Improving remote sensing retrieval of water clarity in complex coastal and inland waters with modified absorption estimation and optical water classification using Sentinel-2 MSI

•A new ZSD retrieval model used for complex coastal and inland waters was proposed.•The model raised the ZSD accuracy by altering absorption coefficient and water types.•The model can be applied to various sensors (MSI, MODIS, MERIS, OLCI and GOCI). The estimation of water clarity in coastal and inland waters from satellite images is limited by significant uncertainties attributable to the optical complexity of these waters. To address this issue, we propose an improved semi-analytical model for retrieving the Secchi disk depth (ZSD) based on the mechanistic model (Lee model) developed by Lee et al. (2015). In the Lee model, a quasi-analytical algorithm (QAA) was employed to estimate the total absorption coefficient (a(λ)) and total backscattering coefficient (bb(λ)) to retrieve the diffuse attenuation coefficient (Kd(λ)). ZSD was then estimated by remote sensing reflectance (Rrs) and Kd. In this study, two improvements were made to the Lee model based on the band specifications of the Sentinel-2 Multispectral Instrument (MSI). First, the reference wavelength (λ0) was shifted from 560 nm and 665 nm to the longer red edge band 705 nm, and the modified empirical formula for estimating a(λ0) was applied to improve ZSD retrieval. Second, waters were classified into different optical water types (OWTs), and the optimal ZSD algorithm was identified for each OWT to further improve the accuracy of ZSD estimation. The proposed model (denoted as QAA_OWT) was calibrated and evaluated with in situ data (N = 399) collected from waters with various optical properties including a lake (Daihai Lake) dominated by colored dissolved organic matter (CDOM), a macrophytic lake (Ulansuhai Lake), and coastal water (the Bohai Sea) dominated by suspended sediments. The validation showed that QAA_OWT improved the accuracy of ZSD retrieval (0.18–5.0 m), with the averaged percentage difference (MAPD) decreasing from 31.9% to 24.9% and the root mean square deviation (RMSD) decreasing from 0.38 m to 0.32 m. The model was further applied to satellite images from the Sentinel-2 Multispectral Instrument (MSI) and independently assessed based on in situ-satellite match-ups (N = 31). The results verified the reliability of the model (MAPD = 31.8%, RMSD = 0.16 m). Time series MSI-derived ZSD maps clearly characterized the spatio-temporal patterns of the water clarity of Daihai Lake. Such maps are of significance in the accurate evaluation of the effectiveness of pollution control measures and