Comparison of skill of two spatial-temporal approaches of daily rainfall simulation across Iran

The aim of this study is a comparison among two multi-site stochastic weather generators for simulation of winter rainfall occurrence across Iran using data of a selected network consisting of 130 rain gauge stations with a historical data of 21 years. The applied approaches included Hidden Markov Model (HMM) as a parametric approach and K-nearest neighbor (KNN) as non-parametric approach. Six stations namely, Bandar Anzali, Sari, Gharakhil Ghaemshahr, Gorgan, Shiraz and Zahedan were chosen respectively as the representative of different climates including very humid, humid, semi humid, Mediterranean, semi dry and dry climates. In comparison of first and second order momentums, results indicated that HMM performed well in almost every station. Data dispersion was examined using box plot and confidence interval analysis. The results revealed better performance for HMM. Regarding probabilities spaces, HMM showed a better performance in simulation of extreme events and higher percentiles of empirical distribution but KNN approach provided better estimations for middle percentiles values. LEPS Score index was used for comparison of cumulative distribution of observed and simulated series which showed more agreement in case of HMM. The spatial correlation was evaluated using Log-odds ratio index, which indicated that KNN model did better. Both approaches performed well in estimation of duration of wet and dry spells though a tendency to overestimate was observed at HMM and a tendency to underestimate viewed at KNN in simulating of wet spells. In general, HMM has more skill in simulation of daily rainfall series which might be attributed to its complex mathematical structure, however relatively good results of KNN approach showed that it can be recommended for less complicated applications.