From Deficiency to Enrichment: Diverging Soil Phosphorus Trajectories Across China's Croplands (1980–2018)

ABSTRACT

Decades of fertilizer intensification have reshaped phosphorus (P) cycling across China's croplands, driving a widespread transition from P deficiency to enrichment while leaving spatially heterogeneous legacies of accumulated soil P. However, the long‐term trajectories of soil available P, their regional divergence, and the drivers regulating P accumulation or depletion remain poorly resolved at the national scale, limiting the development of sustainable and region‐specific P management strategies. Here, we integrate nearly 190,000 soil samples from harmonized national soil surveys with long‐term fertilization experiments to quantify long‐term soil available P dynamics across China at 1 km resolution and to assess the importance of key environmental and management drivers using an interpretable, spatially explicit machine‐learning framework. Soil available P increased sharply from 1980 to 2018, with an overall increase of 20.6 mg kg⁻¹, shifting from widespread deficiency to pronounced enrichment, particularly in eastern croplands. Long‐term experimental evidence demonstrates that sustained fertilizer inputs play a central role in driving soil P accumulation. Interpretable machine learning revealed spatial heterogeneity in the dominant drivers of soil available P change, with management effects strongest in Central China, and climatic and environmental influences more pronounced in the Northwest and Northeast. Sensitivity‐based future projections indicated divergent potential trajectories to 2060, with climatic changes reducing soil available P, whereas management intensification increased soil available P by approximately 13%. These findings show that long‐term soil available P dynamics across China are governed by the combined effects of fertilization legacies, soil‐environmental context, and climate, providing a national basis for region‐specific and sustainable P management.

Harmonized national soil surveys, together with evidence from long‐term fertilization experiments and a machine‐learning modelling framework, were integrated to examine nearly four decades of soil available P dynamics across China's croplands. From 1980 to 2018, China's cropland soils shifted from widespread phosphorus (P) deficiency to pronounced enrichment, with mean soil available P increasing from 7.3 to 27.9 mg kg⁻¹. These changes exhibited pronounced spatial heterogeneity and were jointly driven by climate, environment, vegetation, and management factors. Overall, the study highlights pronounced spatial divergence in long‐term soil available P dynamics across China's croplands and demonstrates that sustainable future P management will depend on strategies tailored to regional combinations of management intensity, climatic constraints, and soil environmental conditions.

1. Introduction

Phosphorus (P) is indispensable for sustaining soil fertility and global food production (Sattari et al. 2016; Zou et al. 2022; McDowell et al. 2024; Ringeval et al. 2025). However, the rapid intensification of agriculture worldwide has raised growing concerns about the stability of P biogeochemical cycling, as increasing fertilizer inputs have disrupted the balance between soil P accumulation, crop uptake and environmental losses (Liu et al. 2016; Margenot et al. 2024). Soil available P serves as a central indicator of P cycling because it governs plant uptake, influences fertilizer use efficiency and reflects the combined action of biological, chemical and hydrological processes (Helfenstein et al. 2024; McDowell et al. 2024). These processes are strongly influenced by climate variability, intrinsic soil properties and management intensity, rendering soil available P highly sensitive to both environmental fluctuations and human‐driven perturbations (Leinweber et al. 2018; Wang et al. 2022). Understanding long‐term soil available P dynamics is therefore essential for revealing how croplands respond to sustained fertilization, climate variability, and land‐use change, and whether they accumulate P, approach equilibrium, or undergo depletion over time. These trajectories have far‐reaching implications for soil fertility, environmental quality and the resilience of agricultural production systems.

China exemplifies the profound challenges of sustaining long‐term soil P status amid rapid agricultural intensification and structural transformation (Alewell et al. 2020; Song et al. 2024; Gong et al. 2025). Decades of rising fertilizer applications, transitions in cropping systems, and the widespread deployment of high‐yield cultivars have cumulatively transformed nutrient balances across major crop‐producing regions (Jiao et al. 2016; Tao et al. 2022; Wang et al. 2024). Intensively managed grain belts across North China, Central China, and Southeast China have experienced substantial P accumulation, reflecting a legacy of sustained fertilizer surpluses under high‐input management (Ma et al. 2018; Gong et al. 2023). In contrast, croplands in Northwest China, characterized by coarse‐textured soils, low fertilizer inputs, and climatically constrained biological activity, continue to exhibit widespread P deficiency (Ma et al. 2018; Li et al. 2026; Zhang et al. 2019; Deng et al. 2024). These contrasting trajectories underscore how climate, soil buffering capacity, topography, and cropping intensity jointly regulate soil P availability, revealing that P dynamics across China are governed by region‐specific processes rather than a uniform national response. Regional studies and long‐term experiments indicate that fertilizer surpluses can drive available P accumulation and legacy P formation, but the extent of this buildup is strongly mediated by soil properties and climate (Gérard 2016; Dai et al. 2020; Shi et al. 2023; Wang et al. 2023; Hong et al. 2025). However, their limited spatial coverage and fragmented geographical representation constrain the extent to which these insights can be extrapolated to China's highly heterogeneous agroecosystems (Tong et al. 2017; Liao et al. 2023).

