Improving crop modeling in saline soils by predicting root length density dynamics with machine learning algorithms

Crop modeling is an effective tool for simulating crop growth under various agricultural water and salinity management practices. However, most crop models fail to describe the root dynamics in response to soil stresses adequately. To address this issue, field experiments were conducted by planting sunflowers in saline soils. Three machine learning (ML) models of random forest (RF), gaussian process regression (GPR), and extreme gradient boosting (XGBoost) were initially introduced for predicting root length density (RLD). Then, by coupling with a crop model SWAP, the soil salt content (SSC), soil water content (SWC), and crop growth indicators of leaf area index (LAI) and dry matter (DM) were simulated. Results show that RF and XGBoost models could predict RLD more accurately than the GPR model, with root mean square error (RMSE) lower than 0.473 cm cm-3. Compared to using a typical cubic polynomial function (CPF) of RLD in the SWAP model, similar SWC and SSC simulation results were obtained based on the ML models. However, for the crop growth simulation, the performances of ML models were significantly better than the CPF. Especially for LAI simulation in the high salinity fields, the relative root mean square error (RRMSE) in the RF model was 0.222–0.282 lower than in the CPF. Moreover, compared to the XGBoost model of RLD, more accurate and stable simulation results of SWC, SSC, and LAI were obtained based on the RF model. These results illustrate that ML models, especially the RF model, can be used to quantify RLD dynamics and improve crop modeling performances.