Application of back propagation neural network in paleoclimate

Abstract

Studies of paleoclimate variations in local regions are seriously restricted by the low resolution and uncertainties of the simulated data at present. In order to apply large-scale modeling data to paleoclimate research in local regions, an effective downscaling model based on three-layer back propagation neural network (BPNN) is developed. Observational and ECHO-G simulated data are employed to train and test the BPNN model. With proper training and validation, BPNN model exhibits its ability to paleoclimate estimation, it is applied to reconstruct monthly (January and July) and annual mean temperature and precipitation in Anhui-Hubei region during the last millennium. The results indicate that BPNN model extracts useful climatic information from observation and simulation and provides fairly accurate paleoclimate estimation. This downscaling method is a successful trial of applying BPNN in local area of paleoclimate modeling, in the meantime, it improves the capacity of researching on paleoclimate variability in local regions using large-scale modeling data.