Neural network models of phytoplankton primary production (1998).

Stazione Zoologica "A. Dohrn" di Napoli, Villa Comunale, 80121 Napoli, Italy

Empirical models of phytoplankton primary production play an important role in oceanographic research, mainly because direct measurements of this variable are difficult and time consuming. First of all, they are needed to fully exploit the large phytoplankton biomass data sets that are obtained by remote sensing, but they are also necessary to carry out instrumental estimates of primary production (e.g. by pump and probe fluorometers) and to post-process the field data.

Many empirical models have been developed and several of them provided useful results, but the oceanographic data sets are growing larger and demand more effective approaches. In this framework, artificial neural networks have been recently proposed as empirical models of phytoplankton primary production. Even though only very simple error back-propagation neural networks were used, they provided very good results and always performed better than conventional empirical models.

Some examples of neural network-based empirical models of phytoplankton primary production will be presented and their results will be compared with those provided by conventional models. These examples will represent a wide spectrum of complexity and of spatial scales, ranging from the local to the global ones. All the models will have a single hidden layer and a single output, whereas hidden and output nodes will have sigmoid activation functions. The neural network training will be performed with a constant learning rate and no momentum term.

One of the main problems in the application of neural networks to ecological modeling is their tendency to overfit the training data patterns. Several strategies will be applied in order to obtain a good generalization (i.e. to train neural network that act as models rather than as memories) and thoroughly discussed.

A major advantage of neural network models of phytoplankton primary production is their capability of incorporating information that is difficult to manage with conventional models (e.g. binary or nominal data, geographical coordinates, etc.), but that tends to co-vary with the output variable. The covariation has not to be linear, because neural networks can deal with non-linear relationships more easily than other empirical models. Some examples of the exploitation of such additional information will be presented.

Finally, the role of sensitivity analysis of the neural network models will be discussed. In particular, sensitivity analysis makes possible to assess the strength of the links between predictive variables and phytoplankton primary production in the observed data sets rather than in the framework of an a priori, simplified theoretical model. Therefore, it can provide a deeper insight into the dynamics of phytoplankton primary production under real world conditions.