A software tool for the estimation of uncertainties and spectral error correlation in Sentinel-2 Level-2A data products

The Copernicus Sentinel-2 (S2) satellite mission acquires high spatial resolution optical imagery over land and coastal areas. Delivering uncertainty estimates and spectral error correlation alongside S2 data products facilitates the constrain of retrieval algorithms, propagate further downstream the retrieval uncertainty, and finally make informed decisions to end-users. This study presents a software (available at Gorroño (2023)) that generates uncertainty estimates and spectral error correlation associated to the S2 L2A data products (i.e. surface reflectance). The uncertainty considers both the Level-1 (L1) uncertainty estimates for top-of-atmosphere (TOA) reflectancefactor and the atmospheric correction. The propagation is performed with a Multivariate MonteCarlo model that effectively accounts for the spectral error correlation between S2 L2A bands. The uncertainty analysis involves the propagation of the L1C TOA reflectance factor through the atmospheric correction using LibRadtran. This propagation accounts for input uncertainty such as L1 TOA reflectance, aerosol optical thickness (AOT) or adjacency correction. On the top of this propagation we also model the contributions from the Lambertain assumption of the correction model and the estimated accuracy of the LibRadTran software. We show results for surface reflectance uncertainty at two different sites. The examples over the Amazon forest and Libya4 desert site illustrate the large variations of the uncertainty levels and spectral error correlation depending on the scene. Furthermore, we include an example on the propagation of surface reflectance uncertainty to spectral vegetation indices. The propagation over vegetation metrics indicates a strong dependence of the error covariance matrix with the phenological cycle and exemplifies how critical is that S2 L2A products include both uncertainty and spectral error correlation since they are effectively the input to many different land surface parameters. Its implementation as an operational per-pixel processing and dissemination of both the uncertainty and spectral error correlation becomes challenging. However, exploring cloud computing and machine learning techniques could become an adequate pathway to minimise these challenges.