A novel hybrid dragonfly optimization algorithm for agricultural drought prediction

Palmer Drought Severity Index (PDSI) is known as a robust agricultural drought index since it considers the water balance conditions in the soil. It has been widely used as a reference index for monitoring agricultural drought. In this study, the PDSI time series were calculated for nine synoptic stations to monitor agricultural drought in semi-arid region located at Zagros mountains of Iran. Autoregressive Moving Average (ARMA) was used as the stochastic model while Radial Basis Function Neural Network (RBFNN) and Support Vector Machine (SVM) were applied as Machine Learning (ML)-based techniques. According to the time series analysis of PDSI, for the driest months the most PDSI drought events are normal drought and mild drought conditions. As an innovation, Dragonfly Algorithm (DA) was used in this study to optimize the SVM’s parameters, called as the hybrid SVM-DA model. It is worthy to mention that the hybrid SVM-DA is developed as a meta-innovative model for the first time in hydrological studies. The novel hybrid SVM-DA paradigm could improve the SVM’s accuracy up to 29% in predicting PDSI and therefore was found as the superior model. The best statistics for this model were obtained as Root Mean Squared Error (RMSE) = 0.817, Normalized RMSE (NRMSE) = 0.097, Wilmott Index (WI) = 0.940, and R = 0.889. The Mean Absolute Error values of the PDSI predictions via the novel SVM-DA model were under 0.6 for incipient drought, under 0.7 for mild and moderate droughts. In general, the error values in severe and extreme droughts were more than the other classes; however, the hybrid SVM-DA was the best-performing model in most of the cases.