Meteorological Drought Modelling by Explainable Artificial Intelligence

The accurate prediction of drought events is important to relieve the negative consequences on water resources, ecosystems, and agriculture. Machine learning algorithms require less time and minimum input and are promising methods to predict drought because it is relatively less complicated than conventional models. In this study, the standardized precipitation index (SPI) is calculated for 3-, 6-, 9- and 12- months by using the monthly precipitation data of the Isparta, Eğirdir, Senirkent, Uluborlu, and Yalvaç meteorology measurement stations in Isparta, Turkey between 1964-2019. The calculated SPI values were used to develop the drought models by machine learning algorithms (Random Forest, K-Nearest Neighbors, Extra Trees Regression, Support Vector Regression, Gradient Boosting, and Extreme Gradient Boost). The R2 and RMSE values are used in the assessment of the model results. In general, these algorithms gave good results for all calculated drought periods. Also, the results showed that the best result is obtained as 0.80 R2 by using Extra Trees algorithms for testing set of 12- month model. The relationship between model results and input parameters was interpreted using the Shapley Additive exPlanations algorithm one of the explainable artificial intelligences. It was seen that the most effective input on the Isparta station SPI values is Senirkent station.