China’s drylands are primarily located in the northwestern inland areas and the northern part of the Qinghai-Tibet Plateau, covering provinces and regions such as Xinjiang, Inner Mongolia, Ningxia, Gansu, and Qinghai (Figure 1). Figure 1 depicts the spatial extent of China’s drylands, highlighting the distribution of major basins (e.g., Tarim, Junggar, Qaidam), deserts, and the division into temperate and high-altitude arid zones. The total area spans approximately 6.6 million km², accounting for 69% of the country’s total land area. This makes it the largest and most complex arid region in the mid-latitude zone globally (Li et al., 2021). According to the United Nations Environment Programme (UNEP) Aridity Index (AI < 0.65), China’s drylands can be divided into temperate arid zones (such as the Tarim Basin and Junggar Basin) and high-altitude arid zones (such as the Qaidam Basin and northern Tibet Plateau). Annual precipitation ranges from 400 mm in the eastern semiarid zones to less than 50 mm in the western hyperarid zones, while evaporation rates reach as high as 2,000–3,000 mm. The predominant ecosystem types are deserts, desert grasslands, and mountain-oasis composite systems, with desert areas accounting for 45.3% of the total drylands, including mobile sand dunes, deserts, and saline-alkali lands. This region is a special adaptation zone for biodiversity, featuring arid plant communities such as Populus euphratica and Haloxylon ammodendron, as well as endemic species like the wild camel (Camelus ferus) and the Gazella subgutturosa.

We used the following formula to calculate vegetation or soil TN and TP stock (Liu et al., 2013): T N  or  T P   s t o r a g e = ∑ i = 1 n D e n s i × A r e a × 10 − 12 i is the ith grid square; Dens _i is TN or TP density (g m⁻²) for the ith grid square calculated, and Area is the area (m²) of each grid square, set by the defined resolution. These calculations were performed using the GIS software package Arcmap Desktop (version 10.2) with the spatial analyst module.

To obtain spatially explicit N and P stock in drylands ecosystems, this study used machine learning models to convert vegetation and soil N and P density from plot scale to regional scale (Figure 2). Machine learning models are powerful tools for predicting regional-scale soil and ecological properties (Beer et al., 2010); compared to scale conversion methods based on classification units, machine learning techniques can reduce the uncertainty in soil carbon and nitrogen stock estimates (Zhang et al., 2025). First, using grid value interpolation sampling. To clarify, the sampling was structured as a grid, and the values were not simply extracted but spatially interpolated. The methodology is now stated as: “The values for the 4,200 sampling sites, which were positioned at grid cell centroids to ensure uniform spatial coverage, were generated by applying Ordinary Kriging interpolation to the original dataset. This approach optimally estimates values at these points by leveraging the spatial autocorrelation inherent in the source data. Outliers were removed to obtain data on vegetation and soil (0–30 cm) N and P density and environmental covariates from 4,200 sites. The method has been clarified in the manuscript as follows: “Outliers were identified and removed based on the interquartile range (IQR) method, which is the statistical criterion underlying the boxplot approach. Specifically, data points falling below Q1 - 1.5 × IQR or above Q3 + 1.5 × IQR were considered outliers and excluded from subsequent analysis. Four machine learning models—support vector machine (SVM), random forest (RF), Gaussian process regression (GPR), and gradient boosted regression tree (GBRT). The predictive variables for the machine learning models included longitude, latitude, Ele, slope, MAT, MAP, NDVI, NPP, Sand, Clay, pH, and SWC. The selection of these 12 variables as predictive variables for the models was primarily based on the following two considerations: (1) the selected variables comprehensively reflect the influence of geographical location, climate, vegetation, and soil physicochemical property on N and P stock in vegetation and soil; (2) previous studies have shown that the selected variables can regulate the magnitude of soil N and P stock by altering N and P input and output processes (Kou et al., 2019; Zhang et al., 2021). Using “leave-one-out” cross-validation (Zhang et al., 2025), this study evaluated the predictive performance of four machine learning models and presented the validation results using 1:1 line scatter plots (Zhang et al., 2025), selecting the model with the best predictive performance for further analysis. Based on the spatially distributed data of vegetation and soil N and P density predicted by the optimal model, the average values of vegetation and soil N density was calculated for each pixel. Then, using the average values of vegetation and soil N and P density and the pixel area, the N and P stock of terrestrial ecosystems in drylands were estimated.

The normality of the data was tested using the Kolmogorov–Smirnov test. Non-normal variables were log-transformed prior to the following analyses (Tian et al., 2022). Subsequently, standard statistical functions were employed to compute the mean, minimum (Min), maximum (Max), standard deviation (SD), and coefficient of variation (CV) for each respective metric (Table 1).

