|Paper title||Establishing an end-to-end uncertainty budget for surface reflectance pre-processing in CCI ECVs|
|Form of presentation||Poster|
Establishing an end-to-end uncertainty budget is essentially required for all ECVs of ESA’s Climate Change Initiative (CCI). The reference guide for expressing and propagating uncertainty consists of the GUM and its supplements, which describe multivariate analytic and Monte Carlo methods. The FIDUCEO project demonstrated the application of these methods to the creation of ECV datasets, from Level 0 telemetry to Level 1 radiometry and beyond. But despite this pioneering work, uncertainty propagation for ECVs is challenging. Firstly, many retrieval algorithms do not incorporate the use of quantified uncertainty per datum. Using analytic methods for propagating uncertainty requires completely new algorithmic developments while applying Monte Carlo methods is usually straightforward but leads to proliferation of computational and data curation resources. Secondly, operational radiometry data are usually not associated with a quantified uncertainty per datum and error correlation structures between data are not quantified either. Deriving this information from original sensor telemetry and an according harmonisation with respect to an SI traceable satellite (SITSAT) reference is a future task.
Nevertheless, it is feasible to explore and prepare ECV processing for the use of uncertainty per (Level 1) datum and error correlation structures among data already now, based on instrument specifications and simplifying assumptions. For the Land Cover ECV of the CCI we developed a Monte Carlo surface reflectance pre-processing sequence, which considers three most significant effects: errors in satellite radiometry, errors in aerosol retrieval, and errors in cloud detection. Error correlations between radiometric data are considered using a simplified correlation matrix with a constant correlation coefficient. Such simplified correlation structure can take account of uncorrelated random noise as well as common systematic errors arising, e.g., from radiometric calibration, which affect climate data sets even in the long term, while all other forms of random error average out sooner or later. Errors in aerosol retrieval are considered in a similar way. Errors in cloud detection affect the land cover classification directly. Omission of clouds degrades the accuracy of the ECV dataset whereas false commission reduces its coverage statistics. Our Monte Carlo pre-processing sequence can simulate random and systematic cloud omission and commission errors.
In this contribution, we explain the concept of our Monte Carlo processing sequence and its computational implementation and present proof of concept by verifying the statistical properties of the created surface reflectance ensemble.