A new automated method for improving georeferencing of nighttime ECOSTRESS thermal imagery

Georeferencing accuracy plays a crucial role in providing high-quality ready-to-use remote sensing data. Georeferencing of satellite imagery is typically based on position and pointing direction of a sensor, which are provided by star trackers and GPS. As the Ecosystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) is not equipped with star trackers, georeferencing of its imagery is based on the inaccurate knowledge about the location of its platform (the International Space Station) and later adjusted by image matching to the Landsat Orthobase. Although the georeferencing accuracy for daytime imagery is relatively high, we have observed that the nighttime imagery in Olkaria (Kenya), exhibits errors of 13.7 pixels on average, but in extreme cases even 62 pixels. Image based georeferencing in nighttime thermal satellite imagery is challenging, dueto complexity of thermal radiation patterns in diurnal cycle and coarse resolution of thermal sensors in comparison to sensors imaging in the visual spectral range. Our paper introduces a novel approach for improved georeferencing of nighttime thermal imagery. We use object based matching of water bodies to an up-to-date landcover reference with high geolocation accuracy. Dynamically changing land cover often renders (static) land cover data bases unusable as a reliable reference basemap. We overcome this issue by automatically creating an up-to-date landcover reference to match acquisition time of the target image. Additionally, we use object based matching to account for lower spatial resolution of thermal sensors, as well as potential sharpness issues. In our method, edges of water bodies serve as matching objects, as they exhibit a relatively high contrast to adjacent areas. Results show that our method improves the existing georeferencing of ECOSTRESS images by 10.6 pixels on average, and an average accuracy of ±3.1 pixels is achieved. The accuracy of our method depends on accurate cloud masks, because cloud edges can be mismatched as water-body edges and included in fitting of transformation parameters. We tested our method on ECOSTRESS imagery, but it is possible to be used with data from other sensors as well.