Recent advances in data‐driven modelling have substantially improved the capacity to characterize soil nutrient dynamics, with machine‐learning approaches increasingly used to integrate diverse environmental and management datasets and generate high‐resolution maps of soil properties (Padarian et al. 2020; Huang et al. 2024; Aramburu‐Merlos et al. 2024). At the national scale, such approaches have already been used to assess soil P dynamics in China. For example, Song et al. (2024) analyzed changes in topsoil available P and total P across China's forests, grasslands, paddy fields, and upland croplands during the 1980s–2010s using repeated soil measurements and machine‐learning techniques, whereas Chen et al. (2023) used data from 91 long‐term experimental sites across 1988–2018 to quantify the transformation from accumulated soil P to available P and its main drivers across Chinese cropping systems. However, these studies mainly quantified nationwide changes in topsoil P status across multiple land‐use types or site‐based relationships between accumulated and available P, and thus provide more limited insight into spatially explicit changes in cropland soil available P over time, regional differences in driver contributions, and future responses to climate and management changes. Addressing these limitations is essential for understanding not only where soil available P has changed, but also why these trajectories differ among regions and where management interventions are most likely to alter future outcomes.

Here, we integrated three harmonized national soil surveys from 1980, 2012, and 2018, comprising 1,693, 35,883, and 151,402 soil samples across China, together with evidence from 23 long‐term fertilization experiments and a machine‐learning modelling framework to examine nearly four decades of soil available P dynamics in China's croplands. We quantified national and regional trends, characterized transitions among soil available P classes, and identified zones of accelerated accumulation. We then used a spatially explicit random forest model with SHapley Additive exPlanations (SHAP)‐based attribution to assess the relative importance and spatial dominance of climatic, environmental, vegetation, and management drivers of annual soil available P change, and to evaluate regional differences in the sensitivity to climatic and management changes. Finally, we projected future soil available P trajectories under contrasting climate and management scenarios to clarify the relative influence of management intensification and climatic constraints on long‐term soil P sustainability. These findings provide a basis for developing sustainable and region‐specific P management strategies under ongoing agricultural intensification and climate change.

2. Materials and Methods

2.1. National Soil Available P Datasets and Long‐Term Fertilization Experiments

We compiled cropland soil available P data for the years 1980, 2012, and 2018 from three major national monitoring programs: the Second Soil Survey, the National Scientific Fertilizer Network, and a national campaign by the Cultivated Land Quality Monitoring and Protection Center, Ministry of Agriculture and Rural Affairs, PRC. These datasets comprised 1,693, 35,883, and 151,402 soil samples, respectively, all collected from the 0–20 cm soil layer, with georeferenced sampling locations shown in Figure S1. Soil available P (Olsen‐P) concentrations were determined by extraction with 0.5 M NaHCO₃ (pH 8.5) followed by colorimetric analysis using the molybdate–ascorbic acid method, ensuring analytical comparability among survey years (Muitire et al. 2025).

The long‐term fertilization dataset was compiled from the national fertilizer monitoring network coordinated by the Ministry of Agriculture and Rural Affairs, covering observation periods from the early 1970s to 2012. We selected 23 representative experiments across China (Figure S2). The monitored cropping systems encompass the major cereals and cash crops cultivated nationwide, including wheat, maize, soybean, and rice, with detailed information for each long‐term experiment provided in Table S1. All experiments followed a standardized experimental protocol consisting of six fertilization regimes: a no‐fertilizer control (CK), nitrogen plus potassium without P (NK), full chemical fertilization with nitrogen, P, and potassium (NPK), manure alone (M), combined chemical fertilizers and manure (NPKM), and chemical fertilizers with straw incorporation (NPKS). Details of the experimental design are provided in the corresponding technical documentation of the monitoring network (Xu et al. 2015). Soil samples were collected from the 0–20 cm soil layer, and soil available P was determined using the method described above. The LTE dataset was used as an independent experimental reference to contextualize the survey‐based soil available P trajectories by comparing temporal changes among contrasting fertilization regimes.

2.2. Spatial Harmonization and Calculation of Soil Available P Dynamics

To visualize the spatial distribution and temporal evolution of soil available P across China's croplands, georeferenced soil observations from the three historical survey years, namely 1980, 2012, and 2018, were interpolated separately onto a 1 km × 1 km grid using ordinary kriging. The resulting interpolated survey surfaces were constrained to cropland areas using a cropland mask, thereby generating spatially continuous maps of cropland soil available P for each survey period. The cropland domain was further divided into six major agricultural regions according to geographical location and administrative boundaries, namely the Northeast (NE), Central China (CC), Northwest (NW), South China (SC), Southeast (SE), and Southwest (SW) regions (Table S2).

Based on the harmonized soil available P, the annual rate of change in soil available P (AP_rate, mg kg⁻¹ year⁻¹) was calculated for each cropland grid cell as follows:

$${\mathrm{AP}}_{\text{rate}}=\frac{P_{\mathrm{t}2}-P_{\mathrm{t}1}}{\mathrm{t}2-\mathrm{t}1}$$ (1)

where $P_{\mathrm{t}1}$ and $P_{\mathrm{t}2}$ denote the soil available P at the initial survey year t1 and the final survey year t2, respectively. Positive rates indicate increasing soil available P, whereas negative rates reflect declining soil available P. AP_rate was calculated for the full period from 1980 to 2018 and for the two subperiods 1980–2012 and 2012–2018 to evaluate temporal shifts in soil available P dynamics.

