APPLICATION OF ARTIFICIAL NEURAL NETWORKS TO PALEOCEANOGRAPHIC DATA

Artificial neural networks are computer systems that have the ability to 'learn' a set of output, or target, vectors from a set of input vec tors. Learning is achieved by self-adjustment of a set of parameters t o minimize the error between a desired output and the actual network o utput. We have explored the potential of this approach in paleoceanogr aphy by application of a neural network algorithm to a problem involvi ng prediction of sea surface-water temperatures from relative abundanc es of planktonic foraminifer species in the southern Indian Ocean. We employed a backpropagation (BP) network to assess how well it was able to predict the actual summer and winter surface-water temperatures. W e compared the results with those obtained from statistical methods pr eviously used for temperature predictions: Imbrie-Kipp Transfer Functi ons, the Modern Analog Technique, and Soft Independent Modelling of Cl ass Analogy. The efficiency of predictions was tested using the Leavin g One Out technique in which each of the observations in the data set is left out one at a time, while the remaining observations are used t o generate a predictor. The accuracy of the predictor is then tested o n the observation left out by comparison with its actual value. A set of tests using 1, 2, 3, 4, 5, and 10 neurons (processing elements) in a 3-layer BP network showed that a network with 3 neurons gave the sma llest errors of prediction for both summer and winter temperatures, 0. 71 and 0.76, respectively. Corresponding errors for the statistical pa ttern-recognition techniques ranged between 1.01 and 1.26 for summer t emperatures and 1.05-1.13 for winter temperatures. Hence, predictions of paleotemperatures from new data on planktonic foraminifer relative abundances in the southern Indian Ocean may be made with a precision o f +/-0.7-0.8 degrees C using the BP network and +/- 1.0-1.3 degrees C using the statistical pattern-recognition procedures. The BP network w as thus the most successful among the methods employed here for temper ature predictions. Artificial neural networks may, therefore, be seen as a viable alternative to more conventional approaches to data analys is in paleoceanography. (C) 1997 Elsevier Science B.V.