Assessment of NIR-red algorithms for observation of chlorophyll-a in highly turbid inland waters in China

Abstract

It has been proven that empirical (two-band) and semi-analytical (three-band) algorithms based on near-infrared and red (NIR-red) wavelengths can be used for estimating Cchl-a in highly turbid productive waters with satisfactory performance. However, the optimal spectral bands and parameters of algorithms vary significantly because of the different optical properties of datasets. Using a comprehensive dataset, we validate and evaluate the applicability of empirical and semi-analytical algorithms for deriving Cchl-a for inland lakes in China. The comprehensive dataset contains 993 in situ samples collected from five inland lakes in China between 2006 and 2013. The optimal algorithms, Rrs(706)/Rrs(685) and [Rrs-1(685)-Rrs-1(707)]Rrs(722), are calibrated using an in situ dataset, with root mean square errors (RMSEs) of 10.66mg/m3 and 8.47mg/m3, respectively. The RMSEs of the NIR-red two- and three-band algorithms for the validation data are 11.1mg/m3 and 8.82mg/m3, respectively. The RMSEs increase to 13.17mg/m3 and 12.58mg/m3 when the algorithms are applied to Medium Resolution Imaging Spectrometer (MERIS) and Geostationary Ocean Color Imager (GOCI) centre wavelengths. The RMSEs for the validation data decrease to 8.80mg/m3 and 7.78mg/m3 when the optimal spectral band (λ1) shifts to 671nm. The RMSEs decrease to 10.03mg/m3 and 9.09mg/m3 as a result of optimization of the model parameters when the algorithms are applied to MERIS and GOCI centre wavelengths. The shifting of the optimal spectral band (the difference between 671nm and 685nm) increases the RMSEs from 8.80mg/m3 to 11.1mg/m3 for the two-band algorithm, and slightly increases the RMSEs from 7.78mg/m3 to 8.82mg/m3 for the three-band algorithm. This indicates that the three-band algorithm is much more suitable for high-turbidity water than the two-band algorithm. Nevertheless, the two-band model can be used for extremely turbid and low Cchl-a waters for analysis of the retrieval results after cluster analysis of remote sensing reflectance. Meanwhile, shifting of the optimal spectral bands (λ1) is highly correlated with the total suspended matter concentration (CTSM) (the Pearson correlation coefficient between λ1 and CTSM can reach 0.95). In conclusion, the results indicate that both the two- and three-band algorithms have high potential applicability for derivation of Cchl-a in high-turbidity inland waters in China.