Comparison of three multi-site models in stochastic reconstruction of winter daily rainfall over Iran

Abstract

Spatiotemporal modeling of daily rainfall may establish the spatial relationships between different stations which is important issue in hydro-climatology and has been addressed by various studies. The aim of this study is to compare three multi-site models, i.e. hidden Markov model (HMM), non-homogeneous hidden Markov model (NHMM), and K-nearest neighbor model (KNN) in stochastic generation of rainfall data over a network consisting of 130 synoptic stations across Iran using 20-year (1991–2010) daily dataset. Regarding the NHMM, the hidden layer parameters were considered time dependent, such that a layer including predictors was added to HMM. General circulation model (GCM) outputs could be selected as the elements of this layer or predictors. A principle component analysis (PCA) was implemented on four standardized variables of HADGEM2 model (pressure, geo-potential height, temperature, and specific humidity) for historical period on 29 grids across Iran. Accordingly, pressure variable has been selected as a predictor layer in the NHMM according to Bayesian Information criteria (BIC). Results showed that the NHMM has a significant lower BIC compared to the HMM, which confirms the time-dependence assumption of Markov transition probabilities. In stochastic generation process by NHMM and KNN models, different criteria such as seasonal mean and variance, empirical distribution of daily precipitation data in winter, probabilities of wet and dry spells in various sequences and spatial dependency have been compared. According to the results, the NHMM had a significantly better performance compared to KNN model. Nevertheless, no model was able to precisely generate rainfall based on all the evaluation criteria. Log-odds ratio as the spatial correlation criterion illustrated that both NHMM and KNN have promising results in reconstructing spatial relationships. Therefore, both models could be considered as multi-site approaches with a good skill in keeping spatiotemporally changing parameters. The findings of this study may improve the seasonal predictability of rainfall characteristics across the country.