2.3. Machine‐Learning Estimation of Long‐Term Changes in Soil Available P

2.3.1. Model Development and Performance

A random forest (RF) model was developed to estimate AP_rate from 1980 to 2018 across China's croplands, based on observation‐derived AP_rate from paired sampling locations between the two years. Prior to model development, all predictor variables were subjected to preprocessing to ensure data quality and consistency. Multicollinearity among predictors was evaluated using both pairwise correlation coefficients and variance inflation factors (VIF). Variables with correlation coefficients greater than 0.7, or those exhibiting the highest VIF values exceeding 10, were excluded to reduce redundancy among predictors (Liu et al. 2021; Li et al. 2022). Finally, 13 predictor variables representing key climatic, environmental, vegetation, and management controls were selected to estimate the AP_rate across China's croplands for the period 1980–2018. Climatic variables included mean annual temperature (MAT, °C) and mean annual precipitation (MAP, mm). Environmental conditions were described by soil organic matter (SOM, g kg⁻¹), soil cation exchange capacity (CEC, cmol kg⁻¹), soil pH, and aluminum (Alo, g kg⁻¹) and iron oxide concentrations (Feo, g kg⁻¹). Vegetation growth was characterized using net primary productivity (NPP, kg C m⁻² year⁻¹) and normalized difference vegetation index (NDVI). Cropland management practices were represented by total carbon (C) inputs (Mg ha⁻¹), nitrogen (N) inputs (kg ha⁻¹), and P inputs (kg ha⁻¹), together with irrigation amount (IRR, mm). Detailed information on these drivers is provided in Table S3. The selected predictors represent key processes governing soil P dynamics, including climatic regulation of biological activity, soil physicochemical controls on P retention and buffering, vegetation‐mediated uptake and cycling, and management‐driven nutrient inputs and redistribution (Deiss et al. 2018; Lambers 2022; Helfenstein et al. 2024; Gao et al. 2026). All covariate datasets were reprojected and resampled using a cropland mask and harmonized to a common coordinate reference system and a uniform spatial resolution of 1 km to ensure consistency within the agricultural spatial framework (Figures S3 and S4). Notably, these datasets were independently developed, ensuring full independence between the predictor variables and the response variable.

To evaluate model generalization, the full dataset was randomly divided into a training subset (70%, n = 1,185) and an independent test subset (30%, n = 508). Continuous predictors were then standardized using the mean and standard deviation derived from the training dataset, and the same transformation was applied to the test data. Hyperparameter tuning was conducted exclusively on the training data using five‐fold cross‐validation, where in each iteration, four folds were used for model training and the remaining fold was used for validation. After identifying the optimal hyperparameters, the final model was retrained using the full training dataset. Model performance was assessed on the independent test dataset using the root mean square error (RMSE), coefficient of determination (R²), model efficiency coefficient (MEC), mean absolute error (MAE), and mean error (ME). During hyperparameter optimization, sampling density bias was corrected using a density‐weighting approach (van Doorn et al. 2024). The inverse of local sampling density was applied as a weighting factor, whereby observations from densely sampled regions received lower weights, while those from sparsely sampled regions were given greater influence. These weights were incorporated both as initial observation weights in model training and in the calculation of the RMSE during hyperparameter optimization. During RF construction, k predictors (k < n) were randomly selected from the n available features to generate decision nodes and subsequent splits, and this process was iterated to grow an ensemble of decision trees. Finally, the trained RF model was applied to the spatial predictor dataset to generate nationwide predictions of the AP_rate. The complete set of optimal hyperparameters for the final selected model is provided in Table S4.

To further quantify prediction uncertainty, we implemented a quantile regression forest (QRF) approach to estimate the conditional probability distribution of AP_rate (van Doorn et al. 2024). Specifically, the median (q0.5), lower (q0.05), and upper (q0.95) quantiles were predicted for each grid cell. Based on these outputs, spatial uncertainty was characterized using the 90% prediction interval (PI90), defined as the difference between the q0.95 and q0.05 quantiles, and the prediction interval ratio (PIR), calculated as the PI90 normalized by the q0.5 prediction. PI90 and PIR represent absolute and relative uncertainty, respectively, enabling a more comprehensive assessment of uncertainty in AP_rate.

2.3.2. Spatial Attribution of Drivers of Soil Available P Change

To elucidate the mechanisms driving the AP_rate across China's croplands, we used an interpretable machine‐learning framework based on SHAP (SHapley Additive exPlanations) to quantify the contributions of climate variability (MAT and MAP), environmental conditions (SOM, CEC, pH, Alo, and Feo), vegetation growth (NPP and NDVI), and cropland management practices (C input, N input, P input, and IRR). SHAP analyses were conducted for both the training dataset and the spatial prediction outputs, where model predictions were decomposed into the marginal contributions of each predictor, enabling the interpretation of predictor effects at the sample level and the characterization of their spatial heterogeneity across 1‐km grid cells. For each grid cell, we calculated the absolute SHAP values (|SHAP|) for the four factor categories and normalized them to derive their relative contribution (RC) as follows:

$${\mathrm{RC}}_i=\frac{\mid {\text{SHAP}}_i\mid }{\sum \mid {\text{SHAP}}_i\mid }\times 100\%$$ (2)

where RC_i denotes the relative contribution of the i‐th driver category (climate, environment, vegetation or management). Using these relative contributions, we determined the dominant driver type for each grid cell. A cell was classified as singly dominated when one category accounted for more than 50% of the total contribution. When no single category exceeded this threshold, the cell was classified as jointly dominated by the two categories with the largest contributions (e.g., “climate‐management” or “vegetation‐environment”). This classification was used to generate a nationwide map of dominant driver types, revealing the spatial heterogeneity in the primary controls on long‐term P dynamics.

