SEMI-SUPERVISED LEARNING FOR CLASSFICATION OF CROP AREAS IN NORTH KOREA

摘要

One of important issues in supervised classification of remote sensing imagery is to collect sufficient numbers of reliable training data. However, to acquire reliable training data requires much cost and time, and may be even impossible in case of classification of inaccessible areas. In this study, iterative classification based on semi-supervised learning using a small number of labelled data is presented to extract reliable training data for classification of inaccessible areas. An initial classification result is first generated by applying a random forest classifier with the small number of training data. The majority vote of random forest classifier is used to extract training candidates with higher certainty measures. The candidates extracted are then used as training data for next supervised classification. To evaluate the applicability of the iterative classification based on semi-supervised learning, a case study of classification of crop areas is carried out using multi-temporal Landsat-8 acquired in Daehongdan, North Korea. From a case study, the accuracy of classification increased as the amount of training data increased through the iterative classification. The semi-supervised approach presented in this study showed a significant improvement of classification accuracy, compared with the classification with a small number of training data. Therefore, it is expected that the iterative classification based on semi-supervised learning could be effectively applied to classify areas in which it is difficult to collect sufficient training data.