A new global review of bird atlases and their contribution to knowledge

ABSTRACT Capsule Over 600 bird atlases projects have been implemented across 93 countries, with at least 380,000 participants. Bird atlases with larger geographical scope had greater research impact but those utilizing online data submission and so higher number of participants had lower research impact. Aims To provide a comprehensive global review of bird atlases, to explore the impact of bird atlases in research, and identify variables that influence impact. Methods A database of bird atlases was compiled. Variables were extracted including: overall survey effort (an index generated using principal components analysis (PCA) comprised of total survey area, number of participants, and number of data records); research impact (an index generated using PCA comprised of bibliometric measures extracted from Scopus and Google Scholar); geographical details; fieldwork, project, and publication timing; fieldwork methods. We then used mixed linear models to explore how these variables differed across atlases, and which were predictors of research impact index. Results As of 2021, over 600 bird atlases projects have been implemented across 93 countries, with at least 380,000 participants worldwide. Total survey area, geopolitical scale, and number of atlas generations had significant positive relationships with research impact. Negative relationships were found between research impact and whether an atlas was published in English and the use of online data submission platforms like eBird. However, we found a significantly positive relationship between atlases using online data submission and our measure of survey effort. Conclusions Bird atlases have been undertaken all around the world at a wide variety of geopolitical scales, and are likely to be influential through widespread impact on knowledge, including research impact and citizen science involvement. Atlases utilizing online data submission generate more data and have a higher level of participant engagement but are less frequently cited by researchers in both scientific and grey literature.

