Using Partial Least Squares-Artificial Neural Network for Inversion of Inland Water Chlorophyll-a

Accurate remote estimation of chlorophyll-a (CHL) concentration for turbid inland waters is a challenging task due to their optical complexity. In situ spectra (n = 666) measured with ASD and Ocean Optics spectrometers from three drinking water sources in Indiana, USA, were used to calibrate the partial least squares model (PLS), artificial neural network model (ANN), and the three-band model (TBM) for CHL estimates; model performances are validated with three independent datasets (n = 360) from China. The PLS-ANN model resulted in accurate model calibration (R-2 = 0.94; Range = 0.2-296.6 mu g/l of CHL), outperforming the PLS (R-2 = 0.87), ANN (R-2 = 0.91), and TBM (R-2 = 0.86). With an independent validation dataset, the PLS-ANN yielded relatively high accuracy (RMSE: 6.12 mu g/l; rRMSE = 42.12%; range = 0.45-97.2 mu g/l of CHL), while TBM yielded acceptable accuracy (RMSE: 8.85 mu g/l; rRMSE = 63.21%). With simulated ESA/MERIS and EO-1/Hyperion spectra, the PLS-ANN also (MERIS: R-2 = 0.84; Hyperion: R-2 = 0.88) outperforms the TBM (MERIS: R-2 = 0.69; Hyperion: R-2 = 0.76) for model calibration. For validation, the PLS-ANN achieves good performance with simulated spectra (MERIS: RMSE = 7.83 mu g/l, rRMSE = 48.79%; Hyperion: RMSE = 6.98 mu g/l, rRMSE = 45.57%) as compared to the TBM (MERIS: RMSE = 10.39 mu g/l, rRMSE = 68.92%; Hyperion: RMSE = 9.54 mu g/l, rRMSE = 65.35%). Nevertheless, considering the large and diverse datasets, the TBM is a robust semiempirical algorithm. Based on our observations, both the PLS-ANN and TBM are effective approaches for CHL estimation in turbid waters.