2.3.3. Climate and Management Sensitivity Analysis

We evaluated the sensitivity of soil available P dynamics to climate variability and cropland management by applying a set of controlled‐change scenarios to the trained RF model at a spatial resolution of 1 km. For climate variables, temperature was increased by 2°C, and precipitation was increased or decreased by 20% in separate scenarios, while all other variables were held constant. For management variables, C input, N input, P input, and irrigation amount were each increased by 20% in separate scenarios, with all other variables unchanged. Each scenario was then applied to the RF model to generate the corresponding prediction of AP_rate.

To evaluate the capacity of management adjustments to buffer climate‐driven changes in soil available P, we quantified a relative adaptability ratio (RAR) for each grid cell as the ratio of the maximum absolute climate sensitivity to the maximum absolute management sensitivity. Climate sensitivity was defined as the absolute change in the predicted AP_rate under each climate scenario relative to the baseline prediction, and management sensitivity was calculated analogously for each management scenario. Based on the RAR values, croplands were classified into four climate‐management response regimes. Grid cells with RAR ≤ 0.5 were defined as strongly management‐buffered, indicating that management adjustments can more than offset climate‐driven changes. Cells with 0.5 < RAR ≤ 1 were classified as weakly management‐buffered, where management responses remain sufficient but approach the threshold of climatic influence. Cells with 1 < RAR ≤ 2 were categorized as weakly climate‐constrained, indicating partial but incomplete compensation by management adjustments. Finally, grid cells with RAR > 2 were classified as strongly climate‐constrained, where climate‐driven changes substantially exceed the buffering capacity of management interventions.

2.3.4. Sensitivity‐Based Future Projections of Soil Available P

To examine the potential future response space of soil available P dynamics, we defined three sensitivity‐based projection scenarios based on the climate and management sensitivity analysis described above. The business as usual scenario (BAU) represented continuation of current climate and management conditions. The climate change scenario (C) used the AP_rate predicted under controlled changes in climate variables, whereas the management change scenario (M) used the AP_rate predicted under controlled changes in agricultural management inputs. These scenarios were sensitivity‐based and included only controlled climate and management changes, without incorporating ongoing or expected policy‐driven changes in fertilizer use, nutrient use efficiency, or manure recycling.

For each projection scenario, the corresponding AP_rate was applied to the baseline soil available P level in 2018 to estimate soil available P in 2060 at 1 km resolution. This projection assumed that soil available P changes linearly over the projection period and did not account for saturation effects. The resulting spatial distributions were mapped to characterize regional patterns and contrasts in future soil P dynamics across China's croplands.

2.4. Statistical Analysis

All statistical analyses were performed using R ver. 4.4.1 (R Core Team, Vienna, Austria). RF and QRF models were implemented using the “ranger” package (ver. 0.18.0). Hyperparameter tuning was conducted using the “mlr3tuning” package (ver. 1.5.1) framework, based on five‐fold cross‐validation implemented in the “mlr3” package (ver. 1.5.0). Model interpretability was assessed by computing SHAP values with the “fastshap” package (ver. 0.1.1). Sankey diagrams were created using the “networkD3” package (ver. 0.4.1), with flow widths proportional to the proportion of grid cells represented by each transition pathway. Boxplots and SHAP beeswarm plots were generated using the “ggplot2” package (ver. 4.0.1). Spatial maps were produced using ArcGIS 10.2 (Esri, Redlands, CA, USA), and all other figures were created in Origin 2025 (OriginLab Corporation, MA, USA).

3. Results

3.1. Spatial and Temporal Dynamics of Soil Available P

We compiled soil available P data from three nationwide soil surveys conducted in 1980, 2012, and 2018, and separately interpolated the point observations from each survey year onto a 1 × 1 km grid using ordinary kriging to present spatial patterns consistently. Soil available P across China exhibited pronounced spatial heterogeneity together with a persistent upward trend over the observation period (Figure 1). National mean soil available P increased from 7.3 mg kg⁻¹ in 1980 to 19.1 mg kg⁻¹ in 2012 and further to 27.9 mg kg⁻¹ in 2018 (Figure 1a–c). At the beginning of the observation period, soil available P levels were generally low nationwide, consistent with widespread P deficiency (Figure 1a). By 2012, soil available P had increased markedly across China's croplands, with a national average increase of 11.8 mg kg⁻¹, followed by further enrichment reaching 20.6 mg kg⁻¹ by 2018. This increase was particularly pronounced in the NE and CC zones, whereas gains in the NW and SW regions remained comparatively modest (Figure 1b,c). Region‐level statistics confirmed significant increases across all agroecological zones (p < 0.05), with pronounced regional contrasts in the magnitude of change (Figure 1d).

FIGURE 1.

