Due to the limitation of labeled training samples, computational complexity, and other difficulties, active learning (AL) algorithms aiming at finding the most informative training samples have been an active topic of research in remote sensing image classification in the last few years. Usually, AL follows an iterative scheme, and the search of new samples relies on the whole image, resulting in that an approach may turn out to be prohibitive when the data sets are huge, e.g., hyperspectral data. Large amounts of unlabeled samples are easy to collect indeed, with respect to the cost of labeled sample collection. However, algorithm complexity, data storage capacity and processing times are also limited. Therefore, a sample set smaller in size, and consisting of the most valuable information, is preferable. In this work, we propose a design protocol to generate a more significant candidate sample set for active learning, aiming at reducing the unlabeled sample search complexity, and eventually improving the classification performance. The basic idea is providing the initial labeled and unlabeled samples that are composed of mixed or pure samples for AL heuristics, to find out which one is better for AL from the low-cost sample design point of view. For comparison and validation purposes, six state-of-the-art AL methods (including breaking ties, margin sampling, margin sampling by closest support vectors, normalized entropy query-by-committee, multi-class level uncertainty and multi view adaptive maximum disagreement based active learning) were tested on real hyperspectral images with different resolution both with and without the proposed sample design protocol. Experimental results confirmed the advantages of the proposed technique.