A comparison of physical-based and machine learning modeling for soil salt dynamics in crop fields

The physical-based and machine learning (ML) models are two distinctive tools for predicting soil salt content (SSC). However, few studies have compared their performances, and it is yet unknown how they differ in terms of simulation accuracy at the field scale. To address this issue, based on a field experiment with sunflowers, a physical-based model and three ML models including distributed random forest (DRF), gradient boosting machine (GBM), and deep learning (Deeplearning) were developed to predict SSC in two common scenarios (A and B). In Scenario A, the SSC was predicted using the training dataset in the in-situ field, while in Scenario B, the SSC was predicted with the training dataset from the other fields. Results show that the physical-based model remains an accurate tool to predict SSC; ML models hold a similar prediction capacity with specific algorithms and input variables. In Scenario A, with limited input variables of the initial status of SSC and related spatiotemporal information, the DRF model achieved better simulation accuracies (R2 higher by 0.05–0.37, NRMSE lower by 0–0.19) than the other two ML models. However, as more input variables were added, the simulation accuracies of the GBM model gradually improved (NRMSE decreased from 0.61 to 0.30) and eventually outperformed the DRF model. Although the variable importance was significant in the Deeplearning model, poor simulation performances were obtained in Scenario A; however, in Scenario B, the simulation accuracy of the Deeplearning model was higher than other ML models, especially for the SSC prediction at the deep soil and during the crop’s late growth period, the median of the NRMSE boxplot approaching 0.31. These simulation results indicated the potential of ML models to substitute the physical-based model for SSC prediction. Still, the preferred ML model differs depending on the prediction scenarios and input variables.