A systematic map of cassava farming practices and their agricultural and environmental impacts using new ontologies: Agri ontologies 1.0 ‐

1. Cassava is consumed by 800 million people and is a staple crop in Africa. Its pro - duction may increase under climate change due


| A systematic map of cassava farming practices and their impacts
Cassava (Manihot esculenta) is a major global crop, with an increasing annual production that totalled >302 million tonnes in 2020 (FAO, 2019).Cassava (also known as mandioca, manioc and yuca) is a perennial plant primarily grown for its starchy tubers, with the majority used for human consumption-either eaten directly or converted into flour (e.g.tapioca, farinha or garri)-and the remainder used as animal feed or biofuel (FAO, 2000).It is primarily grown via subsistence farming in the tropics and subtropics, with Nigeria and Brazil leading production (see Figure 5 for a global yield and production map).Approximately 800 million people consume it, and in Africa, it is the second most important crop after maize, with >40% of the population using it as a staple crop (FAO, 2019).
The importance of cassava as a food crop may increase further as it has the potential to help to mitigate the negative impacts of climate change on food security (Jarvis et al., 2012).For example, climate change is increasing the frequency and severity of droughts (Malhi et al., 2021), with drought-affected areas projected to increase from 15% to 44% by 2100, and the yield of major crops predicted to decrease by >50% by 2050 (Li et al., 2009).Sub-Saharan Africa, which produces >50% of cassava globally (FAO, 2019), is the most vulnerable region to drought (Li et al., 2009).Furthermore, its population is projected to grow by 1 billion people by 2050, which will further increase the pressure on food systems (United Nations Department of Economic and Social Affairs Population Division, 2019).It is imperative that there is investment in crops, such as cassava, that are more resilient to drought and climate change (Jarvis et al., 2012).Cassava's resilience extends beyond drought tolerance too; it can be grown in soils with low fertility and harvested at any time, hence its nickname: 'the drought, war and famine crop' (Burns et al., 2010).Therefore, cassava has the potential to help to mitigate the impacts of climate change on food security.
Given its current and potential future role, it is important that cassava management is evidence based (Sutherland et al., 2004).This will help to reduce yield gaps, which are currently large (yields of the top-producing cassava countries [18 t/ha] are <50% of potential yields [44 t/ha]; Figure 5).Furthermore, some existing farming practices contribute to environmental degradation, such as depleting soil fertility (Reynolds et al., 2015).If cassava production is to be scaled up, evidence-based farming practices will be crucial for doing this as sustainably as possible whilst maximising yields (Sutherland et al., 2004).Cassava management is varied (e.g.intercropping, fertiliser and rotation practices are diverse), and there are thousands of studies testing different management practices.Several international authorities on cassava farming, such as the International Center for Tropical Agriculture (CIAT), the International Institute of Tropical Agriculture (IITA) and the Food and Agriculture Organisation (FAO) of the United Nations, have made evidence-based agricultural extension manuals and work closely with practitioners to disseminate findings (e.g.FAO, 2013;Howeler & Aye, 2017;International Institute of Tropical Agriculture (IITA), 2017).However, there is no systematic map of cassava research (Shackelford et al., 2018).Systematic mapping collates, describes and catalogues evidence (e.g.publications or reports) for a specific topic (James et al., 2016).Systematic maps enable researchers, funders, practitioners or policy makers to identify knowledge clusters or gaps and then direct resources accordingly.For example, researchers can conduct syntheses on knowledge clusters and funders can fund field studies on knowledge gaps (James et al., 2016).
Systematic maps also direct users to relevant studies, which can improve the efficiency and completeness of evidence gathering (James et al., 2016).Here, we create a global systematic map of cassava farming practices to provide an evidence base for sustainable cassava production.
4. This systematic map of cassava farming practices can direct researchers and funders to knowledge gaps that need addressing, and reviewers to knowledge clusters for synthesis.Better research practices should be promoted within cassava research, as poor reporting standards, short study durations and long publication delays result in an ineffective research environment.This systematic map provides an evidence base for cassava production and the ontologies (Agriontologies 1.0) can be applied to other systems to facilitate more efficient and effective synthesis.
agricultural taxonomy, cassava Manihot esculenta, evidence-based management, interactive evidence map, reporting standards, standardised classification system, subject-wide evidence synthesis, sustainable agriculture, systematic map, terminological ontology agriculture

