The impact of climate change on the productivity of conservation agriculture

Conservation agriculture (CA) is being promoted as a set of management practices that can sustain crop production while providing positive environmental benefits. However, its impact on crop productivity is hotly debated, and how this productivity will be affected by climate change remains uncertain. Here we compare the productivity of CA systems and their variants on the basis of no tillage versus conventional tillage systems for eight major crop species under current and future climate conditions using a probabilistic machine-learning approach at the global scale. We reveal large differences in the probability of yield gains with CA across crop types, agricultural management practices, climate zones and geographical regions. For most crops, CA performed better in continental, dry and temperate regions than in tropical ones. Under future climate conditions, the performance of CA is expected to mostly increase for maize over its tropical areas, improving the competitiveness of CA for this staple crop. The authors assess the productivity of conservation agriculture systems for eight major crops under current and future climate using a global-scale probabilistic machine-learning approach, revealing substantial differences in yield gain probabilities across crop type, management practice, climate zone and geography.

C onservation agriculture (CA) is a crop production system based on three principles: minimum soil disturbance (going as far as no tillage (NT)), permanent soil cover with crop residues and diversified crop rotation (with at least three crop species involved) 1 . In compliance with sustainability goals, CA is designed to sustain crop production in the long term while improving crop resilience to climate change and protecting the environment. Benefits of CA have been demonstrated in terms of enhancing soil carbon sequestration, improving soil quality, reducing soil erosion and increasing biodiversity 2,3 . However, since crop yield depends on many interacting factors, including local climate conditions 4 , soil characteristics 5 and management practices 6 , it is difficult to assess the potential of CA to increase agricultural productivity. Besides CA (as defined according to the above principles, which follow the Food and Agriculture Organization's (FAO) approach of CA; ref. 7 ), this issue also applies to its variants such as NT without crop rotations and soil cover, NT with soil cover but no rotation and NT with rotation but no soil cover. On the basis of the findings of recent meta-analyses 5,6,8 , these systems are likely to lead to a yield reduction 6 compared to conventional tillage (CT), except for regions facing water limitations. However, the heterogeneity of the experimental results on CA (and their variants) versus CT is very large and their outcome varies as a function of climate conditions 6 and management practices 6,8,9 . The studies of Pittelkow et al. 6,8 relied on a synthetic aridity index to characterize the climate, which makes it hard to analyse the response of CA productivity to interannual weather variability or to predict the impact of future changes in climate. To date, a comprehensive, global synthesis of the impact of climate change on the productivity of CA with respect to CT systems is still lacking.
Here we compared the crop yields of CA systems (and their variants) versus CT under current and future climate conditions on the basis of a new global database 10,11 of paired yield observations of CA (and their variants) versus CT. This dataset includes the most recently published experimental studies on this topic, a detailed description of agricultural management practices (including crop irrigation, fertilization, weed and pest control, soil cover management and crop rotation) and a broad set of climatic variables from external databases, such as precipitation (P) 12 , minimum air temperature (T min ) 13 , average air temperature (T ave ) 12 , maximum air temperature (T max ) 13 and potential evapotranspiration (E) 14,15 over the crop growing seasons 16 . As an indicator of water availability for crops, the precipitation balance (PB) was defined as (P -E) over the growing season 5 .
A machine-learning model based on random forest 17 was trained and cross-validated on the basis of 4,403 paired yield data of CA (and its variants) versus CT from the dataset. The model was used to map the probability of yield gain from CA-like systems versus CT systems, ( Yield CA or variants Yield CT > 1 ) , considering different agricultural management practices successively (with/without soil cover, rotation, weed and pest control, irrigation, fertilization). The analysis was conducted at a spatial resolution of 0.5° latitude × 0.5° longitude on the basis of current (2011-2020) and future (2051-2060) average climate conditions for eight major crops worldwide (spring barley, cotton, maize, rice, sorghum, soybean, sunflower and winter wheat). For each crop, maps were then produced and compounded to derive the accumulated fractions of the cropping area achieving a given yield gain probability of CA versus CT under current climate conditions. Similarly, the accumulated fractions of the cropping area achieving a given level of change in this probability under future climate conditions were calculated. These fractions show the proportions of cropping area with low to high probabilities of yield gain from CA versus CT and with low to high changes for this probability under climate change. Both proportions were computed at the global scale and across different climate zones 18  Conservation agriculture (CA) is being promoted as a set of management practices that can sustain crop production while providing positive environmental benefits. However, its impact on crop productivity is hotly debated, and how this productivity will be affected by climate change remains uncertain. Here we compare the productivity of CA systems and their variants on the basis of no tillage versus conventional tillage systems for eight major crop species under current and future climate conditions using a probabilistic machine-learning approach at the global scale. We reveal large differences in the probability of yield gains with CA across crop types, agricultural management practices, climate zones and geographical regions. For most crops, CA performed better in continental, dry and temperate regions than in tropical ones. Under future climate conditions, the performance of CA is expected to mostly increase for maize over its tropical areas, improving the competitiveness of CA for this staple crop.
