Performance evaluation of models based on Data Decomposition and GRACE Satellite Products for Groundwater Level Modeling (case study: Aspas aquifer)

Background and Objectives: Excessive extraction of groundwater has caused most of Iran's groundwater aquifers to face a drop in water level in recent years. This has subject caused the use of most of the aquifers to be prohibited, most of the Qanats have dried up and most of the permanent springs have had a significant reduction in their water supply. Therefore, the investigation of the groundwater level should be given more attention. Various methods and tools have been used to investigate this issue. Artificial intelligence models have been used in most of these studies. Among these intelligence models, Support Vector Regression (SVR) model has performed well. In order to improve the performance of these models, in recent years, the use of pre-processing tools and the formation of hybrid models have been considered. One of these tools is complementary ensemble empirical mode decomposition (CEEMD). In this research, the combination of this tool with the SVR model was used to check the groundwater level in the Aspas aquifer. Then their results were compared with the results of the Gravity recovery and climate experiment (Grace) satellite.

Materials and methods: The Aspas subbasin with code 4321 is located northwest of the Tashk-Bakhtegan and Maharlu basin in Fars Province. To check the groundwater level in this sub-basin, the SVR model with 4 kernels include: polynomial kernels, RBF kernel, sigmoid kernel, and linear kernel (Lin) was used. Then discusses the formation of a hybrid model obtained from the combination of CEEMD with the SVR intelligence model. When an initial signal is decomposed using the CEEMD method, and the resulting sub-signals are used as inputs to the SVR intelligence model, the hybrid model of CEEMD-SVR is obtained. Satellite data was used to compare the performance of artificial intelligence models. For this purpose, Grace satellite products with 6 different algorithms were used. The parameters the coefficient of determination (R2), root mean square error (RMSE), and the Akaike information criterion (AIC), was used to examine the efficiency of the methods.

Results: The results showed that intelligent models had better performance than Grace satellite products. Therefore, it is more appropriate to use intelligent models, especially the CEEMD-SVR model, to predict the values of the groundwater level. One of the advantages of using satellite data is that it is available up-to-date. If the satellite data values can be approximated to the observed values (in a similar statistical period) based on a suitable method, the groundwater level data can be estimated in an up-to-date manner.

Conclusion: In this study, the SVR model was used to evaluate the groundwater level changes in the Aspas alluvial aquifer located in the Tashk-Bakhtegan-Maharlu basin. Using observation wells in the area the aquifer groundwater hydrograph was plotted. Changes in groundwater level in the aquifer were estimated using the values of precipitation, temperature, and evaporation parameters obtained from drawing different maps, and groundwater level in the aquifer. The preprocessing tool of CEEMD was used. The results showed that the use of the CEEMD has improved by 3.08% the performance of the SVR model. The GRACE satellite products are used. The comparison of the results of processing algorithms showed that the GFZ processing algorithm had the best performance with a coefficient of determination of 0.71 and an RMSE value of 39.15. In the next step, the performance of the CEEMD-SVR model was compared with the GFZ algorithm. The results showed that the CEEMD-SVR model performed better (R2=0.77, RMSE=25.90) and has the ability to be used for modeling and predicting the groundwater level in aquifers, especially the Aspas aquifer.