Local Fisher discriminant analysis (LFDA) is a feature extraction technique that proved efficient to reduce several types of data and succeeded to outperform many state-of-the-art methods. However, due to its supervised nature, LFDA’s efficiency depends on the available labeled samples and declines dramatically when the latter are very few. Hence, we assume that we cannot resort to LFDA to reduce unlabeled data. In this article, we studied to what extent this assumption is true and questioned the possibility of using LFDA to reduce hyperspectral images (HSIs) with no available ground truth. To study the real impact of the labeled information on LFDA’s performance, we replaced the costly expert-made ground truth by different sets of labeled samples that are generated based on the image’s offered spectral and/or spatial information, with no prior knowledge of the captured scene nor of its classes. Our proposed sets proved able to guide LFDA in extracting relevant discriminating features. This proved that LFDA does not depend only on expert-made labeled information and led us to define the unsupervised image-adapted LFDA (uiaLFDA) that can properly reduce an HSI without requiring its ground truth. To do so and to replace the ground truth that LFDA usually requires to reduce an HSI, uiaLFDA defines its own set of labeled samples by simply gridding the image into cells where each cell is considered a class. Our experiments ran on three HSIs proved that uiaLFDA is as efficient as LFDA and, even better, in reducing unlabeled HSIs.