due to the presence of soil cover and use of crop rotation but does not systematically outperform CT ( Supplementary Fig. 1a1-h1 and 1a4-h4). With well-managed field fertilization and integrated weed and pest control, CA stands high chances of outperforming CT except in tropical regions ( Fig. 1 and Supplementary Fig. 1a2-h2). The performance of CA is also slightly improved in non-irrigated fields, except in tropical regions ( Fig. 1 and Extended Data Fig. 1).
For most of the studied crops, the overall probability of yield gain from CA is higher in continental, dry and temperate regions than in tropical regions (Fig. 1a,d-g and Supplementary Fig. 1a2-g2). The probability of yield gain with CA is particularly low for rice. For this crop, the probability of yield increase/gain is lower than 0.5 (which indicates a higher probability of yield loss) over about 60% of its global cropping area and in about 85% of its cropping area in the tropics ( Fig. 1d and Supplementary Fig. 1d2).
For several crops and climate regions, the estimated effect of climate change on the probability of yield increase with CA is relatively moderate. Over approximately half of the cropping areas, a decrease of up to 10% in this probability is expected, while in the other half an increase of up to 15% may be anticipated (Fig. 2). However, in some important cases, the effect of climate change is stronger, especially for maize in tropical regions where the probability of yield gain with CA increases in about 70% of the cropping area. Besides, for more than 20% of the maize cropping area in this climate zone, the increase on the probability of yield increase is higher than 10% (Fig. 2c). An increase in yield gain is also expected for more than 60% of the cropping area for rice in dry regions and for soybean in tropical regions (Fig. 2d,f). This fraction rises to more than 70% of the cropping area for sorghum in continental regions (Fig. 2e).
Probabilities of yield gain with CA show important geographical variations under both current and future climate conditions for maize (Figs. 3 and 4) but also for other crop species ( Supplementary  Fig. 2). Yield gains with CA systems and their variants are more likely in relatively higher latitude regions (>40°) than lower latitudinal bands for barley, cotton, rice, sorghum, soybean and sunflower ( Supplementary Fig. 3), in line with the results showed on Fig. 1. CA systems and their variants perform better in northwestern America, northwestern India, northern sub-Saharan Africa and southern Russia than in northeastern America, western and central Europe and the central part of sub-Saharan Africa ( Fig. 3 and Supplementary  Fig. 2). For maize, in the absence of fertilizer inputs and integrated weed and pest control, the probability of yield gain from CA is higher than 0.5 in dry areas of western United States, southern Russia, northern India and the North China Plain. The yield gain probability drops to 0.4 in the Laurentian Plateau of Canada, northcentral and northeastern United States and part of western and central Europe (Fig. 3b). For other crop species, CA has a higher chance of leading to a yield loss compared to CT in the tropical regions ( Supplementary Fig. 2a2-h2), southern China (Supplementary Fig. 2e2-h2), northeastern America and western and central Europe ( Supplementary Fig. 2a2,e2,f2,h2). Conversely, when implemented with field fertilization and integrated weed and pest control, CA has a higher chance of outperforming CT in the major cropping areas for barley, sunflower and wheat ( Supplementary Fig. 2a1,g1,h1 and Supplementary Fig. 5).
Our results also show that soil cover has a stronger (and positive) effect on yield gain probability than other management practices, such as fertilizer inputs, weed and pest control and crop rotation. Thus, without soil cover, NT systems show a lower probability of yield gain than full-blown CA ( Supplementary Fig. 2), which is in line with the input importance ranking obtained with our model (Extended Data Fig. 2). Note that, although management practices have a substantial impact on yield gain probabilities, they have limited influences on the geographical variations of these probabilities (Supplementary Fig. 1 and Supplementary Fig. 2).
The maps reporting the differences of yield gain probabilities between current (2010-2020) and future (2051-2060) climate conditions (Supplementary Table 1  To assess the model sensitivity to climate models and climate change scenarios, we plotted the fractions of global cropping areas corresponding to increasing levels of yield gain probability (from −0.1 to 0.2) for four different climate models and RCP scenarios (Extended Data Fig. 3). The choice of the climate models had very little impact on the results (Extended Data Fig. 3i-p), although the results obtained with Hadgem2-es and Ipsl-cm5a-lr are somewhat more extreme than those obtained with Gfdl-esm2m and Miroc5. The sensitivity to the climate change scenarios was more important (Extended Data Fig. 3a-h). Although the main conclusions remain similar across all RCP scenarios, the stronger changes in yield gain probability are obtained under RCP 8.5 compared to RCP 6.0, RCP 4.5 and RCP 2.6 (Extended Data Fig. 3a-h). In particular, the changes of yield gains become larger for maize and rice under RCP 8.5.
