Fan, YZ; Li, SQ; Han, XZ; Stamnes, K (2020). Machine learning algorithms for retrievals of aerosol and ocean color products from FY-3D MERSI-II instrument. JOURNAL OF QUANTITATIVE SPECTROSCOPY & RADIATIVE TRANSFER, 250, 107042.

Heritage atmospheric correction (AC) and ocean inherent optical property (IOP) retrieval algorithms, such as those implemented in NASA's SeaDAS platform, often produce questionable results in complex environments, such as in turbid coastal and inland waters and for heavy aerosol loadings. We present new AC and ocean IOP retrieval algorithms for the Medium Resolution Spectral Imager - II (MERSI-II) onboard the FengYun-3D satellite that solve these problems. The algorithm development is based on extensive radiative transfer simulations for a coupled atmosphere-ocean system in conjunction with machine learning techniques (i.e. multi-layer neural networks) to retrieve ocean color products from MERSI-II sensor data. The final ocean color products include spectral remote sensing reflectances (R-rs(lambda) values), chlorophyll_a concentration (CHL), and IOPs, i.e. absorption by phytoplankton (a(ph)(lambda)), absorption by detritus and gelbstoff (a(dg)(lambda)), and particulate backscattering (b(bp)(lambda)). Spectral aerosol optical depths (AODs) and cloud mask results are also provided. The new machine learning based algorithms are first tested using independent synthetic datasets and show very good performance. The average percentage error (APE) is less than 7% for the R-rs retrievals and less than 5.2% for the AOD retrievals with a 1% uncertainty added to the TOA reflectances. The ocean color products retrieved from MERSI-II sensor data are validated against AERONET-OC field measurements and show good quality. For Rrs retrievals, the APE is around 30% for blue and red bands and 22% for green band. For AOD retrievals, the APE is also around 23% - 30%, and for CHL retrievals, the APE is around 32%. The ocean color products retrieved from MERSI-II sensor data are also cross-validated with the corresponding ones retrieved from Aqua/MODIS data using both NASA SeaDAS package and our machine learning based (DC-SMART algorithms. The results show that these machine learning algorithms completely resolve the negative R-rs(lambda) issue that persists in heritage AC algorithms and significantly improve the quality of retrieved ocean color products, especially in coastal regions. The histograms of the retrieved R-rs, CHL, and ocean IOP products also show good agreement between MERSI-II and Aqua/MODIS ocean color retrievals. (C) 2020 Published by Elsevier Ltd.