Prediction of soil salinity parameters using machine learning models in an arid region of northwest China

Accurate estimation of soil ions composition is of great significance for preventing soil salinization and guiding crop irrigation. The traditional laboratory measurement of ions composition is accurate for calculating soil salinity parameters, but its application is often limited by the high cost and difficulty in long-term in-situ measurement. This study evaluated the performances of three machine learning models, i.e., random forest (RF), support vector machine (SVM) and extreme gradient boosting (XGB), in predicting total dissolved ionic matter (TDI), potential salinity (PS), sodium adsorption ratio (SAR), exchangeable sodium percentage (ESP), residual sodium carbonate (RSC) and magnesium adsorption ratio (MAR) in soils. Soil temperature (T), potential hydrogen (pH), soil water content (SWC) and electrical conductivity (EC) were used as model input variables. Data from 467 soil samples in the Shihezi region of northwest China were used for model training–testing and validation. The results showed that the XGB model performed better when EC, SWC and T were used as input variables, while the RF and SVM models performed well when EC, T and pH were used. The XGB model had overall better performance than the SVM and RF models (with decreases in RMSE by 24.2%–54.8%), while the RF and XGB models showed better generalization capability than the SVM model. The XGB model with EC, SWC and T as input variables could be used to predict all the soil ions composition with coefficient of determination (R2) > 0.770 and residual prediction deviation (RPD) > 1.98, while the RF and SVM models with EC, SWC and pH as input variables could be used to predict TDI (R2 > 0.957, root mean square error (RMSE) < 1.284 g kg−1, RPD > 4.83), PS (R2 > 0.772, RMSE < 0.511 mol L−1, RPD > 2.1) and ESP (R2 > 0.67, RMSE < 9.249%, RPD > 1.74), and the RF model with EC, SWC and pH as input variables could be used to predict RSC (R2 > 0.609, RMSE < 1.060 mol L−1, RPD > 1.60). This study overcame the difficulty of traditional methods in predicting soil salinity parameters, evaluated the performances of different machine learning models, and optimized the input variable combinations. This study can help farmers in regions affected by soil salinization better manage planting practices and improve land sustainability.