In contrast to previous studies on the productivity of CA5, 5,6,8,19 here the probabilities of yield gain resulting from a shift from CT to several variants of CA systems have been mapped for current and future climate scenarios. Thus, our results offer meaningful and new information for policymakers, agricultural extension services and farmers. Relying on a global experimental dataset, we were able to identify favourable and unfavourable climate conditions and geographical regions for the implementation of CA systems for eight major staple crops under current and future climate conditions. Some of the most promising geographical regions in our analysis had also been identified in previous studies 19 but we were able to report information on yield gains as probabilities instead of simpler increase or decrease categories. More importantly, we addressed the impacts of future climate change scenarios and different agricultural management practices on the performance of CA systems and their geographical patterns. However, there are also limitations in our study. Due to the lack of global quantitative data on crop irrigation, this study was taken into account in our models as a categorical variable. Although a large part of the data considered in this study was collected under humid climate conditions, a substantial amount of data was obtained in dry regions under rainfed or irrigated conditions. In this study, climate and soil data were extracted from external databases and not from the collected articles because these data were missing or not reported consistently in the individual articles. Consequently, these data might not always match experimental records. The use of external databases has the added benefit of helping us understand how inter-and intra-annual climate variability affects the productivity of CA systems and their variants versus CT. It is also required to predict the performance of these systems under future climate conditions. Data availability may affect model accuracy. Due to the limited number of observations for these crops, the conclusions obtained for rice and sunflower are more uncertain compared to other crops (Supplementary Fig. 7). Despite these limitations, the results of our model assessment via the cross-validation are overall satisfactory (Methods and Supplementary Fig. 7).
We showed that soil cover has a strong positive effect on probability of yield gain. Without soil cover, NT systems are likely to lead to a yield loss compared to CT ( Supplementary Fig. 2a3-h3). Soil cover reduces soil evaporation and surface runoff and maintains a high level of soil moisture content 20-28 , thus increasing the competitiveness of CA systems especially for dryer climate conditions [29][30][31][32][33][34] . Therefore, keeping the soil covered by crop residues appears to be an important factor for the success of CA systems. However, in practice, maintaining crop residues might be challenging in some regions, such as Africa, where the crop residues are used to feed livestock 35 . In such situations, a possible solution would be to rely on alternative sources of plant materials, such as residues from cover crops, grass, leaf litter from trees, sawdust and so on 35 . Without soil cover, NT has low chance to entail yield gains compared to CT. Under future climate scenarios, the performance of NT is expected to improve in the north of sub-Saharan Africa for maize (Fig. 5c) and sorghum (Supplementary Fig. 4e1) but the probability will remain lower than or close to 0.5.
Although less influential, other farming practices appear to increase the probability of yield gain of CA, in particular fertilization and weed and pest control. Thus, for most crops, CA with field fertilization and weed and integrated pest control outperformed CT in continental, dry and temperate regions but proved less suitable in tropical regions ( Supplementary Fig. 1a1-h1 and Supplementary  Fig. 2a1-h1). This overall pattern is in line with previous work 6 and our results are also consistent with the study of Corbeels et al. 36 who showed that higher productive performance of CA systems in Africa can be expected when CA principles are implemented concomitantly in combination with herbicide application, especially for maize ( Supplementary Fig. 2c1). However, we also need to note that part of the regional variability in the performance of CA and NT might be related to the diversity of farm characteristics in the different regions under consideration, in particular the level of mechanization and size of farm. In addition to increasing crop yield, CA systems can potentially improve biodiversity, increase the soil organic matter and bring about positive environmental externalities such as reduced soil erosion, improved soil quality and enhanced carbon sequestration 2,3,37 . Moreover, CA could improve the resilience of cropping systems to climate change and increase the stability of crop yields 38,39 . Although several variants of CA systems may be associated with a high probability of yield loss in many regions, we also showed that, under future climate conditions and with good agricultural management practices, the relative productive performance of CA is expected to increase for several crop species. This is especially true for maize in tropical regions, which further strengthens the competitiveness of CA for this staple crop. Thus, our results support the idea that CA will be a relevant option for cropping systems in the future, capable of ensuring a long-term, sustainable agricultural production for some key cropping areas 40,41 .

