Sentinel-2A multi-temporal misregistration characterization and an orbit-based sub-pixel registration methodology

The European Space Agency (ESA) Sentinel-2A Multi Spectral Instrument (MSI) provides global coverage 10 m to 60 m optical wavelength data in fixed geolocated tiles. Recently, it became apparent that when multi-temporal Sentinel-2A tile acquisitions are compared, they can be misregistered by more than one 10 m pixel. This study quantifies the multi-temporal misregistration by application of a recently published matching algorithm to all the available Sentinel-2A data sensed in six months (June 2nd to November 30th 2016) over 1400 × 3000 km of southern Africa. The Sentinel-2A L1C tile product processing software was updated by ESA on June 15th 2016. The results of our analysis show that the update reduced the mean multi-temporal Sentinel-2A tile misregistration from about 1.6 10 m pixels (standard deviation >0.9 pixels) to 0.4 10 m pixels (standard deviation of about 0.2 pixels). Future updates to the ESA L1C tile product processing software are expected to further reduce this misregistration error. Meanwhile, users should exercise care when using multi-temporal Sentinel-2A data for applications that require precise sub-pixel registration. In solution, an orbit-based registration methodology that uses a least-squares adjustment applied to matches derived among Sentinel-2A tile acquisitions sensed in different orbits is described and demonstrated. For all the tile acquisitions in an orbit in a Universal Transverse Mercator (UTM) map projection zone, they are registered in the same manner, meaning that there are no relative shifts among the tile acquisitions from the same orbit in the registered data, and that the tile acquisitions that are obscured by cloud or that have a large amount of missing data can still be registered. Over 97% of the study data, including the data sensed before June 15th 2016, were registered. The mean tile registration precision was better than 0.15 10 m pixels (standard deviation 0.06 pixels) and is shown qualitatively to provide multi-temporal data with no visually obvious misregistration at 10 m resolution.