Recursive Ensemble Land Cover Classification with Little Training Data and Many Classes

摘要

Land-cover classification can construct a land-use map to analyze satellite images using machine learning. However, supervised machine learning requires a lot of training data since remote sensing data is of higher resolution that reveals many features. Therefore, this study proposed a method to generate self-training data from a small amount of training data. This method generates self-training, which is regarded as the correct class to consider various times and the surrounding land cover. As a result of self-training conducted using this method, the Kappa coefficient was 0.644 for 12 classification problems with one training data per class.