This paper presents a novel multi-label active learning (MLAL) technique in the framework of multi-label remote sensing (RS) image scene classification problems. The proposed MLAL technique is developed in the framework of the multi-label SVM classifier (ML-SVM). Unlike the standard AL methods, the proposed MLAL technique redefines active learning by evaluating the informativeness of each image based on its multiple land-cover classes. Accordingly, the proposed MLAL technique is based on the joint evaluation of two criteria for the selection of the most informative images: i) multi-label uncertainty and ii) multi-label diversity. The multi-label uncertainty criterion is associated to the confidence of the multi-label classification algorithm in correctly assigning multi-labels to each image, whereas multi-label diversity criterion aims at selecting a set of un-annotated images that are as more diverse as possible to reduce the redundancy among them. In order to evaluate the multi-label uncertainty of each image, we propose a novel multi-label margin sampling strategy that: 1) considers the functional distances of each image to all ML-SVM hyperplanes; and then 2) estimates the occurrence on how many times each image falls inside the margins of ML-SVMs. If the occurrence is small, the classifiers are confident to correctly classify the considered image, and vice versa. In order to evaluate the multi-label diversity of each image, we propose a novel clustering-based strategy that clusters all the images inside the margins of the ML-SVMs and avoids selecting the uncertain images from the same clusters. The joint use of the two criteria allows one to enrich the training set of images with multi-labels. Experimental results obtained on a benchmark archive with 2100 images with their multi-labels show the effectiveness of the proposed MLAL method compared to the standard AL methods that neglect the evaluation of the uncertainty and diversity on multi-labels.
|