Spatial and temporal dynamics of soil available phosphorus (P, mg kg⁻¹) across China's croplands. (a–c) Spatial distribution of soil available P across China's croplands in 1980 (a), 2012 (b) and 2018 (c), derived from harmonized national soil surveys. Six major agroecological regions are indicated: Northeast (NE), Northwest (NW), Central China (CC), Southeast (SE), South China (SC) and Southwest (SW). (d) Regional variation in soil available P across six major agroecological regions of China in 1980, 2012 and 2018. Boxplots show medians (central solid lines), means (red dashed lines), interquartile ranges (boxes), and data ranges (whiskers). Different letters indicate statistically significant differences among years based on one‐way ANOVA followed by the LSD test (p < 0.05). (e) Transitions in soil available P classes across 1980, 2012 and 2018 at the national scale. Flows in the Sankey diagram represent changes in area among five soil available P classes: < 10, 10–20, 20–40, 40–80 and > 80 mg kg⁻¹. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Nationwide transitions among soil available P classes demonstrated a persistent shift from lower to higher soil available P levels during 1980–2018 (Figure 1e). The proportion of low‐available P soils (< 10 mg kg⁻¹) declined sharply over this period, whereas medium and high soil available P classes (20–40 mg kg⁻¹ and > 40 mg kg⁻¹) expanded substantially across croplands. Most soils that were initially P‐deficient had shifted into higher categories by 2012 and continued to show higher soil available P levels by 2018, revealing widespread and sustained enrichment.

To evaluate whether this survey‐based increase in soil available P was consistent with long‐term external P inputs, we analyzed soil available P responses from 23 long‐term fertilization trials distributed across major cropping regions (Figure 2; Figure S2). Soil available P remained consistently low under unfertilized (CK) and nitrogen plus potassium (NK) treatments, indicating negligible P accumulation in the absence of external P inputs. In contrast, treatments receiving mineral or organic P inputs showed pronounced and sustained increases in soil available P. Mineral fertilization (NPK) and organic amendments (M) both elevated soil available P relative to CK and NK, whereas combined mineral and organic inputs (NPKM) consistently produced the highest soil available P levels, indicating the strongest accumulation of the available P pool under sustained positive P inputs. These treatment‐dependent responses were observed across all agroecological regions, with stronger increases in experiments initiated after 2000. Collectively, the clear divergence between zero‐P and P‐input treatments provides strong experimental evidence that sustained external P inputs have been a principal driver of the long‐term increase in soil available P across China's croplands.

FIGURE 2.

Responses of soil available phosphorus (P, mg kg⁻¹) to long‐term fertilization and temporal change. (a) Mean soil available P under different long‐term fertilization treatments in long‐term field experiments. Treatments include CK (no fertilizer), NK (nitrogen and potassium), NPK (nitrogen, phosphorus and potassium), M (manure), NPKM (NPK plus manure) and NPKS (NPK plus straw return). Points indicate treatment means, and whiskers represent the standard errors. Different letters denote statistically significant differences among treatments based on one‐way ANOVA followed by the LSD test (p < 0.05). (b) Regional distributions of soil available P before and after 2000 across six agroecological regions of China: Northeast (NE), Northwest (NW), Central China (CC), Southeast China (SE), South China (SC) and Southwest (SW). (c) Distributions of soil available P before and after 2000 under different long‐term fertilization treatments. Boxplots show medians (central solid lines), means (red dashed lines), interquartile ranges (boxes), and data ranges (whiskers). Different letters indicate statistically significant differences between the two periods based on one‐way ANOVA followed by the LSD test (p < 0.05). Map lines delineate study areas and do not necessarily depict accepted national boundaries.

3.2. Spatial Acceleration of Soil P Change Across China

The spatial distribution of the AP_rate exhibited pronounced spatiotemporal heterogeneity across China's croplands (Figure 3). Higher rates were observed in the NE, CC, and SE regions, whereas comparatively lower rates occurred in the SW and NW (Figure 3a). During 1980–2012, soil available P generally increased at modest rates nationwide, with annual changes mostly below 0.5 mg kg⁻¹ year⁻¹ (Figure S5a). Higher rates were limited to confined areas, primarily in the NE and CC regions. In contrast, during 2012–2018, soil available P accumulation accelerated markedly across large portions of intensively managed cropland regions, with annual rates frequently exceeding 1.5 mg kg⁻¹ year⁻¹, resulting in a pronounced spatial gradient in accumulation rates (Figure S5b). The NE region exhibited the highest AP_rate, whereas the NW region remained relatively low, despite showing an increase relative to the earlier period. Region‐level statistics further supported these patterns, with mean AP_rate increasing significantly after 2012 across all regions (p < 0.05), although the magnitude of increase varied strongly among regions (Figure 3b). These results indicate a marked acceleration of soil available P accumulation in intensively managed cropland regions after 2012, while accumulation rates in SW remained comparatively moderate.

FIGURE 3.

Spatial patterns and temporal contrasts in the annual rate of change in soil available P (AP_rate, mg kg⁻¹ year⁻¹) across China. (a) Spatial distribution of AP_rate across Chinese croplands during 1980–2018. (b) Comparison of regional mean AP_rate between two periods, 1980–2012 and 2012–2018, across six agroecological regions: Northeast (NE), Northwest (NW), Central China (CC), Southeast China (SE), South China (SC), and Southwest (SW). Bars represent mean values, and whiskers indicate standard errors. Different letters indicate statistically significant differences between the two periods within each region based on one‐way ANOVA followed by the LSD test (p < 0.05). Map lines delineate study areas and do not necessarily depict accepted national boundaries.

