A key factor for the success of supervised remote sensing image classification is the definition of an efficient training set. Suboptimality in the selection of the training samples can bring to low classification performance. Active learning algorithms aim at building the training set in a smart and efficient way, by finding the most relevant samples for model improvement and thus iteratively improving the classification performance. In uncertaintybased approaches, a user-defined heuristic ranks the unlabeled samples according to the classifier's uncertainty about their class membership. Finally, the user is asked to define the labels of the pixels scoring maximum uncertainty. In the present work, an unbiased uncertainty scoring function encouraging sampling diversity is investigated. A modified version of the Entropy Query by Bagging (EQB) approach is presented and tested on very high resolution imagery using both SVM and LDA classifiers. Advantages of favoring diversity in the heuristics are discussed. By the diverse sampling it enhances, the unbiased approach proposed leads to higher convergence rates in the first iterations for both the models considered.