Agricultural input shocks affect crop yields more in the high-yielding areas of the world

Modern, industrialized food production depends heavily on off-farm inputs—synthetic fertilizers, machinery, energy, pesticides, seeds and feed—often imported from a limited number of countries. While the impacts of weather extremes and trade disruptions on food supply are increasingly understood, far less is known about how large-scale reductions or shocks in agricultural input availability would affect crop yields and food security. Prior analyses have been regional or focused on specific policy scenarios, leaving a gap in systematic global assessment. This study addresses that gap by quantifying how single and combined shocks to key inputs (N, P, K fertilizers, pesticides, and machinery) would affect yields of 12 major crops worldwide at 5 arcmin resolution. The goals are to identify vulnerable crops and regions, determine which inputs drive the largest yield losses, and provide evidence relevant to food system resilience under geopolitical and supply-chain disturbances.

Previous work has explored food system disturbances from weather and trade shocks (for example, Dall’erba et al.; Ferguson and Gars), but comprehensive global analyses of input shocks are scarce. Beckman et al. used economic models to examine EU Green Deal targets (−50% pesticides, −20% fertilizers), finding heterogeneous outcomes and large wheat production reductions in the EU (−33%). Jansik et al., via expert interviews, estimated that total shocks to farm chemicals, fertilizers and pesticides could reduce Finnish yields by 10–40%. Other studies highlight fertilizer price and availability constraints in sub-Saharan Africa and the importance of balanced nutrient management. Pradhan et al. identified fertilizer and soil quality management as key levers for closing global yield gaps, with pesticides generally less influential at global scale. Comparisons in this paper align with these findings: modeled wheat production changes under a 25% all-inputs shock are broadly consistent with Beckman et al.’s policy scenario outcomes; and literature on P and K supply constraints (e.g., O’Hara et al.) underscores sensitivity of crops like maize to P and K availability. Machine learning (random forests) has outperformed linear models for yield prediction in several contexts (Jeong et al.; Leng and Hall), but prior global RF studies used limited input variables (often climate or N only). This study is among the first to apply RF globally with trade-dependent agricultural input layers.

Study scope: Twelve major food crops (barley, cassava, groundnut, maize, millet, potato, rice, sorghum, soybean, sugar beet, sugarcane, wheat). Baseline yields and harvested areas are from Monfreda et al. (1997–2003 averages). Results are generated at 5 arcmin (~10 km at equator). Inputs and covariates: Synthetic fertilizers (N, P, K) from Mueller et al.; non-mineral (local) fertilizers from EarthStat (held constant across scenarios); machinery intensity proxy from USDA International Agricultural Productivity (country-level horsepower per 1,000 ha; gap-filled with continental averages and non-crop-specific); agricultural labor (FAOSTAT) per 1,000 ha (held constant); pesticides from Maggi et al. (top 20 crop-specific actives) aggregated into four groups (herbicides, insecticides, fungicides, others). Pesticide application rates were rescaled by dividing each grid-cell rate by the global 97.5th percentile for that active, values truncated to [0,1], then summed by group. Irrigation share of harvested area from MIRCA2000 (Portmann et al.). Soil covariates include plant-available soil P (McDowell et al.), soil N (SoilGrids v2), and soil organic carbon (SoilGrids v2), aggregated to 5 arcmin and converted to t/ha. Climate control via climate bins: Each crop’s global area is partitioned into 25 climate bins using Growing Degree Days (with base and cut-off temperatures) and annual precipitation derived from AgMERRA (1990–2010 averages). Bins are defined by quintiles of GDD and precipitation, yielding equal data counts per bin, to isolate input–yield relationships within broadly similar climates. Random forest modeling: Separate RF regressions are built for each crop×climate-bin using yield as the dependent variable and the agricultural inputs plus soil organic carbon as predictors (soil N and P are also described among soil parameters; SOC explicitly included as predictor; other soil variables inform context). To reduce overfitting and spatial autocorrelation, a 1-degree grid partitions data into 12×12 blocks of 5 arcmin cells; 75% of blocks used for training and 25% for testing. RF hyperparameters use defaults based on preliminary tuning: 1,000 trees, nodesize=5, and mtry=2 (R package randomForest 4.7-1.1; R 4.0.4). Model performance is assessed by NSE and RMSE on held-out testing data, and baseline country-level yields are validated against FAOSTAT (1997–2003). Shock scenarios: Seven scenario types at three severities (−25%, −50%, −75%): N-only shock; P-only shock; K-only shock; machinery shock; pesticide shock; combined fertilizer shock (N+P+K); and all-inputs shock (fertilizers+pesticides+machinery). For each crop×bin, model training and scenario prediction are iterated 25 times with different train/test samples to quantify variability; coefficients of variation are computed for baseline predictions. Scenario yields are interpreted as selecting comparable cases within the same climate bin exhibiting similar (reduced) input levels. Production impacts are computed by multiplying scenario yields (t/ha) by baseline harvested area (ha) and comparing to baseline production. Assumptions and boundaries: Inputs and yields largely represent circa-2000 conditions; irrigation infrastructure and non-mineral fertilizer and labor proxies are held constant across scenarios; the model does not capture adaptive responses (e.g., crop switching), long-term soil nutrient dynamics, economic/trade feedbacks, or national input stock buffers.