3.3. Spatial Differentiation of Drivers of Soil Available P Dynamics

We used paired original observations of soil available P from the 1980 and 2018 national surveys to train and evaluate a random forest model capable of predicting AP_rate across China's croplands (Figure S6). The model showed strong agreement with observations (R ² = 0.67) and good predictive efficiency (MEC = 0.60), together with low prediction errors (RMSE = 0.58; MAE = 0.39) and minimal bias (ME = 0.08), indicating robust performance and good generalization capacity (Figure S6a). Among all predictors, CEC, MAT, P input, and pH were identified as the most influential variables, whereas NDVI showed comparatively moderate contributions (Figure S6b). SHAP analysis further revealed the direction of their effects on AP_rate (Figure S6b). Specifically, P input and CEC were generally associated with positive contributions, reflecting their roles in promoting soil P accumulation. In contrast, NDVI exhibited predominantly negative effects. Soil pH and MAT displayed more complex and often nonlinear relationships with AP_rate, highlighting the context‐dependent regulation of soil P dynamics. Spatially explicit predictions revealed pronounced geographic heterogeneity in long‐term soil available P dynamics (Figure S7). The predicted spatial patterns closely mirrored those inferred from the survey data, which exhibited strong regional contrasts in soil P trajectories over the past nearly four decades.

SHAP‐based attribution was used to quantify the spatial distribution of dominant drivers associated with AP_rate across China's croplands, revealing pronounced spatial heterogeneity (Figure 4). Management, environmental, and climatic factors were identified as the dominant contributors across most croplands, either jointly or individually, whereas vegetation‐dominated regimes were relatively rare and largely confined to warm and humid regions. Notably, jointly dominated patterns were widespread, indicating that AP_rate is typically associated with multiple drivers rather than a single contributor. This co‐dominance highlights the inherent complexity of soil P biogeochemical processes underlying long‐term soil available P dynamics. At the regional level, the six agroecological zones exhibited clear differences in the relative contributions of major driver groups (Figure 4b–g). The CC region was predominantly management‐dominated, reflecting the influence of intensive agricultural inputs. In contrast, the NW and NE regions were primarily characterized by higher contributions from climate‐environment (CE) and environment‐management (EM), suggesting stronger constraints imposed by climatic conditions, soil properties, and management practices. The southern regions (SC, SW, and SE) exhibited a more balanced contribution structure, with an increased role of CE.

FIGURE 4.

Spatial distribution and regional contributions of dominant drivers of the annual rate of change in soil available P (AP_rate, mg kg⁻¹ year⁻¹) across China, based on SHAP. (a) Spatial map of dominant drivers of AP_rate across China during 1980–2018. Grid cells were classified according to the relative dominance of four driver categories: Climate (C), vegetation (V), environment (E) and management (M). Jointly dominated categories are denoted by two‐letter combinations, such as climate‐environment (CE) and environment‐management (EM). Cells were defined as singly dominated when one category contributed more than 50% of the total importance, representing more than half of the SHAP‐based relative contribution, or jointly dominated by the two categories with the largest contributions when no single category exceeded this threshold. (b–g) Relative contributions of different driver categories to soil available P change across six agroecological regions: Northwest (NW, b), Northeast (NE, c), Southwest (SW, d), Central China (CC, e), South China (SC, f) and Southeast (SE, g). Map lines delineate study areas and do not necessarily depict accepted national boundaries.

3.4. Climate–Management Sensitivity and Future Trajectories of Soil Available P

We used the climate and management controlled‐change scenarios to quantify the sensitivity of AP_rate to shifts in temperature, precipitation, and key management inputs across China's croplands (Figure 5a). Strongly and weakly management‐buffered regimes dominated large portions of intensively managed croplands, particularly across major agricultural production regions, indicating that adjustments in fertilization and related management practices may offset or even outweigh climate‐induced changes to soil available P dynamics. In contrast, weakly and strongly climate‐constrained regimes were concentrated in environmentally marginal areas, where AP_rate was more sensitive to climate changes than to management changes. This pattern suggests that climatic variability and environmental constraints may play a stronger role than management adjustments in regulating soil available P responses in these regions. Furthermore, sensitivity‐based future projections to 2060 revealed significantly different responses of soil available P across the three scenarios (Figure 5b; Figure S8). Relative to the BAU, the C scenario resulted in significantly lower soil available P levels. In contrast, the M scenario produced the largest national increase, raising soil available P by approximately 13% compared with the C scenario. These findings reveal a clear climate–management gradient in controlling soil P dynamics, driving divergent future trajectories across China's croplands.

FIGURE 5.

Spatial differentiation of management‐ and climate‐dominated controls on soil available phosphorus (P, mg kg⁻¹) across China. (a) Spatial distribution of cropland sensitivity‐risk areas classified according to the relative dominance of management and climate controls on the annual rate of change in soil available P (AP_rate, mg kg⁻¹ year⁻¹). Grid cells were categorized into four types based on relative adaptability to climate and management: Strongly management‐buffered, weakly management‐buffered, weakly climate‐constrained and strongly climate‐constrained. (b) Projected soil available P in 2060 under the business as usual (BAU), climate (C), and management (M) scenarios. Boxplots show medians (central solid lines), means (red dashed lines), interquartile ranges (boxes), and data ranges (whiskers). Different letters indicate statistically significant differences among scenarios based on one‐way ANOVA followed by the LSD test (p < 0.05). Map lines delineate study areas and do not necessarily depict accepted national boundaries.

4. Discussion

We reconstructed continuous annual trajectories of soil available P from 1980 to 2018 at 1 km resolution and used a multi‐decadal network of long‐term fertilization experiments as independent, process‐based evidence to support these reconstructed trends. This integrated approach quantifies temporal changes in soil available P levels and associated fertility classes, with substantial regional variation across China's croplands. Through a spatially explicit modelling approach that integrates climatic, environmental, vegetation, and management variables, we quantify the dominant controls on long‐term soil P change and identify distinct regional control regimes. These regimes reveal systematic spatial contrasts in the sensitivity of soil available P to management adjustments and climatic constraints that have not been characterized previously at a national extent. We further combined climate‐management sensitivity analysis with scenario projections to link reconstructed historical changes with future soil P responses under climatic constraints and management adjustments. Together, these analyses provide a national‐scale perspective on the long‐term evolution, dominant controls, and regional sensitivity of soil P dynamics in China's croplands, offering evidence to inform region‐specific P management under changing environmental and management conditions.

