Predicting agricultural drought in central Europe by using machine learning algorithms

Abstract

Understanding the evolution, mechanisms, and trajectories of agricultural drought is an essential strategy for achieving sustainable crop production. Thus, this research evaluates the patterns and magnitude of agriculture droughts using Standardized Precipitation Evapotranspiration Index (SPEI) from 1926 to 2020 in eastern Hungary, and assess the performance of six machine learning models (Random Forest (RF), Extra Trees (ET), Gradient Boosting (GB), Extreme Gradient Boost (XGB), Support Vector Machines (SVM), and Multi-Layer Perceptron (ANN-MLP)) in predicting agriculture droughts. Results showed a decreasing trend of monthly rainfall and decreasing trend of SPEI monthly values indicating more events in the study area. Furthermore, frequency and intensity analysis revealed a total of 18 % of recorded events were classified as SPEI-3 moderate to extreme drought events (in months) with the highest drought intensity (D1 = −2.43) in January 2007. Overall, 16 years were identified as droughts by SPEI-3 to SPEI-12 in the 1st three decades (1926–1956), 12 drought years from 1957 to 1986 and 21 drought years from 1987 to 2020. Among the six machine learning algorithms, the RF model performed the best in the training phase, with the highest R² = 0.75, lowest RMSE = 0.49, and MAE = 0.4. Like the training stage, RF outperformed among other algorithms achieving the highest accuracy. Overall, the ML models can be ranked as RF > XGB > ETR > GBR > ANN-MLP > SVM. The findings of this research promote RF as a reliable algorithm for predicting SPEI droughts.