The use of volunteers in research (citizen science) has become incredibly popular since the 1970s.This trend has developed in response to a wide range of factors, including large-scale environmental change, budget limitations of government programmes, and an increasing desire from the public to be actively involved in science (Wright et al. 2015).Ornithology is a scientific field well-known for its long tradition of the involvement of amateurs, for whom the observation of birds, or 'birding', is a passion rather than a paid profession (Sullivan et al. 2009).Within ornithology today, there are many highly organized citizen science communities reporting bird observations.One example of citizen science in ornithology comes in the form of bird atlases.Bird atlases combine citizen science, birding, and professional ornithology in a coordinated effort to produce maps of bird distributions at a wide range of spatial scales, from local to international (Donald & Fuller 1998).Although distributional atlases can and have been completed for various groups of organisms, from plants to mammals, birds have become the most common study group, in part because many amateur observers can accurately identify bird species (Beck et al. 2018).
Distributional data, such as those produced by bird atlases, are essential to documenting and conserving biodiversity.Large amounts of data are needed, however, to usefully inform conservation policy and direct conservation action.Despite some shortcomings, a citizen science approach is typically the most practical way to achieve the quantity of data, over a sufficient geographic extent, needed to accurately map species distributions (Tulloch et al. 2013).Bird atlases result in accurate maps without interpolation, distinguishing themselves from more traditional methods of mapping species distributions, which involve drawing lines between localities at the limits of a species' known ranges and filling in the gaps with at best, predictive modelling and, at worst, interpolation (Harrison 1993).
Despite substantial variation in methodology, atlas projects share the ability to generate large, spatially explicit ornithological datasets, detailed maps of species presence/absence, breeding evidence, and/or abundance, as well as detailed habitat information for threatened species (Beck et al. 2018).In addition, many bird atlases were first conducted over 30 years ago, and there is now a tradition of repeating atlases in the same geographical area to quantify change.With the first atlas effort as a baseline, subsequent atlas generations add a new layer of value to the dataset (Gibbons et al. 2007, Beck et al. 2018).All of these outputs were largely unavailable at the scale and resolution produced by atlassing before the 1970s (Donald & Fuller 1998).
Global systematic reviews of bird atlases have been performed previously by Donald and Fuller (1998), Gibbons et al. (2007), and Dunn and Weston (2008).These reviews have documented patterns in the approach of different bird atlases, as well as the application of atlas data and the rate at which atlases are cited in scientific literature.However, the most recent of these reviews was published in 2008 before major developments in citizen science efforts which may have a significant effect on public awareness and the utilization of atlas data, namely the growth of mobile electronic devices and online citizen science platforms for easy submission of species' occurrence observations (Ball-Damerow et al. 2019).The recent development of web-based data-entry platforms like eBird, BirdTrack, Ornitho, and BirdLasser allows volunteers to access, manage, and submit the data they collect in real-time through associated mobile applications on their smartphones (Lee & Nel 2020).Ideally, this development should also serve to increase accessibility and allow data management to occur on a faster timescale allowing more efficient and timely conservation actions.However, there is little research at present of the implications of this paradigm shift on the use of these kinds of data for scientific research applications and, ultimately, conservation action (Sullivan et al. 2017).Even without the use of online data submission, there are conflicting opinions on whether noisy, often unstructured, data that is produced by bird atlases and other citizen science projects effectively facilitate knowledge generation (Isaac et al. 2014, Sullivan et al. 2017, Bayraktarov et al. 2019).
Here we present a global review of bird atlases that have been performed between 1966 and 2021 and quantitatively assess which variables influence the use of atlases in research.Searches of bibliometric databases for the number of times an atlas is mentioned or cited allow us to generate a measure of the 'research impact' of an atlas.Different bibliometric databases, however, have different approaches, strengths, and limitations, and so we use both Scopus and Google Scholar in this investigation.Elsevier's Scopus is a well-established alternative to Thomson Reuter's Web of Science, selectively indexing academic and scholarly publications (Harzing & Alakangas 2016).We chose Scopus over Web of Science due to Scopus' broader coverage of regional journals and inclusion of books (Calver et al. 2017).In comparison to both Web of Science and Scopus, Google Scholar is a much broader search engine and retrieves a lot of 'grey literature' (e.g.technical reports, theses, and preprints) that is not present in either of the other databases (Harzing & Alakangas 2016, Calver et al. 2017).It is important to note that grey literature is not typically considered in traditional scientific citation indices and the usefulness of Google Scholar itself in systematic reviews is still subject to debate (Haddaway et al. 2015).But due to the relevance of grey literature for conservation practitioners (e.g.conservation reports by non-governmental organizations to government, legislation) and because many of the atlases we review here are grey literature documents themselves, we chose to use both Google Scholar and Scopus to increase coverage of literature using atlas data or information (Haddaway & Bayliss 2015).Thus, the research impact of atlases analysed and discussed here does not just cover the peer-reviewed scientific literature, but a broader measure of an atlas's generation of knowledge.
We investigated potential atlas-specific predictors of research impact through linear modelling.The first variable of interest was whether an atlas used online data submission because we perceived that most new atlases are now using this approach to increase engagement, but Lee and Nel (2020) found that the use of BirdLasser for data submission in the Southern African Bird Atlas 2 (SABAP2) changed the way participants behaved in data collection, submitting far more unstructured, ad hoc records.We therefore predicted that the use of online data submission would have a negative relationship with research impact because online atlases may be creating more but less useful data.The second variable was whether an atlas was published in English because language has been identified as a major barrier to the application of scientific knowledge in a complex way.The convergence on English as the lingua franca for global scientific activities means that, in general, information published in English is available to larger groups of users around the world, but, as a consequence, other languages are not equally published and information most relevant to conservation is less accessible to field practitioners and policy makers where English is not used in daily communication (Amano et al. 2016).Due to this barrier to the publication, visibility, and use of non-English science, we therefore predicted that atlases published in English would have a positive relationship with research impact in comparison to atlases not published in, or translated to, English.The third variable was atlas generation (initial atlas, repeated atlas, subsequent repeated atlas, etc.) because repeated atlases are perceived to provide comprehensive data on range change which standalone atlases are unable to provide (Gibbons et al. 2007).We therefore predicted that each generation adds a significant layer of value to the overall dataset, resulting in a positive relationship with research impact.The fourth variable was write-up time, or the number of years in between the end of fieldwork and publication, because increasing time to publish data may result in decreasing relevance of the data to researchers.We therefore predicted that longer periods of write-up time result in significantly lower research impact.The fifth variable was the spatial extent of an atlas because larger atlas project encompass more researchers and the data are likely to be more widely applicable.We therefore predicted that atlases with greater total survey area would have higher research impact.The final variables of interest were minimum atlas grid area (some grid-based atlases utilize multiple sizes of sampling units to perform individual surveys and so the smallest grid size was recorded) and the ratio between minimum atlas grid area and total survey area because these determine the spatial resolution of sampling for an atlas and so the resolution of any conservation recommendations.We therefore predicted that coarser atlases (greater minimum atlas grid area and lower grid: area ratio) would have lower research impact because highresolution data is more tractable for species distribution modelling and conservation planning (Tulloch et al. 2013).We also considered potentially confounding effects, such as the number of years since publication because an atlas is more likely to have been cited as this increases, regional variation (continent of atlas) because Gibbons et al. (2007) found regional patterns in the methodologies of bird atlases, and finally, geopolitical scale because smaller-scale atlases or larger-scale atlases with coarse spatial resolution or patchy effort may not generate the data necessary to be useful for research (Tulloch et al. 2013).

Generating the atlas database: locating atlases
In order to investigate the characteristics and methods of bird atlases, we compiled a global database of all the bird atlases we could find.Extensive Internet searches included using Web of Science, Google Scholar, commercial book websites (e.g.www.amazon.co.uk/ and www.nhbs.com/),Twitter, Facebook, Instagram, and the Global Biodiversity Information Facility (GBIF) to find mentions of atlases.Searches were conducted between January to April 2021.Search terms in different languages were developed throughout the compilation process to find non-English atlases (e.g.'Atlante degli uccelli', 'Atlas de las aves', 'Atlas ptaków').Despite this concerted effort, it is almost certain that atlases have been missed and particularly local atlases with limited online dissemination.But considering the larger dataset we have produced compared to both Dunn and Weston (2008) and Gibbons et al. (2007), when considering atlases prior to their review dates, we suggest that we have sampled extensively and to the most complete degree to date.

