Bayesian neural network modeling of tree-ring temperature variability record from the Western Himalayas

Abstract

A novel technique based on the Bayesian neural network (BNN) theory is developed and employed to model the temperature variation record from the Western Himalayas. In order to estimate an a posteriori probability function, the BNN is trained with the Hybrid Monte Carlo (HMC)/Markov Chain Monte Carlo (MCMC) simulations algorithm. The efficacy of the new algorithm is tested on the well known chaotic, first order autoregressive (AR) and random models and then applied to model the temperature variation record decoded from the tree-ring widths of the Western Himalayas for the period spanning over 1226–2000 AD. For modeling the actual tree-ring temperature data, optimum network parameters are chosen appropriately and then cross-validation test is performed to ensure the generalization skill of the network on the new data set. Finally, prediction result based on the BNN model is compared with the conventional artificial neural network (ANN) and the AR linear models results. The comparative results show that the BNN based analysis makes better prediction than the ANN and the AR models. The new BNN modeling approach provides a viable tool for climate studies and could also be exploited for modeling other kinds of environmental data.