Enhancing a machine learning model for predicting agricultural drought through feature selection techniques

Abstract

This study aims to determine the crucial variables for predicting agricultural drought in various climates of Iran by employing feature selection methods. To achieve this, two databases were used, one consisting of ground-based measurements and the other containing six reanalysis products for temperature (T), root zone soil moisture (SM), potential evapotranspiration (PET), and precipitation (P) variables during the 1987–2019 period. The accuracy of the global database data was assessed using statistical criteria in both single- and multi-product approaches for the aforementioned four variables. In addition, five different feature selection methods were employed to select the best single condition indices (SCIs) as input for the support vector regression (SVR) model. The superior multi-products based on time series (SMT) showed increased accuracy for P, T, PET, and SM variables, with an average 47%, 41%, 42%, and 52% reduction in mean absolute error compared to SSP. In hyperarid climate regions, PET condition index was found to have high relative importance with 40% and 36% contributions to SPEI-3 and SPEI-6, respectively. This suggests that PET plays a key role in agricultural drought in hyperarid regions because of very low precipitation. Additionally, the accuracy results of different feature selection methods show that ReliefF outperformed other feature selection methods in agricultural drought modeling. The characteristics of agricultural drought indicate the occurrence of drought in 2017 and 2018 in various climates in Iran, particularly arid and semi-arid climates, with five instances and an average duration of 12 months of drought in humid climates.