To assess the relative importance of spatial, climatic, soil physicochemical properties, and vegetation on N and P, we used the “Vegan” package to perform variance decomposition analysis. We used the “forward.sel” function of the “Packfor” package to avoid redundancy and multicollinearity in variation partitioning analysis. The piecewise structural equation modelling (piecewiseSEM) was constructed to analyze the direct and indirect pathways through which environmental factors influence N and P. The model accounted for sampling sites (as random effects, reflecting the nested structure of samples within research centers) and reported marginal R ² (the variance explained by predictors when random effects are ignored) and conditional R ² (the total variance explained when random effects are included). Fisher’s C test was used to assess the overall model fit, and necessary model adjustments were made based on its significance (P < 0.05) and goodness-of-fit criteria (0 ≤ Fisher’s C/df ≤ 2 and 0.05 < P ≤ 1.00). To determine the optimal predictors for N or P, piecewiseSEMs were constructed across different data scales, incorporating similar core predictors. Finally, the total standardized effects of each predictor were calculated. All statistical analyses were conducted in R v4.1.0, with piecewiseSEMs implemented using the “nlme” and “lme4” packages.

The 1:1 linear scatter plots (Figures 3, 4) shows that, compared to the other four machine learning models, RF has the highest accuracy in predicting STN, STP, VTN, and VTP (0.89, 0.92, 0.95, and 0.94); however, it has the lowest MAE (3.21, 0.56, 5.37, and 2.23) and RMSE (5.09, 0.79, 7.05, and 2.98). This indicates that the RF model demonstrates the best performance in predicting N and P in arid land ecosystems.

This study comprehensively estimated the N and P density, stock, and coefficient of variation (CV) in drylands (Table 2). The mean density of STN, STP, VTN, and VTP were 168.4, 76.5, 2.7, and 0.4 gm⁻², respectively; the stock was 1111.7, 504.9, 17.6, and 1.7 Tg, respectively; and the CV was 1.25, 0.87, 2.33, and 2.05, respectively.

Overall, both N and P stock in soil and vegetation exhibit great spatial heterogeneity in their distribution, and their distribution patterns share both similarities and differences (Figure 5). Similarities are evident in the following: N and P stock in soil and vegetation exhibit high values in the northeastern part of the drylands (the Greater Khingan Range and the Northeast Plain), while they are low in the central part of the drylands (the Inner Mongolia Plateau). Differences are manifested as follows: N stock is relatively high in the northwestern part of the Qinghai-Tibet Plateau, while P stock in the soil of the Tarim Basin are relatively prominent.

We used the piecewiseSEMs to separately investigate the direct and indirect effects of spatial, climatic, vegetation, and soil environments on STN, STP, VTN, and VTP (Figure 6). Overall, in the model, all environmental factors explained 35%, 27%, 18%, and 15% of the variance in STN, STP, VTN, and VTP, respectively. When considering the ‘random effects’ of sampling locations, an additional 20% of the variance in N and P across the drylands could be explained. In the models, the primary drivers of STN and STP were soil environment factors SWC and clay, respectively; while the primary drivers of VTN and VTP variations were MAP.

Through variance partitioning analysis (Figure 7), we found that space, climate, soil, and plants each explained 42.2%, 37.6%, 33.9%, and 28.2% of the total variations in STN, STP, VTN, and VTP, respectively. Among these, soil had the greatest independent effect on STN and STP variation, explaining 24.4% and 18.4% of the variation, respectively; climate had the greatest independent effect on VTN and VTP variation, explaining 14.9% and 9.8% of the variation, respectively.

This study demonstrates that the random forest (RF) model achieves the best predictive performance in estimating the density of nitrogen (N) and phosphorus (P) in dryland terrestrial ecosystems, as clearly shown in Figures 3, 4. RF, an ensemble learning algorithm based on multiple decision trees proposed by Breiman (2001), offers high accuracy and strong interpretability. The superior predictive performance of RF is reflected in the following aspects:

(1) Effective prevention of overfitting while maintaining high accuracy: Each decision tree is trained using bootstrapped sample subsets and random feature selection. This dual randomness reduces variance among individual trees and the overall model, effectively avoiding overfitting—a common issue in single decision trees—while preserving high predictive accuracy (Zhang et al., 2025). (2) Excellent scalability and capacity to handle complex data: RF is particularly suitable for processing large-scale and high-dimensional datasets. Its parallel computing capability enables efficient and scalable model training even with massive ecological data (Wadoux et al., 2020). (3) Robust prediction results: By aggregating outputs from multiple trees through voting or averaging, RF substantially reduces prediction variance and enhances stability. This “collective intelligence” framework yields more reliable predictions.

The superior performance of RF in predicting N and P in drylands stems from its capability to capture complex nonlinear relationships and interactions among environmental variables, its robustness against outliers and noise in field observations, and its utility in providing ecologically meaningful insights through feature importance analysis. Therefore, the RF algorithm should be more widely applied and further developed in future soil attribute mapping studies.

The spatial distribution of N and P stocks in China’s dryland terrestrial ecosystems exhibits pronounced heterogeneity, with significantly higher vegetation and soil N and P stocks in the northeastern regions compared to the northwest (Figure 5; Table 1). This pattern can be attributed to several factors: Climate differences: MAP is substantially higher in northeastern drylands, promoting vegetation productivity (e.g., aboveground biomass) and microbial mineralization rates, thereby enhancing inputs of organic N and P.

