Comparing the efficiency of several support vector regressions in predicting agricultural drought using teleconnection and climatic parameters in Iranian climatic samples

Drought is a complex phenomenon which is difficult to define. It is a creeping phenomenon that slowly sneaks up and impacts many sectors of the economy, and operates on many different time scales. Although low rainfall is the source of any drought however, delaying rainfall has more effects on water resources and causes a great damage. Therefore, forecasting drought, especially in long term drought, is very important in agricultural water management and planning.Although reduced rainfall is very important on drought occurrence, but increasing or decreasing temperature and evapotranspiration can result in a sever or moderate drought situation. Several indices have been developed based on rainfall and evapotranspiration for drought analysis, however different indices significantly have different results. Because, drought is a multivariate phenomenon and along with the precipitation, the evapotranspiration factor also must be included in drought analysis, especially in the arid and semi-arid regions. The sequence theory, stochastic models, and conceptual models are most popular approaches in forecasting drought events. However, deterministic prediction of drought situation has been more noticeable for researchers in the recent years.Data-driven methods could aspire to extract relations that may further informand augment the current physical understanding. Support Vector Machine (SVM) is one of the data-driven algorithms which has been successfully applied in classification, regression and forecasting in the field of hydrology .It has been emphasized the need of combining an understanding of physics with data mining, not only to avoid generating misleading insight but also to produce new results. There are a number of studies using SVM in drought forecasting. The advantage of SVR is that it could transfer a non-linear problemto a linear problem using the kernel function, and be effective in solving a high dimension problem.

This study investigates the effect of climatic variables (precipitation, maximum and minimum temperature, evapotranspiration) and large-scale climatic signals arising from temperature fluctuations in the Pacific and Atlantic levels (SST) on the SPEI variable. The present study has been conducted in different climatic types of Iran. Accordingly, using the extended De Martone classification climate method, climatic types were identified and sample synoptic stations were selected from 8 climatic types. We applied SVR machine learning algorithm for prediction the SPEI, based on several meteorological predictors during the 1966–2014. In order to learn the SVR model we considered 80% of dataset for training and 20% of the rest as the testing dataset. We also applied the Lasso regression approach to select the important variables affecting the SPEI. In this regard, three different kernel functions were used for training the support vector regression (SVR) including linear, radial and polynomial kernels and the results were evaluated using different criterions.

Lasso regression results showed that in all climates, the most effective drought variable is precipitation. Among the teleconnection signals in all climates in Iran (except semi-arid cold and extra-arid warm), the effect of SST in the North Atlantic and South Pacific is effective on agricultural drought. The results showed that in all climates, the radial function showed the best results (except in the semi-arid cold, extra-arid climates) where the linear function had the best performance. The agreement index between the predictive and predictable values in the best model is 0.912 to 0.731, which is very humid and arid-warm climate, respectively. Therefore, the highest accuracy is related to Ramsar station and the lowest accuracy is related to Bushehr station. To evaluate the performance of the SVR model in predicting drought in training phase, the variables of drought including severity, duration and peak and magnitude were examined. The model is generally underestimated in estimating the severity, magnitude, and peak of drought (except at Tabas and Urmia stations, where the linear function showed different behaviors in the extra-arid warm and semi-arid cold climates). The largest difference between the observed and predicted values of drought severity was observed in the 75th percentile (Bushehr station).

Finally, it can be said that SVR is very efficient in predicting SPEI for most climates. However, its performance in different climates seems to be different, which may indicate the different role of regressions used in training SVM model and different variables. Finally, the results show that SVR is an attractive machine learning tool for drought prediction and can provide useful information for water management in agriculture.