Advanced Machine Learning Model for Prediction of Drought Indices using Hybrid SVR-RSM

Drought, as a phenomenon that causes significant damage to agriculture and water resources, has increased across the globe due to climate change. Hence, scientists are attracted to developing drought prediction models for mitigation strategies. Different drought indices (DIs) have been proposed for drought monitoring during the past few decades, most of which are probabilistic, highly stochastic, and non-linear. The present study inspected the capability of various machine learning (ML) models, including artificial neural network (ANN) and support vector regression (SVR) as original predictive models and optimized by two selected algorithms, namely, particle swarm optimization (SVR-PSO) and response surface method (SVR-RSM) to predict the meteorological drought indices of standardized precipitation index (SPI), percentage of normal precipitation (PN), effective drought index (EDI), and modified China-Z index (MCZI) on a monthly time scale. A novel model named SVR-RMS is introduced by using two calibrating processes given from RSM with two inputs and the SVR by predicted data handled with RSM given from the first calibrating procedure. For evaluating the models, different meteorological input variables in the period 1981-2020 were considered from 11 synoptic stations in arid and semi-arid climates of Iran, which frequently experience droughts. The SPI showed the highest and lowest correlation with MCZI (0.71) and EDI (0.34), respectively. The results of testing dataset (2011-2020) indicated that the SVR-RSM produced superior abilities for both accuracy and tendency compared to other models, while the SVR-PSO model is better than the ANN and SVR. The worst results of drought prediction were obtained for EDI. However, all models provided the acceptable EDI prediction in the high-temperature station of Ahvaz in the south of the country. Application of SVR-RSM as a novel hybrid model can be suggested for predicting the DIs on a short time scale in arid and semi-arid areas.