Data Augmentation Approaches for Satellite Image Super-Resolution

Data augmentation is a well known technique that is frequently used in machine learning tasks to increase the number of training instances and hence decrease model over-fitting. In this paper we propose a data augmentation technique that can further boost the performance of satellite image super resolution tasks. A super-resolution convolutional neural network (SRCNN) was adopted as a state-of-the-art deep learning model to test the proposed data augmentation technique. Different augmentation techniques were studied to investigate their relative importance and accuracy gains. We categorized the augmentation methods into instance based and channel based augmentation methods. The former refers to the standard approach of creating new data instances through applying image transformations to the original images such as adding artificial noise, rotations and translations to training samples, while in the latter we fuse auxiliary channels (or custom bands) with each training instance, which helps the model learn useful representations. Fusing auxiliary derived channels to a satellite image RGB combination can be seen as a spectral-spatial fusion process as we explain later. Several experiments were carried out to evaluate the efficacy of the proposed fusion-based augmentation method compared with traditional data augmentation techniques such as rotation, flip and noisy training inputs. The reconstruction quality of the high resolution output was quantitatively evaluated using Peak-Signal-To-Noise-Ratio (PSNR) and qualitatively through visualisation of test samples before and after super-resolving.