Prediction of Soil Salinity/Sodicity and Salt-Affected Soil Classes from Soluble Salt Ions Using Machine Learning Algorithms

Abstract

Salt-affected soils are related to salinity (high content of soluble salts) and/or sodicity (excess of sodium), which are major leading causes of agricultural land degradation. This study aimed to evaluate the performances of three machine learning (ML) algorithms in predicting the soil exchangeable sodium percentage (ESP), electrical conductivity (ECe), and salt-affected soil classes, from soluble salt ions. The assessed ML models were Partial Least-Squares (PLS), Support Vector Machines (SVM), and Random Forests (RF). Soil samples were collected from the High Valley of Cochabamba (Bolivia). The explanatory variables were the major soluble ions (Na+, K+, Ca2+, Mg2+, HCO3−, Cl−, CO32−, SO42−). The variables to be explained comprised soil ECe and ESP, and a categorical variable classified through the US Salinity Lab criteria. According to the model validation, the SVM and RF regressions performed the best for estimating the soil ECe, as well as the RF model for the soil ESP. The RF algorithm was superior for predicting the salt-affected soil categories. Soluble Na+ was the most relevant variable for all the predictions, followed by Ca2+, Mg2+, Cl−, and HCO3−. The RF and, alternatively, the SVM models can be used to predict soil ECe and ESP, as well as the salt-affected soil classes, from soluble ions. Additional explanatory features and soil samples might improve the ML models’ performance. The obtained models may contribute to the monitoring and management of salt-affected soils in the study area.

Effects of Different Organic Materials on the Remediation and Improvement of Secondary Salinized Greenhouse Soil

“…The results of the study indicated that the RF model outperformed the MLR model in terms of accuracy when applied to both the training and validation datasets, consistent with Wang et al [57] This can be attributed to the effectiveness of soluble major ions complex as a predictor for soil salinity. Additionally, the performance of the MLR model was found to be more suitable in cases of multicollinearity, which was observed due to the high correlation among soluble salt ions [58]. However, both two models performed normal when predicting the soil ESP.…”