Global crop yields can be lifted by timely adaptation of growing periods to climate change

Rule-based mean sowing and maturity dates

Location- and climate-specific mean crop calendars are computed by combining two rule-based approaches published by¹⁹ and²² to simulate sowing and physiological maturity dates of grain crops, respectively. The assumption is that farmers select growing seasons based on the mean climatic characteristics of their specific location and on the physiological limitations (base and optimum temperatures for reproductive growth; sensitivity to terminal water stress) of the respective crop species. Accordingly, they select sowing dates and cultivars with phenologies that, on average, meet these adapted maturity dates.

The climate is classified into (i) seasonality types, based on the coefficient of variation of monthly mean temperature and precipitation and (ii) temperature levels, based on the temperature of the warmest month as compared to the base and the optimum temperatures for the crop reproductive growth. Optimal temperatures for sowing, optimal temperature ranges for grain filling, as well as indicators of soil moisture conditions (based on precipitation/potential-evapotranspiration ratio (P/PET)), are defined as global parameters for each crop (Supplementary Table 1) and used as thresholds to identify the best timing for sowing and for the start or end of the crop grain-filling phase. To cope with fluctuations of daily values around these thresholds, mean daily temperature, precipitation and potential evapotranspiration are derived by linear interpolation between monthly values.

We distinguish between spring and winter crop types. Maize, rice, sorghum, and soybean are simulated as spring crops only, for wheat we simulate both types. For spring crops, farmers sow the crops at the onset of the wet season (first day of the wettest 120 consecutive days), in case of prevailing precipitation seasonality, or on the day of the year when temperatures increase above crop-specific temperature threshold¹⁹ (Supplementary Table 1), in case of temperature-driven seasonality.

For wheat, we distinguish three types: winter wheat with vernalization is chosen if monthly temperatures fall below 0 °C, but winter is neither too harsh (temperature of the coldest month is higher than −10 °C), nor too long (temperatures fall below the sowing temperature threshold (12 °C) after 15th September (North hemisphere) or 31st March (South hemisphere)¹⁹). Winter wheat without vernalization is grown if winters are mild (the temperature of the coldest month is higher than 0 °C) without dormancy. In this case, wheat is sown 75 days before the coldest month of the year. This rule was arbitrarily chosen based on observed wheat sowing dates in mild winter regions. If the conditions for growing any of the winter-wheat types are not met (winter too harsh and too long), then spring wheat (without vernalization) is chosen. Note that the computed sowing dates do not differ between rainfed and irrigated for any of the crops.

The mean maturity date is chosen so that the crop grain-filling phase, the most critical for yield formation, occurs under the least stressful conditions possible in that location and climate as follows. Under precipitation seasonality, grain filling starts towards the end of the rainy season, when a P/PET threshold is crossed. Under temperature seasonality, (a) grain filling of spring crops starts in the warmest month of the year (if summer temperatures are optimal), or right after temperatures return within an optimal range; (b) grain filling of winter crops ends in the warmest month of the year (if summer temperatures are optimal), or right before temperatures exceed the optimal range; (c) eventually, maturity is advanced to escape terminal water stress. Note that the grain-filling phase has a static duration of 60 days for maize and 40 days for all the other crops. This assumption is based on empirical relationships between the total growth period and the post-flowering reproductive phase, showing that the partition between the vegetative and reproductive phase of grain crops follows a saturation curve that levels off after 90–100 days of total growth duration⁵⁴. Different crops are assumed to have only one crop cycle (sowing-to-maturity) per year, therefore neither multi-cropping systems nor crop rotations are accounted for in the decision-making rules. A detailed description of the rules and parameterization can be found in refs. 19, 22.