Soil type differences: Northeastern soils, such as black calcareous and chestnut calcareous soils, are rich in organic matter, and their humus-clay complexes help stabilize P via chelation. In contrast, widespread saline-alkali soils (pH > 8.5) in the northwest facilitate P fixation by calcium and magnesium minerals, and frequent wind erosion exacerbates gaseous N loss (Zhang et al., 2025). Agricultural management differences: Intensive farming systems in the northeast involve higher fertilizer inputs than those in the northwest. These exogenous nutrients are returned to the soil via crop residues and stabilized by soil organic matter, leading to continuous accumulation of N and P stocks in terrestrial ecosystems (Bi et al., 2023).

In dryland ecosystems, soil properties—specifically soil water content (SWC) and clay content—are the dominant factors controlling soil total nitrogen (STN) and soil total phosphorus (STP), respectively (Figures 6, 7).

SWC is widely recognized as a limiting factor in dryland ecosystem processes, strongly regulating plant growth and soil microbial activity, and thereby influencing the accumulation and cycling of N and P (Moyano et al., 2013). It also shapes microbial community structure and regulates key nitrogen transformation processes such as nitrification and denitrification (Chen et al., 2013), further affecting nutrient dynamics, gas exchange, and microbial growth (Mcdaniel et al., 2013). For instance, elevated soil moisture significantly enhances microbial activity—increasing N mineralization rates by 30–50%—and directly enlarges the soil N pool by stimulating organic matter decomposition and biological N fixation (Schimel and Bennett, 2004). Concurrently, improved moisture conditions alleviate plant water stress, increase aboveground productivity by 20%–40%, and enhance litter inputs, indirectly promoting the accumulation of organic N (Austin et al., 2004). In summary, SWC indirectly regulates the stocks of soil N and P by modulating vegetation growth and litter decomposition processes (Zhang et al., 2018; Wang et al., 2020).

Clay content exerts a significant positive effect on STP in drylands. As the proportion of soil aggregates larger than 1 mm increases, P stocks also rise considerably in dryland red soils (Xu et al., 2020), a finding consistent with the present study. Clay particles possess a high specific surface area and strong adsorption capacity, enabling them to immobilize P in soil solution via specific adsorption, thereby converting soluble P into adsorbed forms and reducing plant-available P (Bicharanloo et al., 2022). In clay-rich soils, applied P fertilizers are prone to fixation, lowering P availability. This is mainly due to the high surface area and cation exchange capacity of clay minerals (e.g., montmorillonite and illite), which effectively retain phosphate ions and mitigate leaching (Zhao et al., 2023). Under drought conditions, clay further sequesters soluble P by forming 2:1 layered silicate structures, thereby stabilizing P within soil aggregates (Chi et al., 2024).

Climate, particularly mean annual precipitation (MAP), serves as the principal driver of vegetation total nitrogen (VTN) and vegetation total phosphorus (VTP) in drylands (Figures 6, 7). Precipitation is a major source of soil moisture in these regions, and sufficient rainfall improves soil conditions, supporting the survival and reproduction of plants and microbes. Increased precipitation directly alleviates plant water stress, enhances photosynthetic rates and biomass accumulation, and facilitates root uptake of soil N. For example, in some grassland ecosystems, a 30% increase in precipitation significantly promotes plant root N acquisition (Li et al., 2022). Moreover, precipitation stimulates soil microbial activity, accelerates the mineralization of N and P, and promotes the release of available N and P from organic matter, thereby increasing plant N and P content (Li J. W. et al., 2023). Precipitation also influences the forms and availability of soil P (Yang et al., 2025). Under low rainfall conditions, P tends to form insoluble phosphates, reducing its availability to plants. Increased precipitation can enhance P solubility and release, improving its bioavailability and uptake (Li et al., 2022). In summary, elevated precipitation boosts microbial activity, increases the pool of plant-available N and P, facilitates the movement of soluble nutrients to the root zone, and reduces topsoil nutrient loss via evaporation, collectively leading to higher VTN and VTP.

This study mapped the spatial distribution patterns of N and P in dryland ecosystems, quantitatively estimated the N and P stock in vegetation and soil, and analyzed the main factors driving their changes; however, the results still exhibit a certain degree of uncertainty. First, a large proportion (>30%) of the spatial variation in N and P stock in drylands cannot be fully explained by existing models. Secondly, microbial activity, which is a key factor regulating N and P cycles (Nelson et al., 2016; Li J. S. et al., 2023), could not be included in the analysis of driving mechanisms due to the lack of basic data. Additionally, human activities (such as the expansion of irrigated agriculture) and land-use changes have significantly disrupted soil N and P stock (e.g., land reclamation causes a 20%–30% loss of organic P), which have not been quantified. These missing elements may increase the uncertainty of the results.