Super Resolution of Satellite Imagery using Multi-Sensor Fusion

The commercial small satellite provider Planet is expanding towards global coverage at a daily revisit rate at the cost of lower resolution imagery than satellite providers like DigitalGlobe. The lower resolution diminishes the effectiveness of Planet in many surveillance applications that would greatly benefit from its coverage and revisit rate. We therefore propose a single image super resolution algorithm that increases Planet’s resolution by leveraging the large archive of DigitalGlobe imagery. Our unique approach splits the problem into two convolutional neural networks: modality transfer and super resolution. The modality transfer network learns to map Planet imagery to match DigitalGlobe appearance by training on registered pairs of Planet and downsampled DigitalGlobe. The modality transferred image is then super resolved by a single image super resolution network trained unsupervised on DigitalGlobe imagery. We combine these networks into a single network for maximum benefit. Our approach also explores the ability to synthesize multiple super resolved hypotheses of an object and retrieve similar examples from the training set to support each result and lend confidence to the synthesized details. The algorithm will integrate into Kitware’s VIGLANT system to provide a prototype for analysts use.