| Ontologies for farming interventions and outcomes
The FAIR guiding principles for scientific data management were described in 2016 (Wilkinson et al., 2016) to address the scientific community's growing concern that poor data management was inhibiting data reuse and contributing to a 'significant crisis of reproducibility' (Baker, 2016).These principles outlined four characteristics for data management-that datasets should be 'Findable', 'Accessible', 'Interoperable' and 'Reusable' (Wilkinson et al., 2016).Significant efforts have been made to achieve this, including the development of dataset repositories and citation metrics (Cousijn et al., 2018;Hood & Sutherland, 2021;Konkiel, 2020), journals and funders mandating public data archiving (Mislan et al., 2016) and increased uptake of reporting guidelines or standardised terms for study metadata (Equator Network, 2022;Stevens et al., 2014).Standardised terms can improve reporting standards and address the 'Interoperable' and 'Reusable' aspects of FAIR data management by making it easier to combine datasets (Hopewell et al., 2012;Plint et al., 2006).This facilitates crosscomparison between fields, promotes research efficiency, and enables data reuse for transfer learning, generating synthetic data and synthesis (Todman et al., 2023).Standardised terms can be ordered into multi-level ontologies to show relationships between terms and enable data to be grouped at different levels (Jonquet et al., 2018;Thessen et al., 2015).
There are several ontologies relevant to agricultural and ecological research (Jonquet et al., 2018), including ICASA Data Standards for data from agricultural field experiments (Hunt et al., 2006;White et al., 2013), DarwinCore for taxonomic data (Wieczorek et al., 2012) and AgrO for agronomic data (Aubert et al., 2017).
Here, we use this systematic map of cassava to develop hierarchical ontologies for [1] farming interventions and [2] agricultural, economic and environmental outcomes.These ontologies differ from existing ontologies as they have a broad focus and are intended to be comprehensive while also being general enough to apply to other agricultural systems.
They are also designed to be facilitate quantitative syntheses, such as meta-analyses, by using 'treatment-treatment' and 'treatment-control' categories, and they apply the use of 'not elsewhere classified (n.e.c.)' categories to identify categories that should be added, i.e. if the 'n.e.c.' category has many publications coded in it.In particular, these ontologies were designed to develop case studies for 'dynamic meta-analysis', which is a method of meta-analysis that enables synthesis at multiple levels via an online interactive interface (Shackelford et al., 2020), with the ultimate aim of increasing the relevance and recency of evidence (e.g. via 'living systematic reviews') and facilitating subject-wide evidence syntheses (Martin et al., 2022;Sutherland & Wordley, 2018).For example, standardised ontologies such as this could eventually contribute to a global map of scientific literature on agriculture.

| Objectives of the review
The primary objective of this review was to produce a systematic map of scientific studies about cassava farming practices, with the aim of identifying knowledge gaps and clusters.Our secondary objective was to develop a classification system for [1] farming interventions and [2] agricultural, economic and environmental outcomes.These objectives reflect those in our published protocol (Shackelford et al., 2018), with the exception of primary objective five, which relates to the study designs used and was added during the review.Study design is important for interpreting the quality of the mapped studies and the evidence base as a whole.For example, synthesists that have used the following map to identify knowledge clusters will be able to assess whether their chosen topic has a robust study design (e.g.long study duration and randomised treatments) and may choose to select a different topic if not (Christie et al., 2020).We also expanded objective 1; in the protocol, it was limited to cassava fields, and during the review, we expanded it to whole cassava plants (i.e.we included whole plants in glasshouses).

| Primary objectives
The primary objectives were to answer the following questions:

| Secondary objectives
The secondary objective was to develop and test-using this systematic map of cassava as an example-ontologies of [1] farming interventions (e.g.intercropping or tilling) and [2] agricultural, economic and environmental outcomes (e.g.yield or soil).These ontologies are hierarchical classifications of farming practices and agri-environmental outcomes and are intended to be general enough to be reused for other crops.

| The systematic map
The following systematic map was produced according to the corresponding published systematic map protocol (Shackelford et al., 2018).
This protocol was based on the Collaboration for Environmental Evidence (CEE) guidelines for systematic mapping (James et al., 2016) and the RepOrting standards for Systematic Evidence Synthesis (ROSES) reporting checklist (Supporting Information 1; Haddaway et al., 2018).The published protocol was developed in consultation with our three project partners: the African Cassava Agronomy Initiative (ACAI; Pypers, 2017), the Conservation Evidence group at the University of Cambridge (Sutherland et al., 2021), and the Leventis Foundation, which runs agricultural schools in West Africa (The Leventis Foundation Nigeria [Internet], n.d.).The Leventis Foundation funded this map and uses it to identify knowledge gaps to study in field trials.The Conservation Evidence group uses the map to identify knowledge clusters and conduct syntheses.

