Machine learning algorithms demonstrated promising results for hyperspectral semantic segmentation. However, they strongly rely on the quality of training datasets. As far as the annotation of hyperspectral images is often expensive and time-consuming, only a few thousand pixels can be labeled. In this context, active learning algorithms select the most informative pixels to be labeled. In the machine learning community, recent active learning methods have overcome the performance of conventional algorithms but do not always scale to large remote sensing images. Therefore, we introduce in this paper a preprocessing method that allows the use of computationally intensive active learning algorithms without significant impacts on their effectiveness.