Application of Gaussian process regression to forecast multi-step ahead SPEI drought index

Abstract

Forecasting of drought can be very useful in preparing to reduce its impacts, especially in the agricultural sector. Three machine learning models of MLP neural network, GRNN neural network, and Gaussian process regression were used to forecast the annual drought index (SPEI12) in intervals of 1 to 3 months ahead in 79 Iranian synoptic stations. For test results, the accuracy of the models was assessed based on RMSE and R2. Synoptic stations were divided into 5 clusters (C1 to C5) based on the drought time series using the K-means algorithm. The findings showed that the accuracy of the models has declined in all three approaches, with the predictive period increasing from one to three months ahead. Across all three forecasting intervals and in all 5 clusters, the Gaussian process regression (GPR) model achieved the lowest RMSE and the highest R2 values. Compared with the other two models used in this analysis, the MLP model had worse results, and the GRNN model had a score between the other two models. The models as mentioned above had the best prediction in cluster 1 (southern and southeastern regions of Iran), and the lowest model accuracy was observed in cluster 5 (Caspian Sea Shores).

Keywords

• Drought
• SPEI index
• Forecasting
• Machine learning
• Gaussian process regression