| Searches
We searched for studies using the search string 'cassava OR mandioca OR manihot OR manioc OR yuca'.These are the common synonyms for cassava, but not its products (e.g.tapioca), since we were interested in pre-harvest management of cassava (Gade et al., 2002;Hillocks et al., 2002).As the aim of our systematic map was to map

| Article screening
The records returned from these searches were screened for duplicates in Endnote X8.We screened the remaining records at [1] title and abstract stage and [2] full text stage.The screening process was procedurally independent as the authors that screened the records had not authored articles that would be relevant to include in the map.Report 2010'), we included them.

| Populations/subjects
We included studies that measured the impact of cassava farming practices on agricultural or environmental outcomes.We included studies that looked at other crops in rotation with cassava or intercropped with cassava.We included studies on whole cassava plants in pots, but not modelling studies, studies of cassava pests or natural enemies where cassava was absent or studies on cassava cuttings or germplasm (e.g.laboratory studies of cassava in vitro).We included studies of cassava fields before or after cassava was planted (e.g.prepping the field or cover cropping post-harvest) but not of postharvest management of cassava (e.g.storing).For studies comparing cassava with other habitat types, studies about the conversion of a habitat to cassava agriculture were included, but studies that compared existing cassava fields with other habitat types without cassava were excluded.We did not restrict the timing of studies and the scope was global.

| Comparators
We included studies that compared a population with an intervention with a population without an intervention.This included controlled studies (e.g.comparison of a plot with treatment applied [treatment] and a plot without treatment applied [control]), correlated studies (e.g.comparison of a plot with pre-existing application of a treatment [treatment] and a plot without treatment [control]) and before-and-after studies (e.g.comparison of a plot with itself before [control] and after treatment is applied [treatment]).We included 'treatment-treatment' comparisons (e.g. a comparison of a plot with one fertiliser type applied [control] vs. another plot with a different fertiliser type applied [treatment]).We excluded studies that had no comparators.

| Screening at full-text stage
The publications were included past the title and abstract screening stage, then assessed at the full-text stage and were excluded if 1.A full text of the study could not be found.We searched the publication title in Google, Google Scholar and Web of Science with and without quotations.

2.
A full text of the study was behind a paywall according to the journal subscriptions at the University of Cambridge in 2017, where the study was conducted.
3. The full text was not available in English.
4. The study was secondary literature (i.e.studies that cite other studies as the source of numerical results).
5. The study had no relevant PICO components.We only considered outcomes that were reported via numerical results in figures or tables.
6.The study was a duplicate.In cases where there was a larger book with smaller chapters, we included each as smaller chapters, unless the entire book was written by the same authors.
7. There was no full text (e.g. the study was an audio file or poster).

The study was retracted.
Note that we assessed these exclusion criteria sequentially, sofor example-if a publication was excluded because the full text was not available in English (criteria 3), we did not assess whether it was secondary literature (criteria 4).

| Consistency in article screening
All publications were screened by one person (GES), but a second person (ASCH) screened 10% of publications, selected at random, to check the consistency of the screening process (n = 2038 at title and abstract stage and n = 502 at full-text stage).We discussed any disagreements and clarified the eligibility criteria where necessary.We calculated Kappa scores as a measure of agreement (Cohen, 1960).
Kappa ranges from −1 to 1, with Kappa >0.6 considered as 'moderate' agreement (Cohen, 1960).We used 0.6 as a threshold to indicate agreement.

| Data coding strategy
We coded the interventions (described in 'Agri-ontologies 1.0' below), outcomes (described below), study location(s) (country), study year(s) and experimental design.Experimental design included the following categories, with the option to select multiple categories: before-and-after, blocked (i.e. a controlled study with treatments next to each other spatially), controlled, correlated (sitecomparison), paired (i.e. a correlated study with sites that are selected to have similar characteristics), randomised and replicated.
We only coded outcomes that were reported via numerical results in figures or tables.These were coded using an online web app with drop-down menus (www.metadataset.com).We did not critically appraise the validity of these studies as the map can be used for multiple methods of evidence synthesis which may have different criteria for critical appraisal.We did not contact the authors for any missing or unclear information.Many interventions that we coded were confounded (e.g. for a study that varied fertiliser application in tandem with varying cultivar, we would code this as two single interventions related to 'fertiliser' and 'cultivar').We did not code publications with multiple studies into separate entries, but if the same intervention was tested multiple times in different experiments within one publication, we coded this intervention multiple times within the publication.If the same outcome was tested multiple times within one publication, we coded it once.One author (GES) coded the majority of publications (983/1599).Then we assessed consistency in coding using Kappa scores (described in 'Agri-ontologies 1.0' below).
Once agreement was reached, the second author (ASCH) coded the remaining publications (616/1599).