Our findings demonstrate that nearly four decades of agricultural intensification have driven a sustained and spatially divergent transformation of soil available P dynamics across China's croplands. The widespread rise in soil available P reflects the cumulative legacy of sustained fertilizer inputs exceeding crop P removal, while soil sorption and fixation were insufficient to offset the buildup of labile and plant‐available P (Sattari et al. 2012; Liu et al. 2025; McDowell et al. 2025). This long‐term enrichment was expressed most strongly in the NE, CC, and SE regions, where cropping intensity is high and soils exhibit strong P reactivity (Figure 3; Figure S3). These regional patterns are consistent with evidence from other intensively farmed regions, including the United States, Europe and South Asia, where long‐term positive P budgets have promoted legacy Olsen‐P enrichment in productive agricultural zones, whereas less intensive or environmentally marginal systems show weaker Olsen‐P increases under stronger soil and hydrological regulation (MacDonald et al. 2011; Langhans et al. 2022; McDowell et al. 2025; Figure S3). The SHAP‐based analysis helps interpret the roles of key variables underlying these regional differences. P input acts as a direct source replenishing the Olsen‐P pool, while the negative NDVI effect may reflect stronger vegetation‐mediated P uptake and removal (Lambers 2022; Veneklaas 2022). The positive CEC effect may be explained by stronger surface negative charge and electrostatic repulsion of phosphate ions, which reduce non‐specific P adsorption and thereby contribute to the increase in Olsen‐P (Jiang et al. 2015). The projected decline in soil available P under altered climatic conditions indicates that climatic changes may reshape legacy P trajectories by modulating P transformation processes, crop P removal, and water‐mediated P losses (Ockenden et al. 2017; Bian et al. 2026; Gao et al. 2026). These insights provide a spatially explicit basis for prioritizing P management across regions, helping align nutrient interventions with agronomic demand and environmental risk.

The pronounced spatial heterogeneity in soil P trajectories indicates that P management in China should move beyond uniform national recommendations toward a differentiated regional framework. In management‐responsive regions, where soil P trajectories are strongly shaped by fertilization and other agronomic practices, the results support more targeted interventions such as intensified soil testing, subregional fertilizer recommendation schemes, and nutrient management plans designed to match P inputs with crop demand and existing soil P stocks (Gong et al. 2025; Liu et al. 2025; Van Eynde et al. 2025). In major grain‐producing areas showing persistent P accumulation, management priorities should focus on reducing excessive inputs, drawing down legacy P, and promoting integrated nutrient management to limit further accumulation and reduce runoff‐related pollution risks (Wang et al. 2025). By contrast, in regions with persistent P deficits, especially in parts of the Northwest and Southwest, priorities should shift toward rebuilding soil fertility through improved access to mineral fertilizers, greater use of organic amendments, and practices that enhance P retention and internal recycling (Stewart et al. 2020; Hawkins et al. 2022). The climate‐management sensitivity analysis further refines this regional framework by identifying where management adjustment may be more effective and where greater emphasis should be placed on soil‐condition improvement. This framework provides a spatially explicit basis for prioritizing P management across diverse regional contexts, allowing interventions to be aligned with both agronomic needs and environmental risks. More broadly, it links long‐term soil P trajectories with forward‐looking regional planning, supporting more efficient, resilient, and sustainable P stewardship across China's croplands.

While our study provides a comprehensive assessment of soil available P dynamics across China, several limitations remain. Although harmonized national surveys and long‐term fertilization experiments enabled reconstruction of multi‐decadal soil P trajectories, the underlying observations are uneven in space and time. Spatial interpolation across survey years may therefore introduce heterogeneous uncertainty, especially in environmentally marginal regions with sparse sampling (Liu et al. 2022). However, uncertainty estimated by quantile regression forest (Figure S9) remained within a reasonable range in most areas, supporting the robustness of the main AP_rate patterns. Extrapolation bias was also minimal, as assessed following the method of van den Hoogen et al. (2019), with 98.1% of cropland grid cells falling within the predictor range of the training data. We further acknowledge that change‐based metrics may be affected by regression to the mean and mathematical coupling (Tu and Gilthorpe 2007), although these effects are unlikely to materially alter our conclusions because initial soil available P was not included as a predictor and the sample size was large. In addition, the machine‐learning framework identifies spatial patterns and dominant drivers but does not explicitly represent mineralogical transformations, subsoil P dynamics, or long‐term legacy P mobilization (Wadoux et al. 2020; Ludemann et al. 2024; Zhang et al. 2024), and SHAP‐based attribution reflects relative predictor contributions rather than causality. Finally, the scenario analysis does not capture feedbacks from changing cropping systems, policy interventions, or emerging technologies. Future work should integrate long‐term monitoring, remote sensing, finer‐resolution management data, and process‐based models to better resolve regional P imbalances and soil P resilience under ongoing climate and agricultural change.

