Root length density (RLD) is an indispensable input for driving almost all agro-hydrological models, but it is difficult to measure and has strong plasticity to soil environments, thus it is a challenge to characterize the dynamics of RLD if crops suffering adverse soil stress at field scale. Soil salinity is a major abiotic stress that restricts crop shoot growth and yield formation, but it may stimulate root growth of salt-tolerant crops. In this study, based on 256 datasets of actual root length density (ARLD) for sunflower grown under saline conditions and influencing factors as days after sowing (DAS), root depth (RD), soil salt content (SSC), soil water content (SWC), and leaf area index (LAI), we (1) clarified the limitations of cubic polynomial, exponential, and power elementary functions of root depth (RD) for ARLD prediction; and (2) established and compared three novel machine learning models (MLMs) including gaussian process regression (GPR), multivariate adaptive regression spline (MARS), and random forest (RF) with eight combinations of inputs (COIs). Results show the distribution of the sunflower’s ARLD was significantly different in fields with different salinity levels, the peak value of ARLD in the low-salt field was smaller than 1.2 cm·cm−3, but it could be over 3.0 cm·cm−3 in the high-salt field. Except at the early growth stage (DAS = 28–30), all the elementary functions failed to fit the ARLD accurately with the RMSE ranging from 0.39 to 0.97 cm·cm−3 and R2 lower than 0.38. The higher prediction accuracies for ARLD were obtained in MLMs, especially with the COI of DAS + RD + SSC + LAI. Moreover, the accuracies of RF and GPR models (RMSE ranging from 0.36 to 0.37 cm·cm−3 and R2 greater than 0.73 in the test) were higher than the MARS. The spatiotemporal distribution of simulated ARLD in the GPR model was relatively smooth, while it presented certain discontinuity in the RF model. In general, the RLD plasticity of sunflower in saline soil should not be ignored, the MLMs models, such as the GPR and RF models, are more applicable for RLD prediction than the elementary function models.