Nikolay Khabarov

Global crop–climate impact studies using gridded, process-based models are essential for understanding how climate change may affect agricultural productivity. Due to high computational demand, these models are typically driven by a limited subset of projections from up to five global climate models (GCMs) from CMIP experiments. Yet, there is limited understanding of whether these subsets adequately capture variability in climate drivers within the whole ensemble and how the captured variability in drivers translates into crop yield impacts. Here, we analyze climate and crop yield projections from 29 GCMs contributing to CMIP6 ScenarioMIP that provide all required forcing variables and have been bias-adjusted and downscaled. To address computational constraints, we apply a high-performant crop model emulator to estimate yields of four staple crops across 150 k global locations under the four core emission pathways and for the whole climate ensemble from 1984–2099. We then aggregate results to major regions and use linear programming with the objective to minimize ensemble sizes while bracketing projected changes in temperature, precipitation, or crop yields. I.e., we identify the smallest GCM subsets that capture each variables’ full range. We find that ensemble sizes required to bracket ranges for individual variables increase from temperature (around 5–20 GCMs) to precipitation and further to crop yields (about 10–29 GCMs each), depending on the selection of crops and scenarios, and the strictness of the boundaries. Capturing yield responses across contrasting crops and emission pathways commonly requires the full ensemble. Notably, bracketing climate variables alone fails to capture the range of yield outcomes, highlighting the limited predictive value of individual climatic drivers. Our results emphasize the need for thorough contextualization of GCM selection in impact studies and suggest that emulators can fill a gap in complementing mechanistic model simulations with comprehensive climate uncertainty quantification.

Global crop–climate impact studies using gridded, process-based models are essential for understanding how climate change may affect agricultural productivity. Due to high computational demand, these models are typically driven by a limited subset of projections from up to five global climate models (GCMs) from CMIP experiments. Yet, there is limited understanding of whether these subsets adequately capture variability in climate drivers within the whole ensemble and how the captured variability in drivers translates into crop yield impacts. Here, we analyze climate and crop yield projections from 29 GCMs contributing to CMIP6 ScenarioMIP that provide all required forcing variables and have been bias-adjusted and downscaled. To address computational constraints, we apply a high-performant crop model emulator to estimate yields of four staple crops across 150 k global locations under the four core emission pathways and for the whole climate ensemble from 1984–2099. We then aggregate results to major regions and use linear programming with the objective to minimize ensemble sizes while bracketing projected changes in temperature, precipitation, or crop yields. I.e., we identify the smallest GCM subsets that capture each variables’ full range. We find that ensemble sizes required to bracket ranges for individual variables increase from temperature (around 5–20 GCMs) to precipitation and further to crop yields (about 10–29 GCMs each), depending on the selection of crops and scenarios, and the strictness of the boundaries. Capturing yield responses across contrasting crops and emission pathways commonly requires the full ensemble. Notably, bracketing climate variables alone fails to capture the range of yield outcomes, highlighting the limited predictive value of individual climatic drivers. Our results emphasize the need for thorough contextualization of GCM selection in impact studies and suggest that emulators can fill a gap in complementing mechanistic model simulations with comprehensive climate uncertainty quantification.

Global gridded crop models (GGCMs) are important tools for assessing climate impacts on agriculture, yet significant divergence in their projections limits interpretability, and impact studies often treat GGCMs as black boxes. Targeted ensemble sensitivity analyses are demanding and not transferable to different ensembles. Here, we comprehensively evaluate climatic and soil drivers of crop yield anomalies in a state-of-the-art GGCM ensemble, using maize as a representative crop. Gradient boosting classifiers detect anomalies, SHapley Additive exPlanations (SHAP) values quantify feature importance, and methods are applied to a recent GGCM experiment driven by reanalysis climate data. We find broadly similar climatic drivers across the ensemble, though feature importance distributions differ. Low precipitation dominates under rainfed conditions, while solar radiation typically ranks second, highlighting that drought impacts depend on atmospheric water demand often omitted from sensitivity analyses. In some GGCMs, excess rather than insufficient water drives anomalies. The comparison with climate sensitivities derived from a global yield reference dataset indicates that no single GGCM performs best across all climate regions. Individual GGCMs may respond well to one climatic stressor but not others, underscoring that bulk benchmarking of GGCMs against observations can obfuscate processes requiring targeted improvement. Our findings demonstrate that evaluating opportunistic data, i.e., experiments produced for other purposes, yields vital insights into GGCM divergence in impact studies. Code is publicly available on GitHub to support future attribution analyses and inform broad audiences about drivers of observed results.

