Abbreviating labelling cost for sentinel-2 image scene classification through active learning.

Abstract

Over the years, due to the enrichment of paired-label datasets, supervised machine learning has become an important part of any problem-solving process. Active Learning gains importance when, given a large amount of freely available data, there’s a lack of expert’s manual labels. This paper proposes an active learning algorithm for selective choice of training samples in remote sensing image scene classification. Here, the classifier ranks the unlabeled pixels based on predefined heuristics and automatically selects those that are considered the most valuable for improvement; the expert then manually labels the selected pixels and the process is repeated. The system builds the optimal set of samples from a small and non-optimal training set, achieving a predefined classification accuracy. The experimental findings demonstrate that by adopting the proposed methodology, 0.02% of total training samples are required for Sentinel-2 Image Scene Classification while still reaching the same level of accuracy reached by complete training data sets. The advantages of the proposed method is highlighted by a comparison with the state-of-the-art active learning method named entropy sampling.