Spatial frameworks for robust estimation of yield gaps

Meeting future food demand on existing cropland, without further conversion of natural ecosystems, requires understanding where and by how much crop yields can increase. Yield gap—the difference between actual farm yield and the biophysical yield potential under well-managed, stress-free conditions—is a key indicator of exploitable production potential with current land and water resources. Many global assessments use top-down, gridded crop models with coarse inputs on climate, soils, and cropping systems to estimate yield potential and gaps. Recent country-level studies indicate these top-down approaches may be unsuitable for prioritizing agricultural R&D investments. An alternative bottom-up framework, exemplified by the Global Yield Gap Atlas (GYGA), estimates yield potential and gaps using locally measured weather, finer-scale soils, and validated crop models at representative sites, then upscales to larger areas using production-weighted aggregation. This study explicitly evaluates top-down (GAEZ, AgMIP) versus bottom-up (GYGA) approaches across multiple crops, regions, and spatial scales, and assesses implications for food security planning and AR&D prioritization.

Prior global studies on food security, land use, biodiversity, and climate have relied on top-down gridded modelling frameworks to estimate yield potential and yield gaps (for example, Nature, Nat. Commun., PNAS). However, evidence from specific countries suggests these top-down estimates can be inaccurate for guiding investments. The GYGA bottom-up methodology has been documented as more locally relevant by leveraging measured weather, finer soil data, and models calibrated against high-quality experiments. Independent validations in data-rich regions (e.g., Australian wheat) show bottom-up spatial sampling and upscaling produce estimates consistent with detailed datasets at climate-zone and national scales. Studies also demonstrate that coarse gridded weather and soil inputs can bias simulated yield potential, and that global crop calendars and generic cultivar parameters used in top-down models often misrepresent dominant cropping systems. Despite widespread use of both approaches, a direct, multi-scale performance comparison had been lacking.

The study compares yield potential (Yp for irrigated, Yw for rainfed), yield gaps (Yp or Yw minus actual yield), and extra production potential from two top-down frameworks—Global Agro-ecological Zones (GAEZ) and the AgMIP global gridded crop model ensemble median—against the bottom-up Global Yield Gap Atlas (GYGA). Analyses span three spatial levels: local (buffers), subnational climate zones (CZ), and national/subcontinental. Target crops include maize, rice, and wheat across key regions (e.g., US Corn Belt, Asia lowland rice, Australia wheat, sub-Saharan Africa maize), with broader coverage documented in supplementary materials. Data sources: GYGA uses measured weather (more recent periods), finer-scale soils, and locally validated cropping system data; GAEZ and AgMIP use global gridded weather (1961–1990 for GAEZ; 1980–2010 for AgMIP), coarse soils and cropping calendars. For comparability, top-down grid outputs were aggregated to buffers, CZs, and countries using SPAM 2010 crop-specific harvested area as weights, mirroring GYGA’s upscaling. Actual yields (Ya) were taken from GYGA to avoid bias and allow water-regime disaggregation. Yield gaps were computed as Yp or Yw minus Ya; negative gaps indicate underestimation of potential. Extra production potential was estimated by multiplying yield gaps by harvested area (SPAM 2010). Agreement between approaches was quantified via root-mean-square error and mean absolute error between top-down and bottom-up estimates at each spatial scale and water regime. For food self-sufficiency implications, the study computed self-sufficiency ratios (SSR = production/demand) for five cereals (maize, millet, rice, sorghum, wheat) in sub-Saharan Africa under a 2050 scenario where farmers achieve 80% of Yw (rainfed) and 80% of Yp (irrigated), with no cropland expansion, using UN population projections and IMPACT model per-capita demand. The comparison evaluates how differences in Yp/Yw propagate to SSR outcomes.