| Deviations from the protocol
This systematic map deviated from the protocol (Shackelford et al., 2018) in three respects: 1. We expanded the scope to include coding 'experimental design'.
2. We expanded the scope to include studies on whole cassava plants (e.g.cassava plants in pots, but not leaves in petri dishes), where the original protocol specified that only studies in fields would be included.
3. We did not code outcomes, years or experimental design for the intervention '10.10.10.TT.20.Planting a different variety/cultivar (e.g. a disease-resistant variety)' (i.e.we only coded the intervention).This was due to time constraints.This intervention was the most common (939 intervention occurrences, 24% of intervention occurrences), and it would be difficult to apply a synthesis to this intervention as there are numerous cultivars, most of which are regional.For publications that tested this intervention and others, we coded the outcomes, years and experimental designs of the other interventions.

| Limitations of the method
The following systematic map of cassava is subject to four limitations: 1. We only included publications written in English, which can bias syntheses (Konno et al., 2020).However, in this case, only 1% of full texts were excluded for this reason.We did not have the resources to work in other languages, and the stakeholders involved were working in countries where English is an official language (Nigeria, Ghana, Kenya and the United Kingdom).
2. We excluded publications that were behind a paywall according to journal subscriptions at the University of Cambridge where the study was done.This was 6% of full texts.
3. We excluded publications that we could not find as full texts, which was 24% of records.In many cases, we suspect this was because the full texts do not exist (e.g. the record refers to an unpublished conference paper or a magazine article without an online archive).
4. The searches were conducted at the end of December 2017, so the map is already out of date.This delay was partly due to setbacks caused by the coronavirus pandemic.Nevertheless, this map forms a solid foundation for future updates.

| Data mapping method and analysis
This systematic map of cassava farming practices is freely available online via an interactive format (an 'evidence atlas') on https:// www.metad ataset.com/(registration required) and in spreadsheet form on a research data archive (Hood et al., 2022).Users can select interventions or outcomes at multiple levels and view the relevant publications and their study location (country level).In this paper, we summarise these findings according to this study's objectives outlined above.Data were wrangled, plotted and analysed visually using R version 4.1.1(R Core Team, 2020) with R studio version 1.4.1717(RStudio Team, 2019), using the Tidyverse (Wickham, 2017), rworldmap (South, 2011), simple FeaTures (Pebesma, 2018) and cow-ploT (Wilke, 2019) packages.We used local regression smoothers (LOESS) to visually interpret some temporal trends.Several plots were amalgamated in Inkscape (Inkscape, 2020).

| Agri-ontologies 1.0-Ontologies for farming interventions and outcomes
The intervention and outcome ontologies were developed based on previous work and in consultation with the same three stakeholders that designed the systematic mapping protocol (Shackelford et al., 2018).The ontologies (Figures 1 and 2, listed in full in Supporting Information 3 and 4) are freely available online via an interactive format at https://www.metadataset.com/.These ontologies were designed to be generic enough to apply to other crops and are particularly intended to be used for syntheses, such as systematic maps, reviews or meta-analyses, though they can be used for other purposes too.Both ontologies used 'not elsewhere classified intervention or outcome is missing.We revised and clarified these ontologies as we mapped papers, and they should continue to be revised as necessary.

| Agri-ontologies 1.0 intervention ontology
The Agri-ontologies 1.0 intervention ontology included all in-field/ on-farm management practices, such as applying amendments, planting in a different agro-ecological zone, irrigating, intercropping, planting crop margins or releasing biological control agents (Figure 1).We included four combined interventions: integrated pest management, integrated soil management, organic farming and precision farming.Many interventions were broken into 'treatmentcontrol' or 'treatment-treatment' comparisons at the lowest level (e.g. a study that compared plots with and without fertiliser would be 'treatment-control', whereas a study that compared two fertiliser types would be 'treatment-treatment').This distinction was to facilitate reuse of the systematic map for synthesis, as synthesists may wish to focus on one or other type of comparison with conduct a quantitative synthesis.

