Evaluation of four atmospheric correction algorithms for MODIS-Aqua images over contrasted coastal waters

Abstract The use of satellites to monitor the color of the ocean requires effective removal of the atmospheric signal. This can be performed by extrapolating the aerosol optical properties to the visible from the near-infra-red (NIR) spectral region assuming that seawater is totally absorbent in this latter part of the spectrum, the so-called black pixel assumption. While this assumption is verified for most phytoplankton dominated waters, it is invalid in turbid waters. Consequently, for the past ten years, several algorithms have been developed on alternative assumptions. Studies comparing these algorithms are of great interest for further improvement in water leaving radiance ( L w ( λ )) retrievals from satellite images explaining the focus of the present research. Four published atmospheric correction algorithms for MODIS-Aqua are compared: (1) the standard NIR algorithm of NASA, (2) the NIR similarity spectrum algorithm, (3) the NIR-SWIR algorithm and (4) an Artificial Neural Network algorithm. The MODIS-Aqua estimated normalized L w ( λ ) are validated with AERONET-Ocean Color data and cruise measurements presenting moderately to highly turbid waters. Based on a match-up exercise, the former three algorithms show the best results in the green region of the spectrum (relative error, RE, between 11 and 20%) and the largest errors in the blue and red region of the spectrum (RE exceeding 30%). In contrast, the Artificial Neural Network algorithm performs better in the red band (RE of 22%). The latter tends to overestimate the normalized L w ( λ ) at all wavelengths while the NIR similarity spectrum algorithm tends to underestimate it. Retrievals of aerosol products, such as the Angstrom coefficient, α(531,869), and the optical thickness, τ (869), present RE above 44% and 72%, respectively. The performance of the algorithms is also investigated as a function of water types. For water masses mainly dominated by phytoplankton, the standard NIR algorithm performs better. In contrast, for water masses mainly dominated by detrital and mineral material, the neural network-based algorithm shows the best results. The largest errors are encountered above water masses dominated by high phytoplankton and CDOM concentrations. This work conducted to a number of perspectives for improving the atmospheric correction algorithms.