[Prediction of Spatial Distribution of Cation Exchange in Agricultural Soils Based on Machine Learning]

Cation exchange capacity （CEC） reflects the ability of soil to sequester exchangeable cations and is an important indicator of the fertility and environmental quality of agricultural soils. The indoor titration method for determining soil cation exchange is expensive and cumbersome. To this end, 565 soil samples from the 0-20 cm plough layer were collected from farmland in Ningxia, and the parameters of soil pH, organic carbon, and mechanical composition were determined. A field-scale soil cation exchange （CEC） estimation model was constructed using multiple linear regression and machine learning methods to obtain soil CEC values rapidly and accurately. The results showed that： ① The mean CEC value of farmland soils in Ningxia was 9.39 cmol·kg^-1, with a coefficient of variation of 40.74%. This indicated a high degree of variability, with the spatial distribution of the CEC values generally showing higher values in the periphery of the Yellow River Basin （Ningxia section） and the southern mountainous areas of Ningxia and lower values in the central arid zone and the east-central region. ② The soil parameters selected for modeling the total dataset were as follows： Soil organic carbon, clay content, pH, and sand content were the important factors influencing the CEC of farmland soil in Ningxia, with correlation coefficients of 0.55, 0.72, -0.41, and -0.44, respectively. ③ The results of multiple linear regression modeling showed that dividing the total dataset according to the urban area and constructing a multiple linear-type regression model within the urban area was more conducive to the prediction of the CEC of farmland soils. ④ Compared with the multiple linear regression method, the machine learning method was more effective in the prediction of the total dataset. Further, using the multiple linear regression model as a reference, the prediction accuracy （R²） of the back propagation neural network, convolutional neural network, back propagation neural network optimized by the particle swarm algorithm, convolutional neural network optimized by the particle swarm algorithm, back propagation neural network optimized by the grey wolf algorithm, and convolutional neural network model optimized by the grey wolf algorithm were improved by 13.59%, 30.78%, 18.91%, 35.47%, 20.94%, and 38.91%, respectively. ⑤ The validation results showed that the validation set of the convolutional neural network model optimized by the grey wolf algorithm had an R² of 0.91, an RMSE of 1.07 cmol·kg^-1, and an NRMSE of 11.77%, and the model was close to the very stable level with the best overall performance. In conclusion, the convolutional neural network model optimized by the grey wolf algorithm has high prediction accuracy and strong extrapolation ability, which is a better model for predicting soil CEC at the farmland scale. This result provides a novel idea and solution for the prediction of soil CEC in farmland in Ningxia and the whole country.