An active learning method with entropy weighting subspace clustering for remote sensing image retrieval

Abstract

The increasing volume of high-resolution satellite imagery made image retrieval a demanding research field in the remote sensing (RS) community. The RS data are often complex, with varied temporal and spatio-spectral properties, and are unlabeled, and the process of labeling is relatively expensive. Therefore, developing sophisticated models for retrieving relevant images from RS databases is necessary, especially when few labeled samples are available. Motivated by this fact, we propose a new cluster-guided active learning (CG-AL) framework for remote sensing image retrieval (RSIR). The CG-AL-RSIR utilizes an entropy weighting K-Means subspace (EWKMS) clustering algorithm for influencing an active learning (AL) framework using a support vector machine (SVM). The AL framework efficiently selects the most informative samples for labeling depending on a weighted entropy uncertainty sampling (WtEUS). A bag-of-visual-words representation of the scale-invariant feature transform (SIFT) image descriptors is extracted for the retrieval. The EWKMS clustering addresses the high dimensionality of the RSI dataset efficiently in obtaining quality clusters for labeling The CG-AL-RSIR resulted in an overall accuracy of up to 94.14% and 92.48% for the UCM and SIRI-WHU datasets, respectively, and an ANMRR value of 0.290 for UCM and 0.326 for SIRI-WHU datasets, utilizing less number of labeled samples.