| Agri-ontologies 1.0 outcome ontology
The Agri-ontologies 1.0 outcome ontology included agricultural or environmental outcomes, such as crop yield, crop quality, soil, water, wildlife, pathogens, pests, pollutants, profits, labour and time (Figure 2).We did not include outcomes on livestock, farmer behaviour or physiology that were not directly relevant to quality or molecular biology (e.g.enzyme activities, molecular markers, gene expression or genetic diversity).

| Consistency in data coding
To check the consistency of coding, two authors (GES and ASCH) coded 100 publications selected randomly.We calculated Kappa scores to test the agreement for coding interventions and outcomes.
We discussed any disagreements and revised or clarified the interventions and outcomes ontologies where necessary.As with the article screening, we used a Kappa score of 0.6 as a threshold to signify agreement (Cohen, 1960).

| Searches and screening
Following the searches 36,580 records were identified, 20,380 remained after duplicate removal in Endnote X8, 5020 were included after title/abstract screening, 3515 were retrieved at full-text stage, and 1599 were included after full-text screening (Figure 3, see Supporting Information 2 for included and excluded publications).

F I G U R E 2
The structure of the 'Outcomes' ontology, with a sample section expanded to show the lower (more specific) levels of the ontology.
HOOD et al.
The threshold to indicate agreement in screening was exceeded following the initial 10% at title and abstract and at full-text stage, so we did not screen additional texts (Title and abstract: Kappa 0.69 and 89% agreement, Full text: Kappa 0.76 and 91% agreement).
Considering the global importance of cassava (e.g. in Africa, it is the second most important crop after maize; FAO, 2019), this is relatively few studies.For example, a Web of Science search for publications prior to December 2017 related to maize (search terms: maize OR 'Zea mays' OR corn) and wheat (search terms: wheat OR triticum) yielded 192,745 and 184,662 records respectively, which is more than nine times the records yielded by our search for cassava (20,380 records), which included multiple search engines and synonyms.Cassava research should be prioritised to increase the evidence base for this globally important crop.

| Regional distribution
Of the 1599 included publications, we mapped the country-of-study of 998; 76 were excluded as they did not report the country and 525 were excluded as they only tested the intervention 'Planting a different variety/cultivar (e.g. a disease-resistant variety)' (see Section 2).
These results show that research is dominated by a few countries.This may be partly due to the success of large and successful organisations that study cassava, such as CIAT in Colombia and IITA in Nigeria.The number of publications is also partly reflected in production, as six of the seven top producers (exceeding 1000 t/ year) are in the top 10 countries for research: Nigeria (5798 t/ year), Thailand (3022 t/year), Ghana (1977 t/year), Brazil (1864 t/ year), Indonesia (1802 t/year) and Vietnam (1034 t/year; Figure 5).
However, half of these countries have under 50 relevant publications, and the remaining top producer (the Democratic Republic of Congo: 3863 t/year) has 11 publications (Figure 5).Therefore, many top-producing countries are understudied, and future research should target these regions (Figure 5).This is particularly important in the context of yield gaps; the average yields of the top seven producing countries (18 t/ha) are less than half of potential yields (44 t/ha; Figure 5).Further research in regions with higher yields may help to inform how to close these gaps and broaden the relevance of cassava research.publications on vertebrate control, despite there being a number of vertebrate pests in cassava systems (Cudjoe, 1994).Chemical control (n = 85) has been studied more than biological (n = 39) or physical control methods (n = 43, e.g.weeding), and no publications have F I G U R E 5 World maps using Goode's Homolosine projection, which accurately shows country sizes.Latitude and longitude are shown in grey lines and countries are outlined in black.Brown colour shows (a) the number of publications (n = 1004), (b) average cassava yield (t/ha) and (c) total cassava production (t/year).Data from (b) and (c) are country means from 2016 to 2020 according to FAO (FAO, 2019).In (a) there were four countries in the highest bracket, with three outliers; the black numbers show the number of publications for these countries.F I G U R E 7 A scheme showing the relationship between the highest (least specific) and middle levels of the intervention ontology (see Supporting Information 3 for the full ontology) and the number of publications for each intervention (according to the area of the black points).Some of these intervention terms have been abbreviated (see Supporting Information 3 for the full terms and explanations). of mid-level outcome occurrences).Few publications measured environmental outcomes, with only 24 publications that studied 'Pollutants' and 55 that studied 'Wildlife' (combined 5% of higherlevel outcome occurrences) (Figure 8).

