An active learning approach for supervised classification of remotely sensed data: A review
Abstract
With the increasing resolution of remote sensing images, the classification of features is becoming a tedious task. The number of samples to verify pixel classification are supposed to be more for accurate classification of features. There exists a constant need to reduce the sample size by maintaining the accuracy of supervised classification. For fulfilment of this purpose, semi- supervised classification comes into play. The machine learning-based semi-supervised classification is called active learning. This method uses the initially provided data to create a sample space with unlabelled parameters. Then this active learning approach comes up with different approaches to find the pixels having the least relation to the samples provided. These samples are then labelled by the user to enhance the quality of classification. This method of classification reduces the sample space than used in supervised classification but keeps the accuracy intact. This paper reviews several sampling techniques in active learning currently being used in the classification of remote sensing image. These include committee-based strategies, large margin-based strategies and posterior probability-based strategies. Every sampling method uses different techniques to minimize the sample size and increase the classification accuracy. These active learning approaches lie in the dependent domain. Some strategies may be generalised but some may not. In any case, they will provide a cost-effective way of labelling the samples and leading to good classification.
