Artificial Neural Networks to estimate Paleotemperatures in North Patagonia (Argentina) based on micromammals sequences.

Analía Andrade, Joan Anton Barceló, Florencia Del Castillo


Small mammal’s assemblages from archaeological sites were employed worldwide as proxy data to reconstruct paleoenvironments. However, the scope of these researches allow indirect paleoclimate inferences. The aim of this work is based on the use of neural networks (NN) to predict paleotemperatures during Middle and Late Holocene in northern Patagonia based on the presence/absence of rodent species from stratigraphical sequences.
The study area is the Natural Protected Area of Somuncura, a massive volcanic plateau located in the Extra-Andean Patagonia, in which successive basaltic flows, step-like landform, set an elevational gradient between 600 and 1800 m a.s.l. At present, the annual average temperature decreases with the altitude (linear model, r=-0.96376, m= -0.0061, b= 15.456) but the annual average precipitation remains slow and constant (187 mm). This steep altitudinal gradient configures the assemblage composition of small mammals and plants along the gradient.
Small mammals contained in diverse samples stem from owls’ regurgitate balls and recovered along this gradient, and the temperature estimated at each locality by the regression model were employed as the actual correlate of the network. NN allowed discriminating species that distribute in colder environments (and upper levels) and those from wormer environments (and lower levels). The temperature parameters used to build the network were inferred from the sequence of micromammals with a predicted limit temperature of 9.5°C. Results show that during colder periods micromammals from the current upper levels would have occupied lower levels, while the opposite would have happened in warmer times, micromammals from lower levels have occupied the upper plateau. The model reproduce well the temporal correlation of the observed data and indicates that the spatio-temporal informations and an accurate classification of micromammals is essential for NN predictive time series with incomplete data bases.