Global crop–climate impact studies using gridded, process-based models are essential for understanding how climate change may affect agricultural productivity. Due to high computational demand, these models are typically driven by a limited subset of projections from up to five global climate models (GCMs) from CMIP experiments. Yet, there is limited understanding of whether these subsets adequately capture variability in climate drivers within the whole ensemble and how the captured variability in drivers translates into crop yield impacts. Here, we analyze climate and crop yield projections from 29 GCMs contributing to CMIP6 ScenarioMIP that provide all required forcing variables and have been bias-adjusted and downscaled. To address computational constraints, we apply a high-performant crop model emulator to estimate yields of four staple crops across 150 k global locations under the four core emission pathways and for the whole climate ensemble from 1984–2099. We then aggregate results to major regions and use linear programming with the objective to minimize ensemble sizes while bracketing projected changes in temperature, precipitation, or crop yields. I.e., we identify the smallest GCM subsets that capture each variables’ full range. We find that ensemble sizes required to bracket ranges for individual variables increase from temperature (around 5–20 GCMs) to precipitation and further to crop yields (about 10–29 GCMs each), depending on the selection of crops and scenarios, and the strictness of the boundaries. Capturing yield responses across contrasting crops and emission pathways commonly requires the full ensemble. Notably, bracketing climate variables alone fails to capture the range of yield outcomes, highlighting the limited predictive value of individual climatic drivers. Our results emphasize the need for thorough contextualization of GCM selection in impact studies and suggest that emulators can fill a gap in complementing mechanistic model simulations with comprehensive climate uncertainty quantification.

Shaded Arabica coffee production in agroforestry systems, as opposed to full-sun production, is a nature-based solution improving soil water balance, reducing heat exposure of coffee plants, and supporting sustainable forest management as opposed to deforestation. For this coffee production system in Mexico, which is dominated by smallholders as the largest group of coffee producers, we herein analyze current and estimate future yields. For the first time, to our best knowledge, this is done with a process-based coffee agroforestry model CAF2014 that we adapted for geo-spatial applications and named CAF2014-Rhaobi. Modeling of smallholders’ representative management is based on tree thinning, pruning frequency, and nitrogen supply through fertilizer and litter from nitrogen-fixing shade trees. Modeled historical yields generally agree with the reported numbers; however, there are discrepancies explained by modeling assumptions and simplifications. While shade trees help sustain coffee production, the projected drop in yields under present management is about 30% at the end of the century compared to the present as estimated using an ensemble of CMIP6 SSP5-8.5 climate projections. Economic analysis for three typologies of Mexican small coffee producers (conventional low, high-efficiency, and organic) reveals the major role of farmer associations and organic coffee price premiums in making production economically sustainable. This emphasizes the need for innovative marketing approaches and policies supporting farmers opting for certified production.

Shaded coffee production in agroforestry systems, as opposed to full sun production, is a nature-based solution (NbS) that helps maintain soil water balance and reduce heat exposure of coffee plants. It is part of a range of NbS co-produced with stakeholders in the project SAbERES, which aims at supporting climate change adaptation for small-scale producers in Mexico. For this coffee production system, we analyze current and estimate future yields of small coffee growers in Mexico by employing a process-based coffee growth model CAF2014 adapted for geo-spatial applications and named CAF2014-Rhaobi. A range of climate projections reflecting the SSP5-8.5 scenario until 2100 is taken from an ensemble of five CMIP6 climate models to bracket climate ensemble response.As NbS in agriculture are typically based on complex ecological interactions, a first crucial step in their modelling is the analysis of model sensitivity to its key inputs and validation of its ability to reflect reported yields. Particular attention was paid to the model’s sensitivity to adjustments in plot management such as shade trees pruning, projected changes in precipitation, hydrological soil parameters, and implications of using different soil datasets. The modeling of smallholders’ representative management was carried out based on parametrizations derived from literature. This informed key parameters of fertilizer application including nitrogen supply by litter from N-fixing shade trees and shading cover management, i.e., tree thinning and pruning frequency. Besides the quantification of crop yield changes per se, the project will analyze economic implications based on the spatial distribution of coffee yields and prices as reported by the Mexican Agri-Food and Fisheries Information Service (SIAP).The modelled historical coffee yields are found to be in good agreement with the SIAP reported numbers, while there is a clear overestimation in the south-western part of the coffee producing region of Mexico. This is explained by a range of modeling assumptions and simplifications rendering the model less representative for this region. While shade trees provide some resilience, average drop in shaded coffee yields under present management estimated for the majority of the agro-environmentally diverse coffee producing regions in Mexico across all climate projections is about 25% at the end of the century. There are only few regions that are able to maintain their historical yields. These preliminary results underpin that shade trees as a single NbS do not suffice for climate adaptation in the long run under high warming conditions but will need to be combined with other measures. Future work may include refinement of modeling assumptions based on stakeholders input and analysis of economic implications driven by yield change estimates.