Generating the atlas database: selecting and defining atlases
We largely followed the definition of a bird atlas provided by Dunn and Weston (2008): 'a project that conducts or collates (or both) surveys of bird presence or abundance that includes a spatial mapping component, and covers a significant geographical area, such as a county or local government area, state, province, country or continent'.Most recent atlases are generated using field observations from large numbers of volunteers, and here we specifically exclude any mapping exercise (atlas) conducted entirely by professional ecologists or researchers.For example, one of the earliest bird atlases, the Breeding Birds of North Dakota, field surveys were performed largely by one professional ornithologist, with assistance from other biologists (Stewart 1975).This review therefore focuses only on atlases produced by volunteers but in some cases it was not always possible to determine whether volunteers were the primary fieldworkers.
Each discrete project with its own temporal demarcation, objectives, and survey effort was considered independent, even if they were performed in the same geographical region.For example, Cadman et al. (1987) documented breeding birds in the province of Ontario from 1981 to 1985, Cadman et al. (2007) documented breeding birds in Ontario in 2001-2005, and the ongoing Ontario Breeding Bird Atlas 3 (Purves 2020) will cover breeding birds from 2021 to 2025.These three projects are considered discrete atlases and the latter two are considered 'repeat atlases', in which atlases are performed at the same time of year and geographical location, but at a different point in time, often with slightly different methodology and research objectives.We refer to these discrete projects as different atlas 'generations'.Breeding and wintering atlases were not considered repeats of each other, as these cover different birds and have different objectives.
In some cases, it was difficult to categorize atlases spatially and temporally in the classes we present above, such as fixed survey periods and independent survey effort.We chose to define continuous projects (those with ongoing data collection and no fixed enddate) as atlases if they collected data specifically as an atlas and were separate from other surveys.Many local atlases, like the regional or county atlases produced in Britain, were included in our analysis even though the data produced in these projects were also used to produce the atlases covering the entirety of Britain and Ireland.A regional atlas project was thus considered its own atlas if a separate publication was produced.Indeed, many of these smaller atlases involved differing amounts of fieldwork years and/or smaller grid size than the larger-scale atlas, further increasing their independence.In practical terms, these atlases were also included because they may have different scientific or conservation objectives than the national atlas and their data may be used in different research contexts.These issues of statistical independence were then dealt with by using a mixed model framework (see below).
Like Dunn and Weston (2008), specialized atlases including strictly migratory or seabird surveys were excluded, because these typically involve only expert ornithologists rather than volunteers.This choice does exclude a small number of atlases being produced in the Caribbean, such as the Seabird Breeding Atlas of the Lesser Antilles (Lowrie et al. 2012).Several atlases covering wintering, breeding, and migratory birds were included, because these involved a mix of volunteers and professional ornithologists performing fieldwork.Provisional or pilot atlases, such as the pilot surveys undertaken in Howard and Montgomery counties in Maryland prior to the first Maryland-DC atlas, were also excluded from our database, partly due to a lack of basic information (Klimkiewicz & Solem 1978).Finally, we also stipulated that atlases should be produced from discrete projects that systematically collected and mapped new data.For example, Price et al. (2002) produced a North American atlas from broadscale bird monitoring scheme data, so was excluded.However, it was not always possible to distinguish between projects that produced new data and those that mapped pre-existing data or casual observations onto a grid, and it is likely that some atlases that fell into the latter category were included in our sample.
Table 1.List of variables extracted from each atlas where possible.

Geographical details
Specific location/region Country Continent (Africa, Asia/Middle East, Europe, North America, Australia/ Oceania, South America) Geopolitical scale (regional/local, national, multi-national) Total area encompassed by atlas (km 2 ) Publication details Title of project/publication Year of publication Full reference Write-up time: years between end of fieldwork and year of publication (by calculation) Publication languages (including atlas summaries and available translations)

Organizational details
Host organization name Host organization type (e.g.ornithological society, governmental organization, university/research institution) Whether or not volunteers were used

Fieldwork period
Year fieldwork started Year fieldwork ended Total years of fieldwork (by calculation) Fieldwork season (breeding, wintering, year-round) Whether or not the atlas is continuous

Fieldwork methods
Whether or not the atlas was grid-based Grid units (e.g.minutes, degrees, kilometres, yards) Dimensions of grid cells (e.g. 10 × 10 km, 1 × 1 degree) Area of each grid cell (km 2 ) Grid shape (e.g.square, rectangle, irregular polygon) Total number of grid cells Whether or not online data submission was used Online data portal name

