Developing novel ensemble models for predicting soil hydraulic properties in China’s arid region

•Multiple machine-learning algorithms were used to develop new PTFs for SHP prediction.•1-km resolution maps of 0–2 m SHPs were generated for China’s arid region.•The new PTFs had a higher accuracy than other commonly used PTFs.•Regional terrestrial biosphere modeling by using global SHP datasets should be endorsed with caution. Accurately quantifying the mass (water, nutrients, and carbon) and energy exchange processes between the Earth’s atmosphere, biosphere, and lithosphere requires the accurate parameterization of soil hydraulic properties (SHPs) and their spatial heterogeneity. Because direct measurements of SHPs are difficult, time-consuming, and impossible at larger spatial scales, various pedotransfer functions (PTFs) have been developed in the last few decades, providing divergent estimates of SHPs from readily measurable variables. However, existing PTFs are mostly developed for specific regions and may not be suitable for other pedoclimatic conditions. Here, PTFs were developed using multiple machine-learning algorithms to estimate SHPs and examined the weaknesses and strengths of each method in estimating the average and variability of SHPs across China’s arid region. The optimal PTFs were applied to a 1 × 1 km2 regional map of texture and bulk density, thus producing maps of the saturated hydraulic conductivity (Ks), the parameters of the van Genuchten (VG) formulation, field capacity (θfc), wilting point (θwp), plant available water (θpa), and soil macroporosity (ϕm) in the 0–2 m soil profile throughout the region. The results indicate that the ensemble model with the averaging method (EMA) is the most robust for estimating all SHPs. The EMA-PTFs for Ks yielded the best performance for sand textures, followed by sandy loam, loam, silty loam, silty clay loam, loamy sand, and clay loam textures. There were no significant differences in estimating soil water retention curve parameters among the different soil texture classes. Using the same observed data set, we demonstrated that the new EMA-PTFs outperformed those of existing PTFs, such as Rosetta and HYPRES, with RMSE values decreasing by 25–83 % depending on the SHPs. Furthermore, our SHP datasets exhibited significantly deeper soil profiles and higher accuracy than other available regional and global SHP products. The VG retention parameter α shows the greatest variation vertically throughout the 0–2 m soil profile, followed by ln(Ks), θr, θpa, θfc, θwp, ϕm, θs and n. All SHPs exhibit much