| Interventions and outcomes
Taken together, these results show that there are major knowledge gaps, particularly related to environmental management and outcomes.The majority of research has focussed on production, with minimal research into non-crop habitat, wildlife or pollutants.
It is imperative that this knowledge gap is addressed so cassava production and expansion can be managed as sustainably as possible, as evidence-based management can reduce negative impacts on biodiversity (Sutherland et al., 2004;Walsh et al., 2015).This is likely to be important for maintaining yield stability in cassava too by providing resilience to yield fluctuations, as publications in other systems have shown that biodiversity can achieve this (Dardonville et al., 2022;Haughey et al., 2018).We also identify knowledge clusters, such as interventions related to cultivar selection, harvesting, fertiliser application, intercropping and crop rotations, which can be used for syntheses.Therefore, synthesists should conduct syntheses in knowledge clusters, and funders, stakeholders and researchers should target resources towards improving the knowledge base around environmental management and outcomes in cassava systems.

| Study design
We included four aspects of study design to assess the quality of mapped studies and evidence base as a whole: reporting standards (reporting completeness), experimental design, study duration and publication delay.We tracked reporting standards for country and experimental dates; 24% (n = 384) of publications did not report experimental start dates and 7% (n = 112) did not report country-ofstudy.These are basic experimental details that should be reported in every study so that study context can be properly understood and to facilitate reproducibility and data reuse.We did not track reporting standards for experimental design because having a study comparator was a criterion for publication inclusion during mapping.

F I G U R E 8
A scheme showing the relationship between the highest (least specific) and middle levels of the outcome ontology (see Supporting Information 4 for the full ontology) and the number of publications for each outcome (according to the area of the black points).Some of these outcome terms have been abbreviated (see Supporting Information 4 for the full terms and explanations).Some experimental designs are more biased than others (Christie et al., 2020).Generally, designs that incorporate more of the features listed in Table 1 are considered more robust than simpler designs, and experimental randomised controlled studies and observational before-after controlled or paired/blocked controlled studies are considered to be more robust than simpler controlled or before-after designs, as well as correlational studies (Christie et al., 2020).However, there are other aspects of experimental design (e.g.statistical matching, sample size and plot size) that affect study bias and robustness (Christie et al., 2020).Furthermore, some designs cannot be applied to certain questions (e.g. a study comparing primary forest to another ecosystem cannot be controlled because you cannot create a new primary forest).For publications included in our systematic map, the most common design was a randomised blocked design (n = 1069, 36.23% of experimental design occurrences; Table 1), which is very robust (Christie et al., 2020).
Study duration and publication delay was recorded for 821 publications, as 253 did not report start dates and 525 were not tracked as they only tested cassava cultivars (see Section 2).The study duration was 2.0 ± 0.1 years (mean ± standard error).It has also been decreasing, with visual inspection using LOESS smoothers showing estimated values of 2.7 years in 2000 and 1.3 years in 2017 (Figure 9a).Short study durations are partly due to funding landscapes, where projects are often funded over short periods and researchers are required to demonstrate impact within this time (Grove, 2018).This hinders research, as long-term studies contribute disproportionately both to scientific understanding and to policy in ecology (Hughes et al., 2017;Mills et al., 2015).Long publication delays are also problematic, as they reduce researcher, practitioner and funder access to the latest and most relevant research, which creates an inefficient research environment (Christie et al., 2021).Here, we show that publication delay is 4.0 ± 0.1 years (mean ± standard error), which is longer than delays in the related field of conservation science (3.2 ± 0.1) (Christie et al., 2021).Visual inspection of publication delay through time using LOESS smoothers showed that this has been relatively constant, with some reduction in the last 15 years (approximately −0.7-year delay; Figure 9b).
Based on our findings, we recommend that researchers and other stakeholders (e.g.journals and funders) promote the use of reporting checklists or guidelines by authors, reviewers and editors, to improve reporting standards (Hopewell et al., 2012;Plint et al., 2006;Stevens et al., 2014;Turner et al., 2012).Long-term projects should also be prioritised to increase study duration, as should more ro- a Before-after-control-impact.
b Randomised before-after-control-impact.
c Randomised controlled trial or randomised control-impact.
d Randomised blocked design, including randomised complete blocked design.author guidelines, providing assistance to authors writing in a second language, incentivising peer review, allowing submission to multiple journals, or using pre-print servers; Christie et al., 2021).These results can also be used to direct future resources.For example, after synthesists have identified knowledge clusters using the inter-  Like other ontologies, Agri-ontologies 1.0 can improve reporting standards and increase the interoperability of data (Hopewell et al., 2012;Plint et al., 2006), therefore promoting data reuse for uses such as transfer learning, generating synthetic data and syntheses (e.g.meta-analysis; Todman et al., 2023).Hierarchical classification systems such as this enable the automatic lumping and splitting of evidence, using methods such as 'dynamic meta-analysis' (Martin et al., 2022;Shackelford et al., 2021).Agri-ontologies 1.0 complements existing ontologies, such as DarwinCore (Wieczorek et al., 2012) and AgrO (Aubert et al., 2017).It is unique by having a broad focus, and being comprehensive whilst also being general enough to apply to other agricultural systems.It applies the use of 'not elsewhere classified' categories to identify common categories that are missing and should be added.In addition, it facilitates quantitative syntheses, such as meta-analyses, by using 'treatmenttreatment' and 'treatment-control' categories.
Agri-ontologies 1.0 represents an important step towards 'subject-wide evidence synthesis', as developing subject-wide classification systems facilitates comparison between subjects (Sutherland & Wordley, 2018).For example, if Agri-ontologies 1.0 was used to systematically map the literature on other crops, then these maps would be interoperable, and this could eventually lead to a global map of scientific literature on agriculture.

