Comparison of statistical and artificial neural network techniques for estimating past sea surface temperatures from planktonic foraminifer census data

We present the first detailed and rigorous comparison of six different computational techniques used to reconstruct sea surface temperatures (SST) from planktonic foraminifer census data. These include the Imbrie‐Kipp transfer functions (IKTF), the modern analog technique (MAT), the modern analog technique with similarity index (SIMMAX), the revised analog method (RAM), and, for the first time, a set of back propagation artificial neural networks (ANN) trained on a large faunal data set, including a modification where geographical information was added among the input variables (ANND). By training the techniques on an identical database, we were able to explore the differences in SST reconstructions resulting solely from the use of different mathematical methods. The comparison indicates that while the IKTF technique consistently shows the worst performance, ANN and RAM perform slightly better than MAT and that the inclusion of the geographical information into the training database (SIMMAX and ANND) further improves the accuracy of modern SST estimates. However, when applied to an independent validation data set and an additional fossil data set, the results did not conform to this ranking. The largest differences in the reconstructed SST values occurred between groups of techniques with different approaches to SST reconstruction; that is, ANN and ANND produced SST reconstructions significantly different from those produced by RAM, SIMMAX, and MAT. The application of the various techniques to the validation data set, which allowed comparison of SST reconstructions with instrumental records, suggests that artificial neural networks might provide better paleo‐SST estimates than the other techniques.