Reducing risks of antibiotics to crop production requires land system intensification within thresholds

Human activities in the Anthropocene have boosted socioeconomic development while creating environmental risks, including widespread antibiotic contamination of soils from manufacturing, medical use, and agricultural practices (e.g., manure fertilization and wastewater irrigation). Soil antibiotic pollution can inhibit plant growth, disrupt soil functioning, and reduce crop yields. Despite attention to antibiotic footprints, large-scale risks to crop production remain poorly quantified due to data limitations and scale complexities. Land system intensification—via expansion of arable land and increased inputs such as manure and irrigation—can simultaneously increase yields and exacerbate soil antibiotic pollution. Understanding how these land system regimes influence antibiotic risks across spatial scales, and identifying safe thresholds for intensification, is essential for risk management. This study addresses three questions: (i) Do antibiotic pollution risks reduce crop production at a broad scale? (ii) How do land systems affect soil antibiotic pollution risks with changing spatial scales? (iii) How can land system intensification be managed to balance antibiotic pollution risk and crop yield based on thresholds where risk–yield tradeoffs peak? Using China as a case, the risk quotient (RQ) of antibiotics to arable crop growth is predicted and linked with production of maize, rice, wheat, and vegetables, examining scale-dependent effects and inverted U-shaped relationships between intensification and risk–yield tradeoffs.

Prior studies have documented widespread antibiotic use and emissions to soils via manure and wastewater, with observed ecological impacts on soil organisms and plant growth. Spatial simulations of soil antibiotics exist but often have limited geographic scope and do not directly quantify risks to crop production. Land use and management have been highlighted as key drivers of antibiotic residues in soils, while population density and economic indicators show weaker direct associations, suggesting human impacts are mediated through land systems (land use composition and management practices). The development Kuznets curve hypothesis has been used to describe nonlinear relations between development pressures and environmental outcomes, motivating the exploration of inverted U-shaped relations between land system intensification and risk–yield tradeoffs.

Study scope and targets: Nine frequently detected, toxic antibiotics in Chinese soils were analyzed: tetracycline (TC), chlortetracycline (CTC), oxytetracycline (OTC), doxycycline (DOX), ofloxacin (OFL), norfloxacin (NOR), ciprofloxacin (CIP), enrofloxacin (ENR), and lomefloxacin (LOM). Data collection: A comprehensive measured environmental concentration (MEC) dataset was compiled, primarily from Zhang et al., supplemented by new field sampling in Yunnan and Zhejiang. Samples (generally 0–20 cm depth) were processed and analyzed per Supplementary Methods. Selection criteria excluded experimentally treated samples, severely polluted sites (e.g., hospitals, livestock farms), and assessed spatial autocorrelation (Moran's I = 0.03, Z = 0.22, p > 0.05). The final dataset contained 484 georeferenced locations. Risk assessment: Risks were quantified as risk quotients (RQ = MEC/PNEC). PNECs were estimated via assessment factors from EC50 or LC50 (AF = 1000) or NOEC (AF = 100). Where soil PNECs were unavailable, PNECsoil was derived from PNECwater using equilibrium partitioning with soil–water partition coefficients (Ka). For worst-case characterization, the lowest PNECs and maximum exposure assumptions were used. Cumulative risk per grid was computed by concentration addition as the sum of highest RQs for the nine antibiotics. Predictor variables: Spatially explicit covariates included climate (temperature, precipitation), land use (arable, built-up, natural proportions), anthropogenic factors (livestock densities; population; GDP; nitrogen and phosphorus fertilization; pesticide use), soil properties (clay, organic carbon, bulk density, soil thickness, saturated hydraulic conductivity), groundwater depth, NDVI, and terrain. All datasets were harmonized to 1 km resolution; temporally varying predictors used 2000–2020 series, with limited-change variables assumed temporally stable. Scaling-up and modeling: For each antibiotic, log10(RQ) was modeled using Random Forest (RF) with standardized predictors (Z-scores), up to 1000 trees, implemented via randomForest and caret in R 4.2.1. Data were split 70/30 for training/validation. An ensemble of 500 RF models (Monte Carlo resampling of training data) produced mean predictions and uncertainties (standard deviations), generating 1-km maps of individual and cumulative RQs for soils (excluding bare land and water). Model accuracy was assessed via RMSE and Pearson correlations between measured and predicted values. Land system indicators and scale analysis: Land system composition (proportion arable, built-up, natural) and management (manure nitrogen application rate, kg N/km²/yr; irrigated area proportion) were computed per watershed. Four watershed scales (HydroSHEDS levels renamed: level 1 large to level 4 small) represented grain changes. To assess scale extent effects, 20 subdatasets were created by omitting upper or lower segments (0–90% in 10% steps) along a human footprint gradient. For each subdataset and scale, linear regressions related antibiotic RQs to land use, management, population, and GDP; best models were selected via AICc with model averaging (MuMIn::dredge). Contributions of predictors were expressed as percentages of explained variance from standardized coefficients. Risk–yield relationships and tradeoffs: Crop yields (maize, rice, wheat, vegetables) were related to cumulative RQ using generalized additive models (GAMs) with basis dimension 5 to capture nonlinearities. Risk–yield tradeoff was defined as standardized yield minus standardized cumulative RQ. Land system intensification was computed as the sum of min–max standardized arable proportion, manure rate, and irrigated proportion. A moving-window approach estimated nonlinear relations between intensification and tradeoffs, applying a development Kuznets curve (inverted U) to identify thresholds where the slope equals zero. Uncertainty in thresholds was derived from regression parameter uncertainties (n varies by scale: 17 at level 1; 177 at level 2; 414 at level 3; 818 at level 4).