5. Conclusions

Our study provides a comprehensive reconstruction of nearly four decades of soil available P dynamics across China's croplands and reveals strong regional divergence in both temporal trajectories and dominant controls. By integrating harmonized national soil surveys, long‐term fertilization experiments and a spatially explicit modelling framework, we show that soil P trajectories are jointly regulated by management intensity, soil buffering capacity and climatic conditions. Soil P accumulation has accelerated in high‐input production regions, whereas environmentally marginal landscapes remain more strongly constrained by soil and climate conditions. The results further identify where management adjustments are likely to be effective and where climate and environmental constraints limit its influence. These patterns have direct implications for precision fertilization and regionalized P management: improving fertilizer recommendations, soil testing, and nutrient input combinations should be prioritized in management‐responsive regions, whereas soil improvement and erosion control are likely to be more important in climate‐constrained regions. This framework therefore helps connect long‐term soil P dynamics with region‐specific management priorities.

Author Contributions

Haiqing Gong: conceptualization, writing – review and editing, funding acquisition. Zhenling Cui: supervision, writing – review and editing. Zhen Xu: conceptualization, formal analysis, visualization, writing – original draft. Yulong Yin: writing – review and editing, supervision. Zhong Chen: writing – review and editing. Pengqi Liu: writing – review and editing. Shuaibing Wu: writing – review and editing.

Funding

This work was supported by the National Key Research and Development Program of China, 2023YFD190150101 and National Natural Science Foundation of China, 32402673.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Spatial distribution of soil sampling sites across China's croplands. (a–c) Locations of soil available phosphorus (P) sampling sites used for national analyses in 1980 (a), 2012 (b), and 2018 (c). Each dot represents one sampling site collected from the top 0–20 cm soil layer during the corresponding survey period. The datasets comprise 1,693 samples in 1980, 35,883 in 2012, and 151,402 in 2018, respectively, covering all major agro‐ecological zones across China. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S2: Geographic distribution of the long‐term fertilization experiments across China. Locations of the 22 experimental sites (23 experiments), spanning the major agroecological regions and representing the dominant cereal and cash‐crop production systems. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S3: Spatial distribution of climatic, environmental, and vegetation factors in China's croplands. (a) MAT. (b) MAP. (c) SOM. (d) pH. (e) CEC. (f) Alo. (g) Feo. (h) NPP. (i) NDVI. MAT, mean annual temperature (°C). MAP, mean annual precipitation (mm). SOM, soil organic matter (g kg⁻¹). pH, soil pH. CEC, soil cation exchange capacity (cmol kg⁻¹). Alo, Al oxide concentration (g kg⁻¹). Feo, Fe oxide concentration (g kg⁻¹). NPP, net primary productivity (kg C m⁻² year⁻¹). NDVI, normalized difference vegetation index. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S4: Spatial distribution of management factors in China's croplands. (a) C input. (b) N input. (c) Pinput. (d) IRR. C input, carbon inputs (Mg ha⁻¹). N input, nitrogen inputs (kg ha⁻¹). P input, phosphorus inputs (kg ha⁻¹). IRR, irrigation amount (mm). Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S5: Spatial distribution of the annual rate of change in soil available phosphorus (AP_rate, mg kg⁻¹ year⁻¹) across China's croplands. (a, b) Maps show the spatial pattern of AP_rate during two periods: 1980–2012 (a) and 2012–2018 (b). Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S6: Model performance and SHAP‐based interpretation for predicting the annual change rate of soil available phosphorus (AP_rate, mg kg⁻¹ year⁻¹). (a) Model evaluation on the independent test dataset, showing the relationship between observed and predicted AP_rate. The black dashed line represents the 1 to 1 line. (b) Variable importance and SHAP‐based interpretation of AP_rate. MAT, mean annual temperature (°C). MAP, mean annual precipitation (mm). SOM, soil organic matter (g kg⁻¹). pH, soil pH. CEC, soil cation exchange capacity (cmol kg⁻¹). Alo, Al oxide concentration (g kg⁻¹). Feo, Al oxide concentration (g kg⁻¹). NPP, net primary productivity (kg C m⁻² year⁻¹). NDVI, normalized difference vegetation index. C input, carbon inputs (Mg ha⁻¹). N input, nitrogen inputs (kg ha⁻¹). P input, phosphorus inputs (kg ha⁻¹). IRR, irrigation amount (mm).

Figure S7: Random forest–predicted spatial distribution of the annual rate of change in soil available phosphorus (AP_rate, mg kg⁻¹ year⁻¹) across China's croplands. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S8: Projected soil available phosphorus (P, mg kg⁻¹) in 2060 under contrasting climate and management scenarios. (a) The business as usual (BAU) scenario, representing continuation of current climate and management conditions. (b) The climate (C) scenario, incorporating projected changes in climate variables. (c) The management (M) scenario, reflecting adjustments in agricultural management inputs. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S9: Spatial distribution of prediction uncertainty in the annual rate of change in soil available phosphorus (AP_rate, mg kg⁻¹ year⁻¹) across China's croplands. (a) 90% prediction interval (PI90, mg kg⁻¹ year⁻¹), representing absolute uncertainty. (b) Prediction interval ratio (PIR), representing relative uncertainty. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Table S1: The information on 23 long‐term fertilization experiments.

Table S2: The division of China's six major agricultural regions.

Table S3: Detailed information on all drivers.

Table S4: Optimal hyperparameters of the random forest model.

Data Availability Statement

The datasets and associated resources supporting the main findings of this study are available at https://doi.org/10.5281/zenodo.21213681.