Modeling the activity ratio of soil potassium using machine learning approach

The potassium (K) Quantity-Intensity (Q-I) relationship results in important parameters, including the activity ratio of potassium at equilibrium (ARe^K), which indicates potassium availability in soil. Experiments to observe soil Q-I K relationship parameters are often complex, time-consuming, and do not include environmental variables. This research aims to model ARe^K using a machine learning (ML) approach. ML models applied are Random Forest (RF), Cubist, and Support Vector Machine (SVM) as the primary approaches, with Multiple Linear Regression (MLR) serving as a baseline. The dataset was derived from sixtyone observation points in Brebes, Central Java. The predictors were pH, organic carbon, clay, cation exchange capacity (CEC), exchangeable cations (Exc-Ca, Mg, K, Na), water soluble K, available K, K saturation, potential K, non-exchangeable K (NE-K), elevation, and slope. The response variable was the ARe^K. Variable selection was performed using Pearson correlation to eliminate highly correlated predictors and reduce multicollinearity. Exactly 75% of the data was utilized as the training set and 25% as the test set. Three metrics, i.e., MAE, RMSE, and R², were used in model evaluation. The results showed that the Cubist model could predict ARe^K with high accuracy (R² =0.9437) and low RMSE (0.5701) and MAE (0.3514). Based on the Cubist model, Exc-K, Exc-Mg, CEC, and Exc-Ca were the most important variables for predicting ARe^K. This model can be employed to support site-specific fertilizer recommendation strategies. To improve the performance of the model, it is necessary to add other predictor variables (e.g., soil physical properties, clay minerals, rainfall, temperature and soil moisture).