2014 GSA Annual Meeting in Vancouver, British Columbia (19–22 October 2014)

Paper No. 206-1
Presentation Time: 9:00 AM


ASTRAY, Gonzalo, Department of Geological Sciences, Ohio University, Athens, OH 45701, YANES, Yurena, Department of Geology, University of Cincinnati, Cincinnati, OH 45221, LOPEZ, Dina L., Geological Sciences, Ohio Univ, 316 Clippinger Laboratories, Athens, OH 45701 and MEJUTO, Juan C., Department of Physical Chemistry, Universidad de Vigo, Ourense, 32004, Spain, gastray@uvigo.es

A new environmental predictive model using ANN (Artificial Neural Networks) and stable isotope values of land snail shells to determine atmospheric parameters is presented. The modeled environmental variables are mean annual temperature (MAT) and mean annual precipitation (MAP). To develop the model, we have used published field data collected by Yanes et al. (2009) between 2007 and 2008 from Tenerife Island, Spain, including adult land snail shells of 20 different species from 10 to 2,160 m a.s.l. and MAT and MAP from meteorological stations of the National Meteorological Institute in Tenerife. The oxygen stable isotope values (δ18O) of the shell are affected by temperature, relative humidity, and δ18O values of water vapor, dew or meteoric water. The carbon stable isotope values (δ13C) values primarily reflect variations in the δ13C values of the assimilated organic matter (plants), and in some cases, the substrate carbonate or the atmospheric CO2.

Network models were developed to reconstruct MAT and MAP from the input variables; i) land nail species, ii) terrain elevation where the shells were collected, iii) the S or N slope of the island, iv) the shell δ18O value, and v) the shell δ13C value in the snail shells. We used 186 stable isotope snail data to develop the models, and 20 data to validate them. The best model to predict MAT had 5 input variables, eleven neurons in the intermediate layer, and one neuron in the output layer. The average values for both phases (training and validation) for correlation coefficient (R2) and Root Mean Square Error (RMSE) were 0.9999 and 0.0 oC, and an Average Percentage Deviation (APD) of 0.05%. The best neural network models developed for MAP has a topology of 5 input neurons, ten neurons in the intermediate layer and one neurons in the output layer. These models presented an average R2 of 0.9999, with a RMSE of 0.5 mm. and an APD of 0.08%.

Our results suggest that ANN modeling using stable isotope compositions of land snail shells can be used to predict accurately average environmental conditions.