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/ s41558-021-01075-w.

Methods
Data collection. The literature search was done in February 2020 using the following keywords 'conservation agriculture/ no-till/ no tillage/ zero tillage' and 'yield/ yield change' in the websites ScienceDirect and Science Citation Index (Web of Science). We also collected the papers cited in previous meta-analyses 5,6,8 , supplementing them by the most recently published experimental studies. A total of 1,012 potentially relevant papers were identified by reviewing the title and abstract; these papers were then screened according to the procedure summarized in Supplementary Fig. 6. More details about this screening and selection procedure were presented in previous studies 11,42 . In the end, 422 papers were retained (published between 1983 to 2020). From these papers, we were able to extract 4,403 paired yield observations from NT and CT for eight major crop species, including 370 observations for barley (232 observations for spring barley and 138 for winter barley), 94 observations for cotton, 1,690 observations for maize, 195 observations for rice, 160 observations for sorghum, 583 observations for soybean, 61 observations for sunflower and 1,250 observations for wheat (1,041 observations for winter wheat and 209 observations for spring wheat) in 50 countries from 1980 to 2017 (Extended Data Fig. 4). We also retrieved from the papers information on crop type, year and location of the experiments and agricultural management activities for both NT and CT systems, including crop irrigation (yes versus no), the field fertilization (yes versus no) and the details of the type and amount of fertilizer used, integrated weed and pest control (yes versus no) and the type of herbicide and pesticide used, crop rotation (yes versus no) and details of crop sequence and information of cover crops, soil cover (yes versus no) and the details of residues retention from previous crops or cover crops. On the basis of this information, we were able to define CA as the combination of NT with soil cover and rotation and to distinguish this system from other variants (NT without soil cover and rotation, NT with rotation but without soil cover, NT with soil cover but without rotation). Additional data were extracted from several external databases, pertaining to crop growing season 16 , soil texture 43 and climate factors such as precipitation (P) 12 , minimum air temperature (T min ) 13 , average air temperature (T ave ) 12 , maximum air temperature (T max ) 13 and potential evapotranspiration (E) 14,15 in the growing season 16 in the particular year of the experiment and the precipitation balance (PB) was defined as precipitation minus total evapotranspiration, which indicated the water availability in the growing season. These data were obtained at a spatial resolution of 0.5° latitude × 0.5° longitude and, if the source data were in a finer spatial resolution, they were downscaled to the resolution of 0.5° latitude × 0.5° longitude.

