Unsupervised Learning of Low Dimensional Satellite Image Representations via Variational Autoencoders

The growing number of images acquired by new satellite missions increases the interest of learning low dimensional image representations without human supervision. Variational AutoEncoders (VAE) are one of the most promising strategies marrying graphical models and deep learning. They are able to learn continuous, structured and probabilistic latent spaces encoding the data. In this work, a VAE architecture is proposed and analyzed for multispectral Sentinel-2 images. The regularized β-VAE is studied and compared with the classical auto-encoder strategy. A classification experiment is carried out to corroborate that generated latent spaces can preserve the salient features of the input data. Classification results show that high accuracies can be obtained by using low dimensional latent representation learned by VAE as input data in a standard classification approach.