Application of support vector machine and bayesian network for agricultural drought prediction

Awareness of the drought status and the prediction of its future conditions play an important role in water resources management programs. In this regard, rainfall and temperature variables have a great influence on the severity and duration of this phenomenon. Regarding the status of the Urmia Lake in recent years and the water stress in its watershed, in this study, the drought situation in Saghez synoptic station as one of the important stations of this basin in different time-scales using the Standardized Evapotranspiration Index (SPEI) and SVM model with three linear, polynomial, and radial basis function and Bayesian network (BN) models, were investigated. For this purpose, the SPEI index in the short-term (1 and 3 months), mid-term (6, 12-months) and long-term (24 and 48-months) during the 49-year statistical period for monitoring the drought status at this station was used. Results showed that there was 8 prolonged periods of drought for the years 1962-1968, 1972-1974, 1978-1979, 1980-1982, 1983-1984, 1986-1987, 1999-2003 and 2007-2009 during the statistical period. Then SPEI values were applied to five input models with a delay of 1 to 5 months and SVM and BN models were used to predict drought. The results showed that in both methods, the model with 5-time delay had better performance and the linear basic function in the SVM method was more accurate than the other two functions. Also, the predictive accuracy of these models is directly correlated with increasing the SPEI scale, so that the correlation coefficient in the Bayesian network method at the test stage ranged from 0.174 in 1-month time-scale to 0.985 on a 48-month time-scale and in the SVM method with a linear basic function, it has risen from 1.149 to 0.983.