Comparative evaluation of machine learning models for cation exchange capacity estimation

Cation exchange capacity (CEC) is a crucial indicator of soil fertility and a key parameter for soil databases. However, traditional laboratory methods for CEC determination are time-consuming, labor-intensive, and expensive, limiting their inclusion in many studies. This study aims to employ machine learning models to estimate CEC using low-cost soil properties. The study utilized three machine learning models: Multilayer Perceptron Artificial Neural Network (MLP-ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Random Forest (RF) regression. Soil samples were collected from 232 points at 0-20 cm depth in the Yüksekova plain, representing diverse parent materials and land uses. The CEC values, clay, silt, sand, organic matter (OM) and lime contents, aggregate stability (AS), electrical conductivity (EC), and pH of soil samples were determined. The CEC values, which vary depending on parent material and land use as well as soil properties, ranged from 8,96 meq 100 g-1 to 76,25 meq 100 g-1 with an average of 27,85 meq 100 g-1. Considering the correlation between soil CEC values and soil properties, clay, sand, AS, OM, lime, and EC were used as variables for CEC prediction. The performance of the models was evaluated using three metrics: root mean square error (RMSE), mean absolute error (MAE), and regression coefficient (R2). All three models demonstrated promising CEC prediction performance. ANFIS emerged as the superior model with an R2 value of 0.86, MAE of 1,36, and RMSE of 4,82. MLP-ANN and RF followed with R2 values of 0,77 and 0,75, MAE values of 5,47 and 4,41, and RMSE values of 5,96 and 5,89, respectively. The findings of the study demonstrate that the CEC values of the soils in the study area can be successfully predicted using all three machine learning models used in the study.