Evaluation of machine learning models in predicting drought indicators (Case Study: Ajabshir area)

Drought is one of the destructive phenomena with adverse impacts on water resources and water needs. Machine-learning models are among the helpful tools in time-series prediction that can provide suitable results without the requirements for basic information about a system. In this study, adaptive neuro-fuzzy inference system (ANFIS) and least square support vector regression (LSSVR) models were utilized to predict the standardized precipitation index (SPI) as a meteorological drought indicator and streamflow drought index (SDI) as a hydrological drought indicator for a period (2001-2019). Ajabshir, located in the northwest of Iran, was selected as the study area, where the data of Qaleh Chay meteorological and hydrological stations were used to calculate SPI and SDI, respectively. The precipitation and flow rate data were considered input variables of the machine-learning models in predicting the SPI and SDI, respectively. The results revealed that during the period under review, meteorological drought was more severe in 2004-2011. While in this period, hydrological drought was more severe in 2007-2011 (SPI<-3). Moreover, the prediction results of the indices showed that the performance of the LSSVR model was better than that of ANFIS for both indicators. Using LSSVR, the RMSE and MAPE error evaluation criteria for SPI were 0.74 and 0.59, respectively, while these values for SDI were obtained as 0.62 and 0.46, respectively. The findings of this study show that machine-learning models are suitable tools for predicting drought indicators. Therefore, it is suggested to use such models in predicting drought indicators in other similar regions.