Repeat atlases
Generation (e.g.first atlas, second atlas, third atlas, etc.) Inter-atlas interval (years; by calculation) In many cases, very little information could be sourced on atlases online.Generally, an atlas was included if it could be shown to have been published (i.e. on a booksellers website), was underway presently, or even that a singular species map could be located online (i.e. from an ornithological society's webpage).Unfinished or abandoned atlases were excluded, such as the Yolo County (California) atlas, but completed yet unpublished atlases like the Friuli-Venezia-Giulia (Italy) atlas were included, because sufficient relevant information on these could often be extracted from websites or reviews (Fraissinet 2011).It is likely that some atlases that do not satisfy the requirements listed here were inadvertently included in our analysis due to a lack of information provided online and this represents one of the limitations of our study.

Generating the atlas database: extracting data from atlases
We extracted a several variables from each atlas (Table 1).Variables were extracted from a wide variety of sources, including full atlases, atlas excerpts, reviews, booksellers' descriptions, social media, and news articles.

Generating an atlas survey effort index
Correlation tests were run between all variables prior to statistical modelling.Logged area, logged number of data records, logged number of effort hours, and logged number of observers were highly correlated with one another (Spearman's correlation test >0.60 between all variables).We combined logged area, logged number of data records, and logged number of observers variables using principal components analysis (PCA) to form a single score to measure overall survey effort as an index, with which all three variables showed a positive relationship.The first principal component accounted for 84% of the variance in the effort measurements.Due to low sample size, number of effort hours was excluded from this measure of effort, but number of effort hours was correlated with the principal component effort index by 0.81.The effort index correlated with logged area by 0.88, logged number of observers by 0.94, and logged number of records by 0.93.

Use of atlases in research: citation analysis
Citations are an imperfect but widely used measure of research impact, knowledge generation, and/or relevance (Aksnes et al. 2019).In order to determine the research impact, or the influence of an atlas on scientific and grey literature, all published atlases (n = 549 excluding unpublished works and non-continuous atlas projects currently undergoing fieldwork) were searched for in Scopus and Google Scholar using the title(s) of the publication or project as the search term in quotation marks (e.g.'European Breeding Bird Atlas 2: Distribution, Abundance and Change').For atlases with more than one title (i.e.use of multiple languages), the Boolean operator OR was used; a full table of search terms is provided in online Table S1.Searches were performed in May and June 2021.For Google Scholar, the full text was searched for mentions of the atlas title.The Scopus database does not search the full text, but all fields were searched for atlas titles, which includes numerous aspects of a publication, including the abstract, references, and title.The number of search hits for documents containing the title of an atlas publication was treated as an index of the actual number of citations for an atlas.From this, we derived a citation rate using the number of search hits divided by the number of years since the publication of the atlas.For Scopus, we were also able to collect the maximum yearly citation rate.Citation rates were not calculated for atlases published in 2021 or continuous/online atlases without a publication date.
Ten atlases were randomly selected and all papers produced by the search were inspected to determine whether the atlas was cited in the journal article or referred to within the document.The percentage of erroneous inclusions for a single search query ranged from 0% to 9% (mean % ± SE: 3 ± 1) for Google Scholar and from 0% to 11% (mean % ± SE: 1 ± 1) for Scopus.Only Google Scholar suffered from 'stray citations' (a phenomenon where slight variations in referencing result in duplicate records for the same paper or publication); for this random sample, the percentage of duplicates ranged from 0% to 18% (mean ± SE: 1.7 ± 0.58; Harzing & Alakangas 2016).For Google Scholar, most searches returned the actual atlas publication, but this was not excluded from the index because the repetition was equal across all searches.These errors and biases support our decision to not rely on a single database to investigate citations and research impact of bird atlases.

Generating an atlas research impact index
All citation measures (Google Scholar number of search hits, Scopus number of search hits, Google Scholar average yearly citation rate, Scopus average yearly citation rate, and Scopus max yearly citations) were highly correlated with each other (Spearman's rank P. C. POTOTSKY AND W. CRESSWELL correlation test > 0.70 between all variables).As with our effort index, we combined these variables using PCA to form a single score to measure overall research impact as an index, with all variables showing a positive relationship.The first principal component accounted for 89.5% of the variance in the citation measures.The extracted research impact index correlated with Google Scholar number of search hits by 0.95, Scopus number of search hits by 0.96, Scopus average yearly citation rate by 0.98, Google Scholar average yearly citation rate by 0.88 and Scopus max yearly citations by 0.96.

