The correction of the atmospheric effects on optical satellite images is essential for quantitative and multi-temporal remote sensing applications. In order to study the performance of the state-of-the-art methods in an integrated way, a voluntary and open-access benchmark Atmospheric Correction Inter-comparison eXercise (ACIX) was initiated in 2016 in the frame of Committee on Earth Observation Satellites (CEOS) Working Group on Calibration & Validation (WGCV). The first exercise was extended in a second edition wherein twelve atmospheric correction (AC) processors, a substantially larger testing dataset and additional validation metrics were involved. The sites for the inter-comparison analysis were defined by investigating the full catalogue of the Aerosol Robotic Network (AERONET) sites for coincident measurements with satellites' overpass. Although there were more than one hundred sites for Copernicus Sentinel-2 and Landsat 8 acquisitions, the analysis presented in this paper concerns only the common matchups amongst all processors, reducing the number to 79 and 62 sites respectively. Aerosol Optical Depth (AOD) and Water Vapour (WV) retrievals were consequently validated based on the available AERONET observations. The processors mostly succeeded in retrieving AOD for relatively light to medium aerosol loading (AOD < 0.2) with uncertainties <0.08, while the overall uncertainty values were typically 0.23 ± 0.15. Better performances were observed for WV retrievals with >90% of the results falling within the suggested empirical specifications and with the Root Mean Square Error (RMSE) being mostly <0.25 g/cm2. Regarding Surface Reflectance (SR) validation two main approaches were followed. For the first one, a simulated SR reference dataset was computed over all of the test sites by using the 6SV (Second Simulation of the Satellite Signal in the Solar Spectrum vector code) full radiative transfer modelling (RTM) and AERONET measurements for the required aerosol variables and water vapour content. The performance assessment demonstrated that the retrievals were not biased for most of the bands. The uncertainties ranged from approximately 0.003 to 0.01 (excluding B01) for the best performing processors in both sensors' analyses. For the second one, measurements from the radiometric calibration network RadCalNet over La Crau (France) and Gobabeb (Namibia) were involved in the validation. The performance of the processors was in general consistent across all bands for both sensors and with low standard deviations (<0.04) between on-site and estimated surface reflectance. Overall, our study provides a good insight of AC algorithms' performance to developers and users, pointing out similarities and differences for AOD, WV and SR retrievals. Such validation though still lacks of ground-based measurements of known uncertainty to better assess and characterize the uncertainties in SR retrievals.
Read full abstract