A least-squares spatio-temporal adjustment algorithm (LSSTAA) for radiometric normalization of PlanetScope time series

The commercial PlanetScope constellation provides near daily multispectral 3–4 m spatial resolution imagery for the majority of the global land surface. However, PlanetScope images may have inconsistent surface reflectance among images due to factors including lack of onboard calibration, unreliable atmospheric correction, and detector differences among individual PlanetScope sensors and sensor generations. These inconsistencies should preferably be minimized for time series and large area applications that use multiple images. This paper presents a novel least-squares spatio-temporal adjustment algorithm (LSSTAA) for relative radiometric normalization of PlanetScope image time series. A network, with one or more images selected as references, is constructed by connecting target-reference and target-target images that spatially overlap, and then least-squares adjustment is employed to estimate optimal spectral normalization coefficients (gains and offsets) for all target images simultaneously. Compared with conventional one-to-one normalization approaches that adjust a target image to a reference image, the LSSTAA considers all connected target-target images and target-reference images, and so ensures a high degree of surface reflectance consistency among the normalized images. The LSSTAA is demonstrated considering five sites that have diverse land cover types, using an average of 15 PlanetScope Dove-Classic and Dove-R surface reflectance images per site acquired within a month that have variable spatial overlap and were acquired in periods with both gradual and abrupt land surface change conditions. The LSSTAA is shown to significantly reduce seamlines in site PlanetScope image mosaics, and improve pairwise scatterplots of image time series with more mutually consistent surface reflectance compared with the original unnormalized time series. Quantitative assessment using the relative Time-series Smoothness Index (TSI) shows an overall reduction of TSI values of 59.6 %, 53.4 %, 42.9 %, 50.5 %, and 34.6 % for the green, blue, red, NIR surface reflectance and derived Normalized Different Vegetation Index (NDVI) of LSSTAA-normalized PlanetScope time series compared with the unnormalized data, respectively. The LSSTAA results for 11 images are also compared with the equivalent harmonized surface reflectance images provided by Planet that are generated by normalizing images individually to Sentinel-2 monthly composites. The LSSTAA results are again shown to have more consistent surface reflectance (considering the same qualitative and quantitative measurements) than the Planet harmonized surface reflectance imagery. The limitations and advantages of the LSSTAA and future research directions are also discussed.