Remote sensing of chlorophyll-a concentration for drinking water source using genetic algorithms (GA)-partial least square (PLS) modeling

Abstract

Accurate estimation of phytoplankton chlorophyll-a (Chl-a) concentration in turbid waters through remote sensing is a challenge due to the optical complexity of water constituents. Reflectance spectra and concurrent water quality parameters of 225 samples across the Shitoukoumen Reservoir, the drinking water resource for Changchun City, were used to retrieve Chl-a concentration with high total suspended matter (TSM) during 2006–2008. A combination of genetic algorithms and partial least square (GA-PLS) model was established for Chl-a retrieval through GA to select sensitive spectral variables and PLS for regression. To compare GA-PLS performances, the widely accepted three-band algorithm was implemented for Chl-a concentration estimation. Both GA-PLS and the three-band algorithm have stable performance for the aggregated dataset (R2=0.85 and 0.81; RPD=3.95 and 3.61; relative RMSE=31.7% and 34.2%), with the GA-PLS model performing marginally better. The temporal transferability of the models was validated with the dataset collected in 2006 and 2007 respectively as independent dataset, showing that GA-PLS outperformed the three-band algorithm. Our result also indicated that relative error [(Chl-apredicted−Chl-ameasured)/Chl-ameasured] showed good linear relation to TSM: Chl-a ratio (R2=0.84), which implied that TSM concentration exerted significant impact on the accuracy of Chl-a estimation in this case study. As the results were derived from a large number of samples representing a wide range of spatiotemporal variations of pigment under TSM (3.7–472.8mg/L) concentration influence, the GA-PLS model has great potential for Chl-a estimation for inland waters with similar backgrounds. Nevertheless, the three-band algorithm also has its own merit considering its simplicity for implementation.