Study of a classification model for unlabeled hyperspectral images

With the aim of developing a pixel-by-pixel hyperspectral image classification model trained on unlabeled data, we developed a method to demonstrate the feasibility of such training. To this end, we took a labeled dataset from which we removed the labels, to train our model consisting of a modified transformer for image processing and a K-means model. We then compared the predictions of this model with the original labels of the dataset. The similarities we observed allow us to be enthusiastic about the idea of being able to classify hyperspectral images pixel by pixel, without prior labeling, to derive a meaningful representation.

View the full documentation (PDF in French)