Model training and cross-validation.
The machine-learning algorithm random forest 17 was trained to analyse the yield ratios of NT versus CT as the function of climatic variables, crop types, soil textures and agricultural management activities. The climatic variables during the growing season such as PB, T min , T ave , T max were defined as numerical inputs, while crop type, soil texture and agricultural management activities including crop irrigation, field fertilization, integrated weed and pest control, crop rotation and soil cover management were defined as categorical inputs. The model output was expressed as the probability of yield gain from NT versus CT. When training, each tree in a random forest learns from a random sample of the data points. The samples are drawn with replacement (bootstrapping). Only a subset of all the inputs is considered for splitting each node in each decision tree. Predictions are made by averaging the predictions of all decision trees 44 . The performance of the algorithm was assessed by estimating the area under the ROC curve by leave-one-out cross-validation. The procedure and results of cross-validation are presented in detail in Supplementary Fig. 7 and in ref. 42 . Since crop rotation, soil cover management and other agricultural management practices were included as model inputs, it is possible to map the probability of yield gain for CA (NT with crop rotation and soil cover) and several variants of this system (NT partly or not implemented with crop rotation and soil cover) versus CT. Maps were generated for all crop species at a spatial resolution of 0.5° latitude × 0.5° longitude.
Global projection. The fitted random forest model was used to estimate the probability of yield gain from CA system and their variants versus CT without crop rotation (-R) and without soil cover (-SC) for each grid cell located in cropping regions under current (2011-2020) and future (2051-2060) climate scenarios. This variant of CT was chosen as a baseline comparator because it prevails in the training dataset 10,11 . The monthly average values of the climatic variables (PB, T min , T ave , T max ) were calculated in each grid cell over the two time periods considered and then these data were used to calculate the climatic variables during the growing season on the basis of the crop calendar database 16 (assume no change in current and future scenario). All the climatic data in both current and future scenarios were obtained from four climate models: Gfdl-esm2m, Hadgem2-es, Ipsl-cm5a-lr and Miroc5 and four RCP scenarios: RCP 2.6, RCP 4.5, RCP 6.0 and RCP 8.5. This results in 32 combinations: four climate models × four RCP scenarios × two periods. We mainly focused on the Ipsl-cm5a-lr model and RCP 4.5 scenario in the baseline simulations because of their importance and similar role in the protocol of ISIMIP2b project 45 . However, results from all combinations were analysed and shown in Extended Data Fig. 3. All the climatic data can be downloaded through the website of Lawrence Livermore National Laboratory 46 .
We did not change the categorical inputs describing cropping practices between current and future scenarios. The global soil texture was set on the basis of the HWSD dataset 43 . To compare the performance across different cropping systems, we mapped the probability of yield gain and the change of yield gain probability under climate change with the systems of CA (+F+WD) versus CT-R-SC (+F+WD), CA (-F-WD) versus CT-R-SC (-F-WD), NT+R-SC (+F+WD) versus CT-R-SC (+F+WD), NT-R-SC (+F+WD) versus CT-R-SC (+F+WD), NT-R-SC (-F-WD) versus CT-R-SC (-F-WD), where +/-R indicated crop rotation set as yes/ no, +/-SC indicated soil cover set as yes/ no, +/-F indicated fertilization set as yes/ no and +/-WD indicated weed and pest control set as yes/ no. As for crop irrigation, it was set on the basis of the crop irrigation mask from MIRCA2000 dataset 47 . When more than 50% of the area in a grid cell was under rainfed conditions for a given crop in the MIRCA2000 database, this cell was then considered as non-irrigated for this crop and vice versa. See Supplementary Table 1 for the details of model settings. The model outputs were mapped at a spatial resolution of 0.5° latitide × 0.5° longitude on the basis of the MIRCA2000 crop mask database 47 . Accumulated area fractions under different levels of yield gain probability and different levels of probability change between current and future scenarios were computed at the global scale and in different climate regions.
Climate regions. The term 'global' indicated the global cropping region for each crop. Climate was classified according to the Köppen-Geiger classification 18 and its nomenclature (Supplementary Fig. 8). The term 'tropical climate' included the regions with the climate types Af, Am, As and Aw (ref. 18 ). The term 'dry climate' included the regions with the climate types BWk, BWh, BSk and BSh (ref. 18 ). The term 'temperate climate' included the regions with the climate types Cfa, Cfb, Cfc, Csa, Csb, Csc, Cwa, Cwb and Cwc (ref. 18 ). The term 'continental climate' included the regions with the climate types Dfa, Dfb, Dfc, Dfd, Dsa, Dsb, Dsc, Dsd, Dwa, Dwb, Dwc and Dwd (ref. 18 ).
Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The dataset 10 is made fully available and described in a data paper 11 . The dataset can be accessed using the following link: https://doi.org/10.6084/m9.figshare.12155553.

Code availability
All the R and MATLAB codes are available in figshare and can be accessed through the link https://doi.org/10.6084/m9.figshare.14427977 or on request from the corresponding author.

Data
Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: -Accession codes, unique identifiers, or web links for publicly available datasets -A list of figures that have associated raw data -A description of any restrictions on data availability The code and dataset used in this study are made fully available online. Field-specific reporting Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Ecological, evolutionary & environmental sciences study design
All studies must disclose on these points even when the disclosure is negative.

Study description
The data come from the systematic review of crop yield in conservation agriculture vs. conventional tillage system, the details is described in our data paper, Su

Sampling strategy
The procedure to data screening was presented in detail in our data paper, Su, Y., Gabrielle, B. & Makowski, D. A global dataset for crop production under conventional tillage and no tillage systems. Sci Data 8, 33 (2021). Papers not reporting yield data for CT and NT systems were excluded, as well as papers reporting experiments on reduced tillage (RT) systems. Papers reporting only mean yield data across different years or sites were also excluded. We then checked whether information on fertilization, weed and pest control, crop irrigation, crop rotation and crop residue management were reported for both CT and NT practices.

Data collection
The data collection procedure was presented in detail in our data paper, Su The procedure to data exclusion was presented in detail in the data paper, Su, Y., Gabrielle, B. & Makowski, D. A global dataset for crop production under conventional tillage and no tillage systems. Sci Data 8, 33 (2021). When the experiment's climate and soil data are not available in the external datasets, it was removed from the dataset.