Modelling atlas characteristics and research impact: statistical analysis
All analyses were performed using R 4.0.5 (R Core Team 2021).Summary statistics are presented as means ± 1 standard error.We tested for differences in effort index between atlases using online data submission and those using manual data submission with a linear mixed model with years since publication (measure of temporal variation) and continent (measure of regional variation) as fixed effects.Atlases for which there were missing data for any of the variables in the models were excluded from analysis.We repeated this analysis looking at differences in effort index between atlas generations.For atlas generation, we pooled fourth and fifth-generation atlases together (due to low sample size) and treated generation as a continuous numeric variable.Atlas projects that had no repeat (as of early 2021) were still treated as firstgeneration atlases.
For the second part of our analysis, we used mixed linear models to explore the influence of the eleven variables that were hypothesized to influence research impact index.The research impact index was log transformed to obtain normally distributed residuals, as was total survey area.Variables of interest, including whether an atlas used online data submission, whether an atlas was published in English, atlas generation, number of fieldwork years, write-up time, minimum atlas grid area, and atlas grid: total survey area ratio were inputted into a full model and no model selection was undertaken in order to include potentially confounding effects, such as the number of years since publication, regional differences (continent of atlas), and geopolitical scale.Atlas generation and geopolitical scale were also treated as a continuous variable.No interactions between variables were considered in order to avoid over-parameterization of the model.In order to increase sample size due to atlases with missing data, a model where effort index was replaced with logged survey area was produced, because these two variables were shown to be highly correlated in our PCA.Both models (high and low sample size) are presented to show that the overall results remain the same, although with some differences in statistical significance as might be expected with changes in statistical power.Possible non-linear effects for research impact due to citation curve were considered by running analyses with atlases published later than 2019 excluded and there was no difference in model results for the high sample size model.For the low sample size model, the results also remained the same, but only geographical scale remained statistically significant.Model fit was assessed by visual inspection of residuals plotted against fitted values and quantile plots and all were reasonable after log-transforming research impact index.Effect sizes for variables of interest were calculated by dividing the predicted value over the range of the data in the high sample size model, with all other parameters set to average values.Finally, predicted values were plotted using the ggplot2 package in R (Wickham 2016).

Atlas characteristics
The database contains information on 603 atlases (see online Appendix S1 for details).Most atlases have been published as large books or journal articles, but a small percentage (3%) have been published as webpages only.Some have gone unpublished entirely, and some are in preparation for publication (as of early 2021).Descriptive statistics for the database are shown in Table 2. Sample sizes vary considerably and only 5% of atlases had all fields filled in for the variables listed in Table 1.

Geographical distribution of atlases
Atlases were carried out in 93 countries (Figure 1), on a large range of geographical scale, from a 1 km 2 university campus (Matos & Luís 2007) to an entire continent (Keller et al. 2020).In terms of geopolitical scale, 83% of atlases were on a regional or local scale (mean area covered in km 2 : 56,393, range: 1-3,960,000, n = 455), 15% on a national scale (mean area covered in km 2 : 52,693, range: 61-7,688,287, n = 91), and only 2% on a multi-national scale (mean area covered in km 2 : 2,068,772, range: 41,445-3,960,000, n = 12) (Figure 2).Despite particular effort placed on finding as many atlases as possible on a global scale, 89% of atlases were from Europe or North America (n = 536), with 104 atlases from within Italy alone.Following Italy, the top producers were the United States (93), the UK (81), and France (46).In Europe and North America, 86% of atlases were regional or local, whereas only 54% of atlases elsewhere (n = 67) were regional or local (Figure 2(a)).For example, forty US states and Puerto Rico have carried out at least one statewide atlas, as well as numerous repeat atlases in the same geographic area and smaller-scale county atlases.

Temporal and seasonal distribution of atlases
The database contained atlases published between 1970 and 2021.Twenty-two atlases were still collecting data as of early 2021, with six continuous projects (e.g.SABAP2 and the Nigerian Bird Atlas Project).Atlases have started data collection in every decade since the 1960s, with two peaks in the 1980s and the early 2000s (Figure 2(b)).Most atlases collect data for five years, but the range is wide, with some data collection spread over several decades.
Most atlases do not collect data on distribution throughout the entire year: 70% of atlases focus only on breeding distribution, 22% mapped seasonal (i.e.breeding and wintering, or wet and dry season) or year-round distribution, and 8% mapped wintering distribution.There were no significant differences in area covered, length of fieldwork, grid size, or effort index between breeding and wintering atlases (details in online Table S2).

Sampling and spatial resolution of atlases
Most atlases use a grid system to structure sampling.The area, shape, size, and measurement unit of the grid cell vary to differing extents (Table 2).Where the grid shape was known (n = 501), 98% used square or rectangular blocks, but a few atlases used irregular polygons based on physical or ecological structures (e.g.Maffei et al. 2001) or hexagons (Castro-Prieto et al. 2020).Most grid-based atlases used kilometres, miles, hectares, or yards to measure cells.For 18.5% of grid-based atlases areas were defined by minutes or degrees, which mean the area of the grid cell may vary significantly throughout the survey area, depending on latitude.Only four atlases (<1%) were found to not use a grid cell system at all, but collected data for point localities (e.g.Barrett et al. 2003).