• Across four case studies and multiple scales, AgMIP estimates of yield potential are on average 60% lower than GYGA, leading to much more conservative extra production estimates. GAEZ agrees better with GYGA at national/subcontinental levels but still shows discrepancies from −50% to +30%, which are larger at local scales (−95% to +480%).
• Specific examples: In the US Corn Belt, GYGA Yp is 8% higher than GAEZ and 63% higher than AgMIP for maize. For rainfed wheat in Australia, GYGA Yp is 46% higher than AgMIP. For lowland rainfed rice in Asia and maize in sub-Saharan Africa, top-down methods show narrow Yp ranges across climate zones and often reverse rankings relative to GYGA.
• Top-down approaches frequently produce negative yield gaps (implying Yp/Yw < actual yields), indicating underestimation of potential: at local scale, GAEZ negative yield gaps occur in 13% of maize, 3% of rice, and 3% of wheat locations; AgMIP negative gaps occur in 39% (maize), 45% (rice), and 25% (wheat). GYGA reports no negative yield gaps.
• Food self-sufficiency in sub-Saharan Africa (SSR by 2050 with 80% of potential on existing cropland) differs markedly by approach: GAEZ SSR = 1.36 (surplus), GYGA SSR = 1.03 (near balance), AgMIP SSR = 0.96 (deficit). At national levels, AgMIP SSR deviates from GYGA by −24% to +39%, and GAEZ predicts surpluses for most countries whereas GYGA and AgMIP often indicate deficits.
• The large uncertainties in top-down estimates have major implications for prioritizing AR&D investments and for studies on land use, climate, and biodiversity that rely on production potential metrics.

Findings show that the spatial framework choice strongly affects estimated yield potential, yield gaps, and derived metrics like SSR, thereby influencing AR&D priority setting and policy decisions. The prevalence of negative yield gaps in top-down outputs signals substantial underestimation of potential and cautions against using such estimates for investment targeting. Likely causes include reliance on coarse, gridded weather and soil datasets, misrepresentation of cropping systems (cropping intensity, sowing windows, water regimes), and generic crop model parameters not tailored to local cultivar responses. Bottom-up methods, grounded in measured data, locally validated cropping system characterization, and model calibration against well-managed experiments, provide more accurate and locally relevant estimates and can be validated by local experts. While top-down approaches offer consistency and lower effort for global coverage, their uncertainties necessitate integrating bottom-up estimates and datasets to improve robustness. Even when aggregate subcontinental estimates appear similar, large within-region discrepancies persist, underscoring the need for fine-scale accuracy to guide interventions.

The study provides a comprehensive, multi-scale comparison showing that widely used top-down frameworks (GAEZ, AgMIP) often underestimate yield potential and produce implausible negative yield gaps, leading to divergent and potentially misleading conclusions about extra production potential and future food self-sufficiency. Incorporating bottom-up, data-rich approaches such as GYGA can substantially reduce uncertainty and improve the relevance of estimates for AR&D investment decisions and policy planning. The authors advocate wider adoption of bottom-up frameworks using primary weather data, finer-scale soils, and accurate cropping system information, while recognizing opportunities for complementarity: leveraging bottom-up results to inform and validate top-down models, and using top-down methods in data-scarce regions until better local data become available. Robust, bottom-up-informed estimates are essential for credible foresight on food security, land use, and climate change impacts.

• Data requirements: Bottom-up approaches demand granular measured weather, detailed soils, and accurate cropping system data, limiting application in data-scarce regions.
• Temporal mismatches: Yp/Yw simulations use different historical periods across datasets (GAEZ 1961–1990, AgMIP 1980–2010, GYGA more recent), potentially affecting comparisons.
• Potential author bias: Although authors contributed to GYGA, they state neutrality and full transparency for replication; nevertheless, inherent bias cannot be entirely dismissed.
• SSR scenario assumptions: No cropland expansion, fixed net planted area, and no explicit inclusion of future genetic gains in Yp (or negative climate change impacts beyond SSP2 demand trajectory) may affect absolute SSR levels.
• Top-down input constraints: Coarse gridded weather, soil maps, and generic crop calendars/parameters in top-down models contribute to uncertainty; resolving these requires data not always available globally.