and other surface materials such as vegetation and impervious surface .In this case , the DT classifier could not effectively distinguish shadow and water because the decision rule and threshold were determined by analyzing the spectral curves of different samples .RF classifier was based on the ensemble of the results derived from multiple decision tree classifiers , which was more robust than one decision tree classifier .In this study , when there were only a few training samples , results that were more accurate were derived from RF classifier compared with the results from DT classifier .When the size of training samples lay in the range of 250 and 400 , no significant difference was found between the results derived from these two algorithms.This indicates that RF classifier could be used to deduce the cloud -cloud shadow -water mask successfully .