Organizational aspects of atlases
For 65% of atlases, volunteers were explicitly stated as being the primary data collectors.Where known (n = 547), atlases were organized and run by ornithological societies (62%), governmental organizations (12%), universities and other research institutions (11%), naturalist groups and conservation NGOs (7%), museums (5%), park and reserve organizations (2%), and project-specific groups (1%).Many atlases were highly collaborative, involving multiple organizations of differing types, but these percentages consider only the main organizing body.

Atlas effort
The total number of observers for the atlas database was 389,126 (n = 319 atlases).This value is a minimum, because many atlases reported minimum figures or did not report number of observers at all.The total number of data records reported was 126,912,197 (n = 168 atlases).Again, this is a minimum, as most atlases did not report summary statistics.Finally, only 67 atlases reported effort hours of participants, but the total number of hours invested in atlas projects was 3,587,087, or over 400 full calendar years of survey effort.On average, each atlas project represents 53,539 effort hours, or six full calendar years of survey effort (Table 2).Table 3 shows a ranking of the top ten atlases for effort index.

Online data submission
Where known (n = 550), 20% of atlases used some form of online data submission, whether this was by submitting data sheets online after a field survey or by submitting data in real-time from mobile data submission applications.Some of the most commonly used data portals or applications were BirdTrack, eBird and the individual country or local portals hosted through NaturaList (e.g.www.faune-france.org).Atlases using online data submission had a significantly higher measure of our effort index, even when controlling for years since publication and regional differences, but with much residual variation (Table 4).

Repeat atlases
In our database, 26% of atlases were repeat atlases.We found that effort index significantly increased with atlas generation, even when accounting for temporal and regional variation (Table 5).Furthermore, thirdgeneration atlases had significantly higher effort index values than second-generation atlases (Table 5).

Factors predicting atlas research impact
Log survey area and geopolitical scale were significant positive predictors of our research impact index in the high sample size model (overall R 2 = 0.56, Table 6, Figures 3(a,b)).Atlas generation was a marginally significant positive predictor.English as a publication language and the use of online data submission were significant negative predictors of research impact index (Table 6 and Figure 3).The effect size for using online data submission was a decrease in the average predicted research impact index with all parameters in Note: this considers local atlases as well, so for some countries, this may mean that only a small geographical area of the whole country has been covered by an atlas.On the same note, some European countries have not produced their own national atlas but have participated in the continental atlas (EBCC) and are still shaded here.Antarctica excluded.Mercator projection.the high sample size model set to their average value by 7.8%.For using English as a publication language, the average predicted research impact index with all parameters in the high sample size model set to their average value decreased by 4.5%.In addition, research index marginally significantly decreased with atlases produced in Asia and the Middle East (Table 6 and Figure 3(d)).No other variables had a statistically significant effect (Table 6).The results were the same for the low sample size model except that only effort index, online data submission, and geopolitical scale remained statistically significant.

