Estimation of soil cation exchange capacity using Genetic Expression Programming (GEP) and Multivariate Adaptive Regression Splines (MARS)

- The soil CEC was estimated by using the Genetic Expression Programming (GEP) approach.
- We employed Multivariate Adaptive Regression Splines (MARS) to estimate the soil CEC.
- Readily measurable soil properties were used as inputs in GEP and MARS.
- Both GEP and MARS estimate CEC more accurately than the traditional regression models.
- The soil organic matter and pH have the most and the least effect on CEC, respectively.

SummaryThe soil cation exchange capacity (CEC) is one of the main soil chemical properties, which is required in various fields such as environmental and agricultural engineering as well as soil science. In situ measurement of CEC is time consuming and costly. Hence, numerous studies have used traditional regression-based techniques to estimate CEC from more easily measurable soil parameters (e.g., soil texture, organic matter (OM), and pH). However, these models may not be able to adequately capture the complex and highly nonlinear relationship between CEC and its influential soil variables. In this study, Genetic Expression Programming (GEP) and Multivariate Adaptive Regression Splines (MARS) were employed to estimate CEC from more readily measurable soil physical and chemical variables (e.g., OM, clay, and pH) by developing functional relations. The GEP- and MARS-based functional relations were tested at two field sites in Iran. Results showed that GEP and MARS can provide reliable estimates of CEC. Also, it was found that the MARS model (with root-mean-square-error (RMSE) of 0.318 Cmol+ kg−1 and correlation coefficient (R2) of 0.864) generated slightly better results than the GEP model (with RMSE of 0.270 Cmol+ kg−1 and R2 of 0.807). The performance of GEP and MARS models was compared with two existing approaches, namely artificial neural network (ANN) and multiple linear regression (MLR). The comparison indicated that MARS and GEP outperformed the MLP model, but they did not perform as good as ANN. Finally, a sensitivity analysis was conducted to determine the most and the least influential variables affecting CEC. It was found that OM and pH have the most and least significant effect on CEC, respectively.