Simulated crop calendars reflect current farmers’ management

Simulated historical crop calendars, driven by the bias-corrected climate dataset WFDEI²³, largely agree with observations^11,12,13. We compare results both at the country and grid-cell level because, although the observed crop calendars used here are gridded datasets, their underlying sources are often reported per country. The country-level comparison highlights that the agreement is good for most countries, importantly, including those with large cropland area. The area-weighted Mean Absolute Error (MAE) is close or well below 30 days for all considered crops (Fig. 4). The simulated crop calendars compare well with the observed data also at the grid-cell level. Large areas, including major agricultural regions of importance for global yields, show deviations within ±15 days for both sowing and maturity dates (Supplementary Table 2 and Supplementary Figs. 21–24). However, evaluating the accuracy below 30 days is limited by the time resolution of the observations, which is either (i) monthly¹¹ and converted by us into daily values, by taking the mid-day of the reported month, or (ii) daily^12,13, but resulting from averages over large time windows (often > 1 month). Overall, the accuracy of the model is in line with the original evaluations of this rule-base method^19,22, as well as with other studies simulating average growing periods across large regions^18,20.

Simulation of daily crop phenology and yields with the LPJmL crop model

We perform a modeling experiment across the global land grid at 0.5° × 0.5° resolution. We used the LPJmL5 crop model^24,25 to simulate daily growth and phenological development of five crops, driven by climate projections from four General Circulation Models (GCMs) GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR and MIROC5 under the Representative Concentration Pathways 6.0 (RCP6.0) as provided in bias-adjusted form from the CMIP5 archive by the ISIMIP2b project⁴². Irrigated and rainfed production systems are simulated separately on their current harvested areas¹¹, which is also used to compute total crop yields at grid-cell and global scale, as the product of yield by crop-specific area. A first 5000-year spin-up simulation is used to initialize all model pools (e.g., soil carbon and nitrogen content). A second spin-up simulation of 390 years is used to introduced effects of historical human-driven land-use change on these pools. A change in cropping area for the future scenarios is not considered in this study.

Phenological development is simulated based on the thermal-time model, including the effect of vernalization. All crops are assumed to be insensitive to photoperiod, due to a lack of parameters for multiple-crops and global-scale simulations. Previous global studies^15,18 that have focused on maize and wheat only, have found lower performances in the growing-period simulations when using a photo-thermal model, compared to a temperature-only driven approach and thus recommend caution when using the photoperiodic response. State-of-the-art global crop models^13,16 also typically do not consider sensitivity to photoperiod or assume that the photoperiodic response of the cultivars chosen in each location are perfectly tuned to the given conditions.

Sowing dates are prescribed based on the external rule-based algorithm. Crop cultivars are parametrized based on the phenological units required to reach the corresponding maturity dates (TU_req, °C days). In line with¹⁵, TU_req are derived consistently with the phenological module of the crop model LPJmL for each grid cell, crop, and rule-based computed growing period from the respective climate input. They are calculated as the sum of daily mean air temperature increments above a crop-specific base temperature (TU) (Supplementary Table 1) between rule-based sowing and maturity. In addition, winter-wheat cultivars require effective vernalization days (VU_req), that range between 0 (mild winters) and 70 (cold winters), depending on the temperature of the 5 coldest months (Eq. (1))^15,18.

where T_m is the mean temperature of the month.

From the day of sowing, effective TU for phenological development are accumulated daily, as the difference between the mean air temperature on that day and the crop-specific base temperature for phenological development (Eq. (2)). The vernalization effectiveness is computed daily by a scaling factor (0–1), which is then multiplied to the TU (Eq. (2)). For crops that are insensitive to vernalization, VU_d is set equal one.

where the scaling factor VU_d is computed by a three-stage linear response function with a range of optimal temperatures (Eq. 3). Temperature for effective vernalization range between −4 °C and +17 °C, with an optimum range between 3 °C and 10 °C.

In this study, we have removed the effect of vernalization on slowing down TU accumulation until 10% of the total vernalization requirements is reached. In this way, the crop can accumulate both vernalization units and heat units in fall, so that there is some leaf growth before winter (in LPJmL, the LAI curve depends on accumulated heat units).

The LPJmL model simulates phenology as one single phase from emergence to maturity. Although the flowering stage is not simulated as an explicit break point, the fraction of above-ground biomass that is allocated to the storage organs (fHI) depends on the phenological progress (fTU_req, fraction of TU_req that have been fulfilled), with the bulk of the storage organs start filling up after 40% of TU_req have been reached (Eq. (4)). In line with this, the LAI curve reaches a plateau when 45% (wheat) or 50% (other crops) of the TU_req are fulfilled, which could be considered a proxy of the flowering stage.

