Improved teleconnective predictability of monthly precipitation amounts using canonical correlation analysis

Abstract

This study was aimed at evaluating the application of the canonical correlation analysis (CCA) to predict monthly precipitation amounts (predictands) by benefitting from 17 large-scale climate indices (predictors) in Iran. Monthly precipitation data, covering the period of 1987–2017, were collected from 100 weather stations across the country. Monthly precipitations were predicted using the multiple linear regression (MLR) models, based on the 1- to 6-month lead times of the original and canonical predictors. The cross-validation was conducted to compare the prediction skills of the two sets of MLR models constructed on the basis of the original predictors (MLOrigPr) and the canonical predictors (MLCCAPr). The analyses revealed the dominant teleconnections and that there are the interannual variations in responses of precipitation to them suggesting that a signal only is not sufficient to achieve a robust understanding of the associations. At the 1-month lead time, the MLR models based on the canonical predictors outperformed those based on the original predictors. However, the skill of both models was reduced by increasing the lead times up to 6 months. Averaging on all stations, around 61.4% and 26.3% of the observed values, falls into the cross-validated 95% prediction intervals of the MLCCAPr and MLOrigPr models, respectively. Furthermore, the MLCCAPr models were found to be more spatially universal than the MLOrigPr ones and decrease multicollinearity symptoms strengthening the predictions. These findings corroborated the advantage of using the CCA in improving the teleconnective predictability of precipitation in Iran.