SUMMARY AND CONCLUSIONS

By means of the application of the ANN BP algorithm the input signal from calcareous nannoplankton assemblages was predicted in terms of two parameters, SST and ¹⁸O. The same technique could be applied to predictions of any type of physio-chemical variables. The number of samples in both examples was small, but the error estimates obtained from the ANN suggest that even when using small samples highly reliable results can be obtained.

The ability to learn from the examples reveals the goal of the analysis architecture. The examples could also be represented by nonlinear and nonhomogeneous data. In the specific case of a nannoplankton assemblage, the examples could comprise absences of some species in some samples. In the case of the samples from the California Bight, the network has realized the best solution by choosing the samples in such a way as to have represented in the training session all the seasons comprising the time period between October 11, 1991 and July 20, 1992 (the italic word in Table 1). So, in this case, the training set includes the months of October, January, March, May, June, and July, whereas the test set consists of December, March, April, May, June, and July. This arrangement was made from the neural network automatically, without any information during the initialization phase. Thus, the ability of the network to learn from a small, random data set allows the organisation of the data in the correct order and with a correct interpretation.

In the case of the Mediterranean Sea, the network has selected the best configuration based on training and test sets established by the program, and the final chart (Figure 5) shows a good relationship between predicted and observed values. In particular, this chart documents a very similar pattern of change in observed and predicted ¹⁸O values with increasing core depth. This demonstrates the optimum choice of the neural network concerning the training test-set subdivision. In this case, there was a large number of `zero' values in the distribution of E. huxleyi in the lower part of the core. However, this problem did not in any way harm the final result. The neural approach reveals a highly satisfactory result despite the occurrence of many `zero' values as well as with consideration of the relatively few observations in this dataset.

The application of ANN to nannoplankton data could be potentially useful for studies of paleoproductivity and paleoclimate, paleobiogeographic patterns, and paleo-oceanography. For example, some areas of nannoplankton research where ANN might be used involve predictions of sea surface-water temperatures (SST), nutrient and micronutrient composition, species distribution and algal blooms from the analysis of modern calcareous nannoplankton distributions in parts of the World Ocean.