Evaluation of effective parameters for predicting the potassium grade of saline water by using support vector machine and random forest algorithms (case study: playa of Khoor and Biabank area city, Isfahan province)

Introduction:

With the increase in the world population, one of the important issues in the field of agriculture is increasing the production of agricultural products, and potassium is one of the most widely used elements to increase crop yield. For this reason, the demand for potassium fertilizers increases. One of the main sources of potassium fertilizers is underground water one of the important issues in saline water extraction is the amount of potassium grade of saline water. conventional methods of grade estimation, such as geometric and geostatistics techniques, cannot accurately estimate the grade value and have low accuracy. one of the novel solutions to estimate the grade of minerals is Machine learning algorithm, which perform evaluation and determination of the grade of mineral resources with high accuracy.

Objective:

The aim of this research is to Evaluation of effective parameters for predicting the potassium grade of saline water by using of machine learning algorithms (random forest and support vector machine) as new, low cost and cost effective methods and determining the effective parameters (independent variables used ) with the greatest influence measuring the potassium grade in order to improve the utilization of potassium reserves and reduce executive, operational and laboratory costs.

Materials and method:

In this research is to use support vector machine (SVM) and random forest (RF) algorithms in order to predict and prioritize the effective parameters on the potassium grade of groundwater in playa Khoor and Biabank in Isfahan province. For this purpose, 55 different parameters were measured in 12 boreholes (sampling locations).The parameters measured as independent variables include the percentage of saturation moisture core at 15 different depths (sp1 sp15), the apparent specific gravity of the core at 15 different depths (pb1 pb15), the porosity of the core at 15 different depths (n1 n15), polygon area (ap), underground water depth (duw), salt layer depth (dsl), surface layer potassium (slp), brine density (d) and the amount of calcium (Ca), magnesium (Mg), sodium (Na), chlorine ( Cl) and the dependent variable were also the potassium grade in the brine (Potassium Grade). three parameters n, sp and pb which were measured in 15 different depths; They were converted into an equivalent parameter using the principal component analysis (PCA) method. Also, three measured parameters, Ca, Mg, and Na were entered into the model with the sodium absorption ratio (SAR) formula. A total of 10 measured parameters were entered into the model as independent variables to predict the grade of potassium. Both RF and SVM models were implemented in Python programming language based on the relationship between dependent variable and independent variables. In different kernels of the SVM algorithm, in order to prevent the collinearity of independent parameters, all the different combinations of independent variables ( 2 to the power of 10 different combinations) considering the variance inflation factor (VIF) less than 8 and the highest coefficient of determination and the lowest MSE error are checked and the best combination were chosen. Permutation Feature Importance (PFI) and Recursive Feature Elimination (RFE) methods were used in the RF model to prioritize and select parameters for modeling.

Results and discussion:

The parameters effective in predicting the potassium grade of the both in the RF algorithm and the linear function of the SVM algorithm were sp, ap, duw, slp, SAR and n, sp, duw, and SAR respectively, which led to the best results (high determination coefficient and low error). Based on the results, the accuracy of the model (explanation coefficient) for the RF model and SVM (linear function) was 0.99 and 0.97, respectively, which indicates the good accuracy of both algorithms. Effective parameters in choosing suitable areas for drilling in order to extract potassium from saline water play a significant role and prevent repeated and time consuming tests in the laboratory, and the developed models can be used for this purpose

Conclusion:

Machine learning algorithms are one of the most important techniques for evaluating mineral grade estimation . Given that, a large part of the country consists of arid and semi arid areas, where there are many playas that are rich in underground saline water that have good and suitable reserves of potassium and because in playa, the conditions are unpredictable and the environment has high complexity، Effective parameters in choosing suitable areas for drilling in order to extract potassium from saline water play a significant role.