Iterative active learning with diffusion geometry for hyperspectral images

We propose an active learning algorithm for labeling hyperspectral images (HSI). Pixels with ambiguous class affinity are iteratively estimated using geometric and statistical properties of the data. These pixels are then labeled with ground truth data, yielding a small but potent set of labeled pixels, which are consequently used to label the remaining data. The proposed method enjoys quasilinear complexity in the number of sample pixels, as well as competitive empirical performance on real HSI. Substantial improvement in labeling accuracy compared to unsupervised learning and existing active learning methods is observed with just a few well-selected label queries.