Multivariate Drought Forecasting in Short- and Long-Term Horizons Using MSPI and Data-Driven Approaches

Abstract

A simultaneous survey of several types of droughts, such as meteorological, hydrological, agricultural, economic, and social droughts, is possible by using the multivariate standardized precipitation index (MSPI). In this study, the accuracy of four artificial intelligence (AI) methods, including the generalized regression neural network (GRNN), least-square support vector machine (LSSVM), group method of data handling (GMDH), and adaptive neuro-fuzzy inference systems with fuzzy C-means clustering (ANFIS-FCM), were investigated in forecasting the MSPI of three synoptic stations (Jolfa, Kerman, and Tehran) located in the arid-cold climate of Iran. The data used was monthly precipitation and belongs to a 30-year period (1988–2017). MSPI values were calculated in five time windows, including the following: 3–6 (MSPI3–6), 6–12 (MSPI6–12), 3–12 (MSPI3–12), 12–24 (MSPI12–24), and 24–48 (MSPI24–48). The period of 1988–2016 was considered for training (75%) and testing (25%), and 2017 (12 months) was used for long-term forecasting. The methods were evaluated by the root mean square error (RMSE), mean absolute error (MAE), Willmott index (WI), and Taylor diagram. In the short-term forecasting phase, results showed that the methods had their best performances in forecasting multivariate drought types of groundwater hydrology-economic-social (MSPI24–48), agricultural-groundwater hydrology (MSPI12–24), surface hydrology-agricultural (MSPI6–12), soil moisture-surface hydrology-agricultural (MSPI3–12), and soil moisture-surface hydrology (MSPI3–6), respectively. Also, among the mentioned methods, the weakest accuracy was reported for GRNN with an RMSE=0.673, MAE=0.499, and WI=0.750 (related to MSPI3–6 of the Kerman station); the most accurate performance resulted from the GMDH with RMSE=0.097, MAE=0.074, and WI=0.989 (related to MSPI24–48 of the Jolfa station). In spite of the acceptable performance of the models in short-term forecasting, by increasing the forecasting horizons, the models’ errors were increased in the long-term forecasting phase. The models could have acceptable long-term forecasts for just two months (or in some exceptional cases, three months) ahead. Further, according to the investigations, it can be shown that the methods show better performances in mountainous arid-cold regions, compared to desert arid-cold regions. As a theoretical study of multivariate drought forecasting, the AIs have promising results, and this research can be extended for the other regions.