Abstract
We present a new operational algorithm for the retrieval of water quality from optical remote sensing data for both clear and turbid waters. It contains an array of neural networks providing input for the Levenberg–Marquardt multivariate optimization procedure as the final retrieval tool. With a given accuracy threshold, the developed algorithm is sufficiently robust to data with noise up to 15% for certain hydro‐optical conditions. To avoid inadequate retrieval results, the algorithm identifies and eventually discards the pixels with inadequate atmospheric correction and/or water optical properties incompatible with the applied hydro‐optical model. This procedure also identifies coccolith expressions. Examples of practical applications of the developed algorithm are given.
Acknowledgments
This work was carried out under the projects supported by INTAS (Project 03‐51‐4494 ‘Synergy’), INCO‐COPERNUCUS (Project FP6‐003605 ‘EcoMon’). The authors would like to thank Professor N. Filatov (Northern Water Problems Institute, Petrozavodsk, Russia) and Dr R. Schuchman (Altarum, Ann Arbor, MI, USA) for logistic support and fruitful discussions, as well as the Russian R/V ‘Ecolog’ and Western Michigan University scientists and technicians for water sampling and laboratory analyses. We also extend our gratitude to Mr Are Folkestad (Nansen Environmental and Remote Sensing Centre, Bergen, Norway) for his valuable assistance in training some of the neural networks used in the present study as well as taking part in fruitful discussions.