Spatio-temporal patterns of extreme climate events have been extensively studied, yet two questions remain underexplored: Do such events occur regularly, and how do regularity patterns change under global warming? We address these questions by investigating dominant periods in crop failure, heatwave, and wildfire data. Here, we show that under pre-industrial conditions dominant periods emerge in 28% of cropland exposed to crop failure and 10% of wildfire-affected areas, likely related to climatic oscillations such as the El Niño-Southern Oscillation, while heatwaves occur irregularly. The number of dominant periods increases by 2–13% during the transition from the pre-industrial era to the anthropocene. In the anthropocene, the occurrence of extreme events shifts towards monotonic growth, replacing previous natural regularity patterns. Linearly de-trended projections reveal an additional shift towards smaller dominant periods due to climate change. These shifts in regularity are crucial for adaptation planning, and our method offers an additional approach for studying extreme events.

Global gridded crop models (GGCMs) are simulation tools designed for global, spatially explicit estimation of crop productivity and associated externalities. Key areas for their application are climate impact and adaptation studies. As GGCMs are typically computationally costly and require comprehensive data pre- and post-processing, GGCM emulators are gaining increasing popularity. Earlier emulators have typically been published pre-trained on synthetic weather and management combinations. Here, we present a novel computational pipeline CROp Model Emulator Suite (CROMES) v1.0 that serves for flexibly training GGCM emulators on data commonly available from GGCM simulations. Essentially, CROMES consists of modules to (1) process climate data from daily resolution netCDF files to (sub-)growing season aggregates as climate features, (2) combine various feature types (climate, soil, crop management), (3) train emulators using machine-learning algorithms, and (4) produce predictions. Exemplary, we apply CROMES to train emulators on simulations for rainfed maize from the GGCM EPIC-IIASA and climate projections from a single GCM to subsequently test their skill in predicting crop yields for unseen climate projections from other GCMs. Depending on the training and target data, the regression statistics between GGCM simulations and predictions across all points in time and space are in the ranges R² = 0.97 to 0.98, slope = 0.99 to 1.01, and intercept = −0.06 to +0.06. The RMSE ranges between 0.49 and 0.65 t ha⁻¹. Spatially, patterns are evident with lowest performance in (semi-)arid regions where aggregation of weather data may result in higher information loss while permanent crop growth limitations may hamper evaluation statistics as well. The gain in computational speed for predictions is at more than an order of magnitude with time required to produce target features and subsequent predictions at about 30min on common hardware. We expect CROMES to be of utility in covering more comprehensively uncertainty in climate impact projections, evaluations of adaptation options, and spatio-temporal assessments of crop productivity.

Global gridded crop models (GGCMs) are simulation tools designed for global, spatially explicit estimation of crop productivity and associated externalities. Key areas for their application are climate impact and adaptation studies. As GGCMs are typically computationally costly and require comprehensive data pre- and post-processing, GGCM emulators are gaining increasing popularity. Earlier emulators have typically been published pre-trained on synthetic weather and management combinations. Here, we present a novel computational pipeline CROp Model Emulator Suite (CROMES) v1.0 that serves for flexibly training GGCM emulators on data commonly available from GGCM simulations. Essentially, CROMES consists of modules to (1) process climate data from daily resolution netCDF files to (sub-)growing season aggregates as climate features, (2) combine various feature types (climate, soil, crop management), (3) train emulators using machine-learning algorithms, and (4) produce predictions. Exemplary, we apply CROMES to train emulators on simulations for rainfed maize from the GGCM EPIC-IIASA and climate projections from a single GCM to subsequently test their skill in predicting crop yields for unseen climate projections from other GCMs. Depending on the training and target data, the regression statistics between GGCM simulations and predictions across all points in time and space are in the ranges R2=0.97 to 0.98, slope=0.99 to 1.01, and intercept=-0.06 to +0.06. The RMSE ranges between 0.49 and 0.65 t ha^-1. Spatially, patterns are evident with lowest performance in (semi-)arid regions where aggregation of weather data may result in higher information loss while permanent crop growth limitations may hamper evaluation statistics as well. The gain in computational speed for predictions is at more than an order of magnitude with time required to produce target features and subsequent predictions at about 30min on common hardware. We expect CROMES to be of utility in covering more comprehensively uncertainty in climate impact projections, evaluations of adaptation options, and spatio-temporal assessments of crop productivity.