• Model performance and national risk patterns: Ensemble RF models showed good agreement between modeled and observed RQs (correlations 0.48–0.68; differences generally <1 RQ unit). The average cumulative soil RQ across China was 6.1 ± 2.1. Highest cumulative risks were concentrated in central and eastern China. Ensemble relative uncertainty averaged ±26%, with notable uncertainty in major grain-producing regions (Huang-Huai-Hai and Northeast Plains).
• Compound-specific risks: Ofloxacin contributed the highest individual risk (RQ 0.76 ± 0.37). Approximately 11.4% of land was contaminated by more than one antibiotic compound.
• Crop yield sensitivity to antibiotic risk: Nonlinear relations were observed between cumulative RQ and yields of maize, rice, wheat, and vegetables. When cumulative RQ exceeded thresholds of approximately 8.30–9.98, yields declined substantially, falling below yields at zero risk, indicating that antibiotic pollution can offset benefits of intensification.
• Scale-dependent drivers: Across 20 subdataset analyses and four watershed scales, antibiotic RQs correlated strongly with land system indicators (land use, management) and weakly with population and GDP, underscoring mediation via land systems. With downscaling, the contribution of arable land to explaining RQs increased from 12.9 ± 3.2% to 59.5 ± 12.7%. Manure application was a consistent positive driver (increasing risk in 16 of 20 analyses at small watershed level). Management effects dominated at large scales, while land use composition gained importance at smaller scales.
• Risk–yield tradeoffs and thresholds: Tradeoffs (yield minus risk) exhibited convex, inverted U-shaped relations with land system intensification across scales. Thresholds at which tradeoffs peaked were crop- and scale-specific. Vegetables and wheat showed higher thresholds than maize and rice, especially at small scales. Example threshold ranges: manure fertilization—vegetables 523.5–1801.6 kg N/km²/yr; wheat 672.1–1186.9 kg N/km²/yr. Irrigated area proportion—vegetables 54.4–61.5%; wheat 50.0–61.7% (small scales); rice 46.0–66.5% (large scales). Arable land proportion—vegetables 34.4–39.1%; wheat 32.7–44.9% (small scales). These indicate higher resistance of vegetable and wheat production to antibiotic risk under intensification at small scales.
• Variance explained by land systems: Land systems accounted for a substantial fraction of variability in antibiotic risks across spatial scales (reported overall range 21–66%).

The study demonstrates that soil antibiotic pollution is a widespread, quantifiable risk to crop production at national scale, and that these risks are predominantly governed by land system intensification rather than population or economic indicators per se. By coupling ensemble RF risk mapping with crop yield analyses, the work shows that beyond cumulative RQ thresholds (~8.3–10), antibiotic pollution can negate yield gains from intensification. The observed scale dependence—management dominating at larger extents and land use composition gaining importance at smaller watersheds—clarifies why relationships vary with spatial grain and extent. This aligns with the notion that antibiotics enter soils chiefly via land management pathways (manure, wastewater irrigation), while land use expansion increases the spatial footprint of exposure. The inverted U-shaped risk–yield tradeoffs provide actionable thresholds for manure application, irrigation extent, and arable area proportion, particularly highlighting that vegetable and wheat systems can tolerate higher intensification before tradeoffs turn negative, especially in small-scale watersheds. These findings directly address the research questions by (i) confirming the negative impact of high antibiotic risk on yields, (ii) attributing risk variability to land systems with clear scale effects, and (iii) identifying intensification thresholds to guide management for balancing production and environmental safety.

This work delivers the first broad-scale, high-resolution assessment linking soil antibiotic pollution risks to crop production across multiple spatial scales in China. It shows that cumulative RQ thresholds around 8.30–9.98 mark substantial yield reductions, and that land systems—particularly manure application, irrigation, and arable land extent—are the primary, scale-dependent drivers of risk. The study operationalizes risk–yield tradeoffs via inverted U-shaped relations with land system intensification and provides quantitative thresholds that can inform sustainable intensification strategies. To achieve win–win outcomes, intensification should remain below crop- and scale-specific thresholds, with particular attention to small-scale watershed management for vegetable and wheat systems. Future research should incorporate experimental validation, mechanistic biophysical models, and explicit accounting of antibiotic resistance dynamics and biotic interactions (soil and plant microbiomes), along with improved monitoring to refine risk estimates and management thresholds.

(i) Risk assessments may be conservative because the modeling assumed maximum exposure and minimum tolerance (worst-case scenarios), potentially overestimating risks; however, this provides precautionary guidance. (ii) The analysis focused on crop exposure risks and did not incorporate antibiotic resistance development or detailed biotic mediators (soil and plant microbiomes), partly due to data limitations, although correlated environmental covariates were included. (iii) Unreported antibiotics were assumed absent; some may be present below detection limits. (iv) Results are based on modeling without dedicated experimental validation, limiting mechanistic inference; enhanced biophysical modeling and additional datasets are needed for multi-compartment risk predictions.