Earth-INN: Invertible Neural Networks (INNs) for Accommodating Earth Observation

Earth observation satellites offer a rich and diverse data source with significant variations in spectral, temporal and spatial domains. A key objective in the field is to develop source-agnostic and robust methodologies to facilitate flexible and efficient utilization of the data from various sources. Earth-INN project aims to train invertible neural networks to translate imagery from one earth-observation  satellite (e.g. Sentinel-1) into imagery of another satellite (e.g. Sentinel-2) given same location and time. In the process, we hope to learn the global distribution of satellite data over from different sources. This would enable us to fill coverage gaps in a satellite source using another satellite source, allowing for source agnostic earth observation.