Discussion
In this study, we were able to identify and characterize 603 bird atlases, involving at least 700,000 citizen scientists.In comparison to the previous reviews performed by Gibbons et al. (2007) and Dunn & Weston (2008), we were able to identify 185 atlases that had been initiated since 2005, a similar number of atlases compared to Gibbons et al. (2007), and at least 152 more atlases from 2006 and before compared to Dunn & Weston (2008).The average profile of a bird atlas globally is a breeding bird atlas organized by an ornithological society that collects data for six years, with about 1200 participants each contributing about 70 h of total effort, takes three years to compile and analyse data for publication, over an area of about 8000 km 2 with a minimum grid size of 70 km 2 .
A small number of countries dominate the production of bird atlases in terms of numbers of atlases produced, namely Italy, the United States, and the United Kingdom (Figure 1).In addition to long traditions of amateur participation in ornithology in these countries, this pattern may be due to the regional trends towards the production of urban atlases or atlases produced for states, counties, or provinces (Figure 2).It is a note for future research to investigate whether urban or smaller-scale atlases have relatively more or less utility for conservation, whether in terms of legislation, species conservation status, or habitat management.One potential strength in local atlases is the ability to tailor data collection protocols to generate the most useful data in the local conservation context.For example, for the Kerala Bird Atlas in India, participants were asked to note the presence of invasive plant species in order to provide additional ecological data (Praveen & Nameer 2021).Although atlases have been performed on all continents except Antarctica, there are still large gaps in survey effort, particularly for Asia, northern Africa, Central America, and parts of South America.We suggest that effort and funding should be focused on improving capacity in these regions to develop and implement bird atlas projects on at least a national geopolitical scale, particularly because citizen science initiatives can be a significant resource in countries with insufficient professional scientific capacity (Barnard et al. 2017).Another important process in developing capacity is the identification of champion individuals and relevant institutions to lead bird atlas projects.An example of these processes working successfully is the ongoing Nigerian Bird Atlas (Tende   Gibbons et al. (2007) found that atlas initiation had peaked in the 1980s, but we find nearly as many atlases began fieldwork in the 2000s (Figure 2(b)).We suggest that any troughs and peaks in the initiation of bird atlases are associated with the interval between repeat atlases (and the smaller-scale county or regional atlases which may run concurrently with these national or international repeated atlases), which is, on average, about 14 years (Table 2).In terms of seasonal distribution, most atlases document breeding birds; one of the benefits of this approach is that breeding bird atlases concentrate fieldwork into the shorter breeding season, in which highly mobile species are easier to locate, and have lower overall cost.They may also keep participants at a higher level of engagement for a shorter amount of time, with the added benefit of documenting breeding-specific behaviours.The documentation of breeding ranges and breeding-specific behaviours add additional biological meaning and context to atlas maps.
Before assessing the research impact of atlases, it is important to consider how research impact might be measured.How to deal with citation data and what database to use to evaluate scientific impact is a point of major discussion in the literature (Harzing & Alakangas 2016, Waltman 2016, Martín-Martín et al. 2018, Aksnes et al. 2019).Previous research has concluded that Google Scholar is not suitable to be used alone in systematic reviews (Haddaway et al. 2015).However, the clear need to incorporate grey literature, which often constitutes the majority of information available for conservation decision-making, into any measure of research impact for atlases led us to seek a method to use Google Scholar in conjunction with more rigorous bibliometric databases like Scopus (Haddaway & Bayliss 2015).In this paper, we combined five different measures of citations (Google Scholar number of search hits, Scopus number of search hits, Google Scholar average yearly citation rate, Scopus average yearly citation rate, and Scopus max yearly citations) into one   index, there is not a large difference between any of the citation measures used in this paper.The methodology used here combining citation measures from Google Scholar and Scopus with a PCA allows a multifaceted but straightforward approach to measuring research impact, or a measure of research impact that incorporates both academic and grey literature.Adopting this approach for future such reviews may remove concern about the validity of different databases being used or not by demonstrating that they give very similar relative values; however, this approach may only hold true in this sample of search terms and further research is needed.
The research impact modelling led to four key significant results.First, we found that the research impact of an atlas significantly increased with increasing survey area and geopolitical scale (Figures 3  (a,b) and Table 6).These two variables correspond to spatial coverage.Atlases performed on a national or international scale had a significantly higher research impact, as well as atlases with increasing geographical extent.This relationship is best illustrated by the first European Breeding Bird Atlas (Hagemeijer & Blair 1997), which had the highest research impact index by far (nearly double that of the second highest impact atlas), despite its coarse spatial resolution (50 × 50 km grid) and vast geographical diversity in terms of cultural, linguistic, economic and scientific contexts.In addition, this publication has already been established as a fundamental reference for the entirety of Europe (Herrando et al. 2019).
Second, we found that repeated atlases had significantly greater research impact (Figure 3(c) and Table 6).We also found that repeated atlases have significantly higher survey effort (Table 5).Both results suggest that repeating atlases provides increased value, in terms of both higher participant engagement and increased scientific yield with each generation.This finding provides quantitative justification for repeating atlases.
Third, an unexpected result was that English as a publication language for atlases was a negative predictor for research impact (Figure 3(d) and Table 6), given that English is considered the common mode of international scientific communication and language is often credited as a significant barrier to the transfer of scientific knowledge (Amano et al. 2016).It is also of interest that only 38% of atlases in our database could be shown to have been published in English or translated into English alongside another primary language.This may be largely influenced by the dominance of regional or local European atlases, particularly given that over 100 atlases were produced solely in Italy, none of which were found to have been translated to English.This is still meaningful however, as this relationship may also mean that data produced by bird atlases is being used mainly by local researchers who understand the publishing language and knowledge transfer is occurring within one country or language community.Our results suggest that where data can best be applied locally for conservation ends, then outputs are best in local languages.
Fourth, the use of online data submission was found to be a significant negative predictor of research impact, even with the number of years since publication controlled for in our model (Figure 3 and Table 6).In our database, 20% of atlases used online data submission and we expect this percentage will only continue to increase in the coming years.The results of this study quantify and highlight a key trade-off that occurs in citizen science.Although the use of online data submission contributed to significantly lower measures of research impact, atlases using online data submission also had significantly higher measures of survey effort (Table 4).Essentially, atlases using online data portals have higher engagement with participants but generate fewer research outcomes.As our survey effort index is highly correlated with the number of data records and effort hours, we suggest that the use of online data submission in bird atlases has led to a real or perceived prioritization of data quantity over data quality, resulting in reduced scientific yield.Online data submission platforms like eBird and BirdLasser vastly increase data collection potential without higher levels of structure or data curation, representing a potential barrier to the efficient use of atlas data in science which has not yet been sufficiently explored (Bayraktarov et al. 2019, Lee & Nel 2020).Our results suggest that the organizers of atlas projects using online data submission should make data management for clear scientific or conservation outcomes a greater priority, as well as fostering a culture amongst participants of strict adherence to data collection protocols.
The research impact modelling also showed that some variables were not important.Atlases were not less likely to be cited depending on the spatial resolution, temporal resolution, or temporal relevance: we did not find the minimum grid size and/or minimum grid size: total survey area ratio, length of fieldwork in years, or write-up time (number of years between the end of fieldwork and publication) to be significant predictors of research impact.Tulloch et al. (2013), in contrast, found that increasing spatial resolution was significantly related to the higher impact of an atlas on scientific literature, although regional variation or geopolitical scale were not accounted for in their modelling and they used a different sampling method to produce citation rates.Tulloch et al. (2013) also found no significant relationship between temporal resolution and scientific impact, suggesting that longer atlas projects do not necessarily produce more research.It is interesting that write-up time, essentially the time taken for collation, analysis, and synthesis of data produced by the atlas into a published book, was not a predictor of research impact.In some cases, write-up equalled or far exceeded the length of fieldwork, possibly decreasing the data's use for practical application in conservation, given the current rate of changes in breeding distributions of birds in response to environmental change.We expect that online data submission will have significant impact on write-up time, because computerization allows analysis to occur concurrently with data collection, although the sheer amount of data produced through this method may hinder this process.
Finally, we faced similar issues to previous reviews in generating sufficient sample sizes for modelling, as with Dunn & Weston (2008).We therefore urge atlas projects to maintain their online presence after the atlas is completed, to publish the data collection and analysis protocols, survey effort statistics and to make raw survey data more easily accessible.
The analysis presented here should form the basis of a continuing assessment of the performance of bird atlases, particularly as data technologies change and subsequently pose or deconstruct barriers to its use as a primary tool in science and conservation.The next stage of research would be to quantify real conservation action that results from or is supported by data produced by bird atlases, because it may not be the case that research impact and conservation action are directly linked (Meijaard & Sheil 2007, Knight et al. 2008, Shanley & López 2009).Overall, we have seen that bird atlases have proliferated across the globe over the past sixty years, with more atlases being performed with more participants and more data produced.The factors which determine the research impact of an atlas are not necessarily obvious or simplistic, but certain characteristics should be embraced by future atlases and these include repeating atlases over the same area at regular intervals, covering larger geographical areas, as well as publishing atlases in the local language rather than English.The use of online data submission needs to be carefully managed by atlas organizers because we have shown that this approach undermines, rather than enhances, the research impact of an atlas.However, the significant engagement benefits, and potentially much wider conservation gains through empowering and informing citizen scientists, provided by online data submission, may mean this is an acceptable trade-off.
hours Number of observers per km 2 (by calculation) Effort hours per observer (by calculation) Total number of grid cells Whether or not online data submission was used Online data portal name