| CON CLUS IONS
Here, we provide two advancements in the field of evidence synthesis.We provide the first example of an evidence atlas for an entire crop and we develop hierarchical classifications of farming interventions and outcomes tailored to evidence synthesis.The results of our searches indicated that there has been little research on cassava compared with other major global crops: maize and wheat.
Furthermore, we found regional knowledge gaps and clusters.Three countries (Nigeria, Colombia and Brazil) accounted for nearly half of the country occurrences, and many top-producing countries are under-researched (e.g. the Democratic Republic of Congo produces 3863 t/year and has 11 publications).In Africa, cassava is the second most important crop after maize (FAO, 2019) and its importance may increase further as it is better adapted to the impacts of climate change than other crops (Jarvis et al., 2012).Therefore, increasing the evidence base around cassava in under-researched regions should be prioritised to broaden the relevance of cassava research and reduce existing yield gaps there.
We also found knowledge gaps and clusters in the interventions and outcomes that were tested.The majority of research has focussed on production, with minimal research into non-crop habitats, wildlife or pollutants.Improving the knowledge base around environmental interventions and outcomes will facilitate evidence-based management of cassava, and therefore, production and expansion that is more sustainable and reduces negative impacts on biodiversity (Sutherland et al., 2004;Walsh et al., 2015).Such environmental benefits will likely also promote yield stability in cassava, as biodiversity provides resilience to yield fluctuations in other systems (Dardonville et al., 2022;Haughey et al., 2018).This may help to address existing yield gaps (Figure 5).In terms of knowledge clusters, synthesists can use this map to identify well-researched topics; for example, we found many publications for interventions related to fertiliser use and crop rotations.Stakeholders should aim to increase the knowledge base in environmental topics in particular to facilitate sustainable management of cassava.
There were several concerning trends in terms of study design.
Reporting standards were poor, and we recommend that reporting checklists/guidelines are promoted within this field to improve this (Hopewell et al., 2012;Plint et al., 2006;Stevens et al., 2014;Turner et al., 2012).Long-term projects and more robust experimental designs should also be prioritised, and systemic changes are needed across academia to address the problem of long publication delays (Christie et al., 2021).Given the importance of cassava as a staple crop, we urge researchers, funders, policymakers and other stakeholders to use our systematic map to guide syntheses and fill knowledge gaps with primary studies.Doing this could help to close existing yield gaps, increase food security and reduce the impacts of cassava production on the wider environment.

