Comparison of the accuracy of SST estimates by artificial neural networks (ANN) and other quantitative methods using radiolarian data from the Antarctic and Pacific Oceans

• Main Authors: Gupta, S.M., Malmgren, B.A.
• Format: Article in Journal/Newspaper
• Language: English
• Published: The society of Earth Scientists 2009
• Subjects: surface temperature; radiolarian ooze; artificial neural networks; sea surface temperatures; Antarctic; Antarctic Ocean; Pacific; The Antarctic; Antarc*
• Online Access: http://drs.nio.org/drs/handle/2264/3346

The quality of quantitative estimates of sea surface water temperatures (SSTs) is evaluated for different techniques, Imbrie-Kipp transfer functions (IKTF), the modern analog technique (MAT), weighted-averaging partial least squares (WAPLS) regression, the maximum likelihood (ML) method, and artificial neural networks (ANNs), based on radiolarian faunal abundance data from surface sediments from the Antarctic and Pacific Oceans. Recent studies have suggested that ANNs may represent one of the most optimum procedures for estimates of paleo-SSTs. It is therefore, employed ANNs together with these other methods to estimate SSTs during February-April (TFA) and August-October (TAO) at a mean water depth of 40 m, wherein radiolarian abundances generally coincide with chlorophyll-a maximum. We used CLIMAP’s modern Antarctic radiolarian core top data and Pisias et al.’s Pacific Ocean core top data for the analyses. A portion of the datasets (75%) was used for training of ten ANNs per season, and estimates of error rates (root-mean-square-errors of prediction, RMSEPs) were made from the remaining observations, constituting an independent holdback (HB) set. The same training and HB sets were used for estimates of RMSEPs for the other methods. For the ANNs the RMSEP in the Antarctic Ocean dataset is as low as approx. 1.3 degrees C for both TFA and TAO. In comparison, RMSEPs for the other techniques for TFA are higher in ranging between 1.8 and 2.0 degrees C, whereas those for TAO are similar (1.4-1.5 degrees C). Correlation coefficients (r:s) between observed and predicted SSTs using the ANNs are 0.97 for both seasons. In the Pacific Ocean dataset, RMSEPs derived from the ANNs are considerably lower for both seasons, 1.5 degrees C for TFA (1.8-2.2 degrees C for the other methods), and 1.4 degrees C for TAO (the other methods 1.7-1.9 degrees C). ANN-derived correlation coefficients (r:s) between observed and predicted SSTs are 0.98 for both TFA and TAO in the Pacific Ocean. Comparison of residual (estimated-observed) SST maps suggests that MAT and ANN produced lesser geographic trends than those of the other methods