Visualization-Based Active Learning for the Annotation of SAR Images

Active learning has gained a high amount of attention due to its ability to label a vast amount of unlabeled collected earth observation (EO) data. In this paper, we propose a novel active learning algorithm which is mainly based on employing a low-rank classifier as the training model and introducing a visualization support data point selection, namely, first certain wrong labeled (FCWL). The training model is composed of the logistic regression loss function and the trace-norm of learning parameters as regularizer. FCWL selects those data points whose labels are predicted wrong but the classifier is highly certain about them. Our experimental results performed on different extracted features from a dataset of SAR images confirm at least 10% improvement over the state-of-the-art methods.