1.
Which studies have measured the impacts of cassava farming interventions (e.g.intercropping/tilling) on agricultural, economic and environmental outcomes (e.g.yield/soil)?2. Which interventions and outcomes have been studied?3. Which countries have been studied and when? 4. What is the distribution and abundance of studies between different interventions, outcomes, countries and years?In other words, where are the knowledge gaps or knowledge clusters in this map? 5. Which study designs have been used (metrics include reporting completeness, experimental designs, study durations and publication delays)?
At title and abstract stage, the study was excluded if the title or abstract:1.was not in English 2. the study had no relevant 'PICO/PECO' components (where 'P' = populations/subjects, 'I/E' = interventions/exposures, 'C' = comparators and 'O' = outcomes; James et al., 2016).The interventions/exposures and outcomes refer to the ontologies, which are described in '2.2 Agri-ontologies 1.0' below, and the populations/ subjects and comparators are defined directly below.If the title and abstract were too ambiguous to know whether the study had any or all of the relevant PICO/PECO components (e.g.'Annual (n.e.c.)' categories to capture interventions or outcomes that did not fit into existing categories.These categories result in a comprehensive map, as all publications are included, and they facilitate updating the ontologies, as having many publications in 'n.e.c.' categories can indicate that a new category should be created if a commonF I G U R E 1The structure of the 'Interventions' ontology, with a sample section expanded to show the lower (more specific) levels of the ontology.
Figure 6a).Together, these seven interventions accounted for 56% of intervention occurrences.The remaining interventions had fewer than 84 publications each (Figure6a).
Downloaded from https://besjournals.onlinelibrary.wiley.com/doi/10.1002/2688-8319.12249by University of Reading, Wiley Online Library on [23/06/2023].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License studied biological control of weeds (e.g.undersowing; Figure 7).Few publications have looked at 'Non-crop habitat management' (n = 67, e.g.planting in margins; Figure 7).In terms of outcomes, the majority of publications measured outcomes related to 'Crops' (n = 913, 58.3% of high-level outcome occurrences) and 'Soil' (n = 264, 16.8% of high-level outcome occurrences; Figure 8).Within the 'Soil' category, 'Soil structure, functions, & chemistry' was measured more than 'Soil organisms' (17.8% of occurrences in the 'Soil' category).Outcomes in other categories were rarely measured, with the exception of 'Money' (n = 100, 4.2% F I G U R E 6 Barcharts showing the most studied (a) 20 interventions and (b) 20 outcomes classed at the lowest (most specific) level of the intervention and outcome ontologies.The total number of (a) interventions (n = 3890) (b) and outcomes (n = 4236) is greater than the number of publications included in the map (n = 1599) as many publications tested multiple interventions and outcomes.Colours show the categories of the (a) interventions and (b) outcomes at the highest level in our ontology.Some of these intervention and outcome terms have been abbreviated (see Supporting Information 3 and 4 for the full terms and explanations).
Downloaded from https://besjournals.onlinelibrary.wiley.com/doi/10.1002/2688-8319.12249by University of Reading, Wiley Online Library on [23/06/2023].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Crop yield, biomass, & propagules -686 Crop quality (e.g.taste & chemical composition) -433 Crop damage, infection, & infestation -218 Crop growth, size, & survival -320 Soil structure, functions, & chemistry -245 Soil organisms (incl.microbes) -47 Water quantity (use/loss) (e.g.irrigation, runoff) -33 Water quality (biocides, nutrients, sediments) bust experimental designs.Publication delays have multiple causes, and changes are needed throughout the research process to address these causes (e.g.simplifying formatting requirements, improving TA B L E 1 All possible experimental designs for studies with comparators: controlled, correlated and before-after.Designs that are not shown (e.g.randomised correlated studies) are not possible.Bars and numbers indicate the number of interventions that were tested using each design (n = 2951).Some study designs have commonly used names, which are referenced in the 'design' column.
ventions and outcomes maps, they can interrogate these data online (www.metadataset.com) to see whether their chosen topic has a robust study design.If not, they may wish to choose a different F I G U R E 9 Scatterplots showing the (a) study duration and (b) publication delay through time (year of publication).The colour and point area shows the number of publications with that value (maximum: [a] 16, [b] 10).Blue lines show local regression smoothers (LOESS) with a span of 1 and shaded areas show 95% pointwise confidence intervals.Zero values show durations <1 year.
26888319, 2023, 2, Downloaded from https://besjournals.onlinelibrary.wiley.com/doi/10.1002/2688-8319.12249by University of Reading, Wiley Online Library on [23/06/2023].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 26888319, 2023, 2, Downloaded from https://besjournals.onlinelibrary.wiley.com/doi/10.1002/2688-8319.12249by University of Reading, Wiley Online Library on [23/06/2023].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License topic.Similarly, funders and field researchers can use these results to identify topics that require further research, either because they have been overlooked or because they have been tested with biased examples, Supporting Information 3 and 4 for the full ontologies, or www.metad ataset.comfor an interactive online format).