• Model performance: Across crops and climate bins, 95% of RF models have NSE>0.65 and 79% have NSE>0.75, indicating good to very good predictive skill; baseline country-level yields align well with FAOSTAT (R^2>0.8 weighted by production for all crops).
• High-yield systems are most vulnerable: Within bins, grid cells with the highest baseline yields experience the largest declines under shocks; low-yield areas often change little and can sometimes increase, indicating potential yield gap patterns in low-input regions (notably sub-Saharan Africa and South Asia).
• Fertilizers dominate yield losses: Shocks to N, P, K—especially combined fertilizer shocks—drive most of the yield reductions. For many crops, fertilizer-shock impacts resemble those from all-inputs shocks, underscoring fertilizer importance.
• Pesticides generally have smaller global effects on yield decreases, although specific commodities (e.g., sugarcane) show notable production sensitivity.
• Climate-bin heterogeneity: For maize, temperate-like bins (e.g., 6, 9, 11, 12, 13) are highly sensitive to P shocks; bin 5 to K shocks; bin 3 to machinery shocks. Some bins (e.g., 21) show little sensitivity across shocks.
• Production impacts (all-inputs shock): Effects are nonlinear with shock size. For wheat, a 25% all-inputs shock reduces global production by ~15%, and a 50% shock by ~20%. Maize shows the largest global production declines: >25% at 50% shock and nearly 40% at 75% shock. Abstract-level estimates indicate up to −26% (maize) and −21% (wheat) for 50% shocks in all inputs.
• Geographic hotspots under all-input shocks (−50%): Significant declines in the western United States (barley, maize, potato, wheat), northern Argentina (barley, maize, millet, potato, sorghum, soybean), and Central/Western Europe (barley, maize, potato, wheat, sugar beet). Rice declines are pronounced in Thailand, Vietnam, and southern India.
• Country-level impacts: Large relative production losses (>50%) in Denmark, Oman, the UK, New Zealand, and Saudi Arabia. Among top producers (Brazil, China, India, Thailand, USA), the USA shows the largest relative decline (−28% across the 12 crops) and the largest absolute decline (−140 Mt), followed by Brazil (−114 Mt).
• Exporter-specific impacts at 50% all-inputs shock: Maize—Argentina −47%, Brazil −36%, China −16%, France −32%, USA −34. Rice—India −7%, Thailand −8%, USA −27%, Vietnam −29. Soybean—Argentina −22%, Brazil −13%, USA −22. Wheat—Australia −1%, Canada −29%, Germany −48%, France −39%, USA −19%, Russia −11.
• Some low-input areas show yield increases under shocks, consistent with yield-gap patterns and potential rebalancing of nutrient stoichiometry and management within climate bins.

The analysis directly addresses the question of how shocks or reductions in key agricultural inputs affect global crop yields and production, revealing that the most productive, high-input systems—the world’s breadbaskets—are disproportionately vulnerable. Consistent with prior work on yield-gap closure, fertilizers (N, P, K) emerge as the primary drivers of shock-induced yield losses, while pesticide shocks generally have smaller global impacts on yield decreases. The climate-bin approach highlights that vulnerability is context-specific: sensitivity to specific inputs varies across temperature–precipitation regimes and crops. Comparisons with existing studies show broad agreement in magnitude and direction (e.g., Beckman et al. for EU/US/China wheat outcomes under reduced inputs), enhancing confidence in the findings. The results imply that widespread fertilizer supply disruptions or deliberate reductions would materially curtail global production—especially maize and wheat—amplifying food security risks through higher prices and potential export restrictions, with pronounced impacts on importer-dependent regions (e.g., MENA for wheat, Central America for maize, West Africa for rice). The observed increases in some low-input regions suggest untapped potential if management imbalances are corrected, but they do not imply that reducing inputs generally raises yields; rather, they reflect intra-bin exemplars of higher efficiency. Overall, the findings underscore the centrality of fertilizer security and balanced nutrient management to global food system resilience.

This study provides the first global, high-resolution assessment of how single and combined shocks to key agricultural inputs affect yields and production across 12 major crops. Using a climate-bin constrained random forest framework, it shows that high-yielding breadbasket regions are most at risk, with fertilizer shocks driving the bulk of yield losses. Under a 50% all-inputs shock, maize and wheat production could decline substantially (maize by >25%, wheat by ~20%), with disproportionate impacts in North America, Western Europe, parts of South America, and Southeast/South Asia. To enhance resilience, regions highly dependent on synthetic, often imported fertilizers should diversify and, where feasible, substitute with sustainable local nutrient sources (e.g., organic fertilizers), while improving nutrient balancing, soil health, and irrigation management. The maps and scenario data can inform national and global risk assessments, contingency planning, and policy on strategic input reserves and diversified sourcing. Future research should: (1) incorporate additional inputs (e.g., seeds), agronomic practices, and long-term soil nutrient dynamics; (2) couple biophysical shock impacts with economic/trade models to capture price effects, substitution, and adaptive responses; (3) represent national input stockpiles and logistics; and (4) refine gridded input datasets to reduce spatial autocorrelation and improve model fidelity, particularly in low-yield, data-sparse regions.

• Temporal representativeness: Most input and yield datasets reflect circa-2000 conditions; relationships are assumed stable over time.
• Data granularity: Limited unique values for fertilizer rates (often subnational/county-level averages) and country-level machinery proxies may induce spatial autocorrelation and constrain model precision.
• Model scope: Random forests may perform poorly at extremes and cannot extrapolate beyond training ranges; low-yield, data-sparse regions are less well captured. The model cannot account for adaptations (e.g., crop switching), labor reallocation, or management changes following shocks.
• Short vs long term: Short-term buffering (e.g., soil P pools) and long-term soil fertility feedbacks are not modeled; irrigation infrastructure is held constant.
• Economic and trade dynamics: No integration with market responses, trade re-routing, or policy measures; no data on national reserve stocks of inputs, which could mitigate impacts.
• Pesticide aggregation: Pesticides are grouped and rescaled; active-specific effects and interactions are simplified, potentially obscuring nuanced outcomes.