Figure 1 .
Figure 1.The number of atlases performed in each country.

Figure 2 .
Figure 2. (a) Number of atlases performed around the world with geopolitical scale, n = 603.(b) The decade in which atlases were initiated with geopolitical scale, n = 572.

Figure 3 .
Figure 3. Modelled relationships for the final, high-sample size model (Table 6) of research impact for 339 bird atlases and (a) log area, (b) geopolitical scale, (c) atlas generation, (d) language of atlas publication, and (e) continent.Lines represent the predicted values from the high sample size linear model using median values for all other variables with continent set to Europe and publication language set to English.Red lines correspond to atlases utilizing online data submission and black lines correspond to atlases which did not utilize online data submission.Shaded areas or whiskers indicate one standard error.

Table 2 .
Descriptive statistics for key variables from 603 bird atlases.Indicates median as opposed to mean.Note: continuous atlas projects were cut off at 2021 to calculate fieldwork duration. *

Table 3 .
List of the top 10 atlases with highest effort index (n = 133) with fieldwork season, fieldwork time period, effort index value, and whether the atlas project utilized online data submission.

Table 4 .
Relationship between effort index with online data submission, number of years since publication, and continent.P < 0.001, adjusted R 2 = 0.27.The intercept for continent is set to Africa.P values less than 0.05 are in bold.etal. 2016).However, the complexity engendered by the scale, effort, and organization needed for a firstgeneration national bird atlas in countries affected by political and/or socioeconomic instability should not underestimated.

Table 5 .
Relationship between effort index with atlas generation (first column includes all generations, the second column compares second and third-generation atlases), number of years since publication, and continent.
research impact index through PCA.Scopus and Google Scholar produce different citation counts: Google Scholar consistently produced higher numbers of citations, which is expected as more low-impact documents and grey literature are included in this database (Martín-Martín et al. 2018).However, given that the first principal component, (our 'research impact index'), captured 89.5% of the variance in all citation measures, and all citation measures were highly correlated with each other and our research impact