Abstract
Drought is a climate risk that affects access to safe water, crop development, ecological stability, and food production. Therefore, developing drought prediction methods can lead to better management of surface and groundwater resources. Similarly, machine learning can be used to find improved relationships between nonlinear variables in complex systems. Initially, the standardized precipitation evapotranspiration index (SPEI) was calculated, and then using large-scale signals such as large-scale climate signals (the North Atlantic Oscillation, the Arctic Oscillation, the Pacific Decadal Oscillation, and the Southern Oscillation Index), along with climatic variables including temperature, precipitation, and potential evapotranspiration, predictions were made for the period of 1966-2014. Several new machine learning models including Least Square Support Vector Regression (LSSVR), Group Method of Data Handling (GMDH), and Multivariate Adaptive Regression Splines (MARS) were used for prediction. The results showed that in estimating SPEI in moderately arid climates, the GMDH model with criteria (RMSE = 0.26, MAE = 0.17, NSE = 0.95 in validation) under scenario S1 (included all variables plus the SPEI of the previous month) performed better, while in arid and cold climates, the LSSVR model (RMSE = 0.22, MAE = 0.18, NSE = 0.95 in validation) under S1, and in arid and hot climate, the LSSVR model (RMSE = 0.29, MAE = 0.19, NSE = 0.93 in validation) under scenario S2 (included meteorological variables plus the SPEI of the previous month) had higher prediction accuracy. Although the MARS model was less accurate in validation, it showed higher accuracy during calibration compared to the other two models in all climates. The results showed that using large-scale signals for predicting SPEI was beneficial. It can be concluded that machine learning models are useful tools for predicting the SPEI drought index in different climates within similar ranges.