Evaluation of Split-Brain Autoencoders for High-Resolution Remote Sensing Scene Classification

Self-supervised methods are interesting for remote sensing because there are not many human labeled datasets available, but there is practically unlimited amount of data that can be used for self-supervised learning. In this paper we analyze the use of split-brain autoencoders in the context of remote sensing image classification. We investigate the importance of training set size, choice of color space and size of the model to the classification accuracy. We show that even with small amount of unlabeled training images, if we finetune the weights learned by the autoencoder, we can achieve almost state of the art results of 89.27% on AID dataset.