Crop biomass growth is simulated by daily carbon accumulation and allocation to different plant organs (roots, leaves, storage organs, mobile reserves, and stem). The fraction of carbon allocated to each pool is a function of the fraction of completed phenological progress. Water stress increases allocation to the roots and reduces allocation to the leaves. The daily Net Primary Production (NPP) is the result of the Gross Primary Production (daily gross photosynthesis) reduced by the respiration costs. Gross photosynthesis is simulated as a function of absorbed photosynthetically active radiation, CO₂ atmospheric mixing ratio, air temperature, day length, and canopy conductance. Photosynthesis rate is given by the minimum between light-limited and Rubisco-limited photosynthesis rates, with distinguished pathways for C₃ and C₄ crops. Respiration is tissue-specific and it is also driven by temperature. If accumulated NPP is insufficient to satisfy all organ demands, allocation follows a hierarchical order from roots, to leaves, to storage organs, and consequently penalizing the harvest index. Crops are subject to yield failure due to frost events (daily minimum temperature <−5 °C) occurring during grain yield formation (50% <fPHU <95%).

All management inputs other than sowing dates and cultivars have been provided according to global gridded datasets (0.5° × 0.5° spatial resolution) of current and historical practices. Land-use datasets for the period 1510–2015, of crop- and irrigation-specific area fraction and of mineral nitrogen fertilizer annual application rates (kg N ha⁻¹ year⁻¹), were based on the LUH2v2⁵⁵ dataset. This has been dis-aggregated and remapped using the MADRaT tool⁵⁶, in order to match the crop classification to the Crop Functional Types simulated by LPJmL²⁴. The crop residue management input dataset (1850–2015) was obtained from ref. 56, which estimates crop residue removal rates from FAOSTAT national statistics on residue-related management practices (e.g., burning on field, household fuel and livestock fodder). The crop-specific spatial and temporal patterns of tillage and no-tillage systems for the period 1973–2010 were derived from ref. 57.

Crop-model skills are preserved when driven by simulated crop calendars

We run LPJmL with both simulated and observed historical growing periods, forced by the historical observation-based climate dataset WFDEI²³, and test model outputs against a standard benchmark for global gridded crop models evaluation⁵⁸. Forcing LPJmL with simulated crop calendars preserves the model skills in simulating global spatio-temporal patterns of crop yields, as indicated by the three metrics to compare simulated and observed national yields (correlation coefficient, the root mean squared difference, and the variance) (Supplementary Fig. 25).

Data processing and analysis

We assess the future adaptation potentials over the present-day cropland¹¹. To quantify the performances of cropping systems, we compute grid-cell level 20-years area-weighted average crop yields (t ha⁻¹) in each simulation period and GCM. We quantify climate change impacts with and without adaptation as the difference between future and historical yields. Furthermore, we isolate the adaptation effect from that of climate, by computing yield differences between contrasting management scenarios (e.g., with and without adaptation) under the same future period (2080–2099) for each climate scenario separately, where crops experience the same climate change scenario but are grown in different growing periods. As crop yield interannual variability metric, we use the coefficient of variation computed as year-to-year yield deviation normalized by the mean yield over the same period (1986–2005 for the reference scenario and 2080–2099 for all other scenarios). We account for uncertainties from climate scenarios by expressing the evaluation metrics at the aggregated level (e.g., global yield losses) as the mean and range across the four GCMs. In the main text, spatial patterns of growing period and yield changes are shown as the mean change per grid cell across four GCMs. The results for the individual GCMs are reported in the Supplementary Figures.

We compute the area-weighted mean absolute error (MAE) at the grid-cell and country levels as (Eq. (5)):

where n is the number of grid cells (countries), A_i is the crop-specific area of grid cell (country) i, S_i, and O_i are the simulated and observed average dates (sowing or maturity) of grid cell (country) i in Julian days.

For data processing, we used R⁵⁹ and R-packages for handling netcdf4⁶⁰, performing computation^61,62, and plotting results⁶³. The country borders displayed in all figures that include maps were based on the Natural Earth dataset, which is available at naturalearthdata.com.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Source: Ecology - nature.com