Using neural networks to predict surface zooplankton biomass along a 50 degrees N to 50 degrees S transect of the Atlantic

Four Atlantic transects between the UK and the Falkland Islands were carrie d out during spring and autumn as part of the Atlantic Meridional Transect (AMT) programme. These 50 degreesN to 50 degreesS transects cross several o cean regions. An optical plankton counter (OPC-1L) sampled continuously alo ng the transects from the ship's uncontaminated seawater supply, giving a s urface distribution of zooplankton abundance and size. Measurements of unde rway fluorescence-derived chlorophyll, sea surface temperature and salinity were also taken from the uncontaminated seawater supply. The relationship between zooplankton biomass and these variables was investigated using mult iple linear regression and neural network techniques. In the analysis, loge -transformed biomass was used to reduce the influence of extreme values. Tw o transects were used to develop the models, and two to test the generaliza tion capabilities of the models. Multiple linear regression could explain u p to 55% of the observed variation in the transformed biovolume, and demons trated the association of hydrographic variables and diel migration within the surface zooplankton community. Neural networks, starting with the same set of variables, were able to explain up to 78% of the variability, showin g an increased performance over the multivariate analysis. An optimized mod el accounted for 77% of the variance in the original data. However, it show ed greater generalization capabilities (R-2 = 0.47) when applied to new dat a sets than either the original neural network model (R-2 = 0.31) or the mu ltiple linear regression model (R-2 = 0.34). This study highlights the non- linear nature of the parameters' associations with the zooplankton biomass and their variability between oceanographic regions.