Journal of Plankton Research Vol.24 no.12 pp.1289-1303, 2002
© Oxford University Press 2002
Identifying characteristic chlorophyll a profiles in the coastal domain using an artificial neural network
1 Oceanography Department and 2 Zoology Department, University of Cape Town, Rondebosch 7701, Cape Town, South Africa
* Corresponding Author: anr{at}mail.pml.ac.uk
3 Present Address: The Sir Alister Hardy Foundation for Ocean Science, The Laboratory, Citadel Hill, Plymouth Pl1 2pb, UK
To estimate primary production in the marine environment, knowledge of the vertical distribution of phytoplankton is needed. The measurement of ocean colour by satellite remote sensing makes it possible to map the near-surface phytoplankton distribution, although the subsurface vertical structure cannot be measured. In this study, we investigated the shape of vertical chlorophyll profiles from the Benguela upwelling system seasonally and with respect to environmental variables such as surface temperature and chlorophyll, and mixed layer and water column depth. We used an artificial neural network technique called a self-organizing map to identify characteristic classes of vertical chlorophyll profiles, and then classify existing profiles into these representative classes. The self-organizing map identified a continuum of patterns, ranging from those with small deep peaks (~1 mg m-3, 45 m) to those with large near-surface peaks (~9 mg m-3, 7 m). Although profile shape varied seasonally, profiles were very variable within each season, making chlorophyll profiles averaged seasonally meaningless. A canonical succession in profile shape following upwelling of cold water in spring and summer could be identified, with large surface peaks in cool water and small deep peaks in warm water. The approach presented here can be used in a semi-quantitative manner to predict the subsurface chlorophyll field from known (water column depth) or easily measured variables from satellites (surface temperature or surface chlorophyll), as the relative frequency of each characteristic profile under different environmental conditions is presented. This approach enables prediction of profile shapes in the dynamic coastal domain and thus superior regional estimates of primary production.