No one-size-fits-all solution to clean GBIF

Choice of study taxa

This study is the outcome of a workshop held at the Federal University of Rio Grande do Norte in Natal, Brazil in October 2018 which gathered students and researchers working with different taxonomic groups of animals, fungi, and plants across the Neotropics (Fig. 1). Each participant analysed geographic occurrence data from their taxonomic group of interest and commented on the results for their group. Hence, we include groups based on the expertise of the participants rather than following an arbitrary choice of taxa and taxonomic ranks. We acknowledge a varying degree of documented expertise and number of years working on each group. We obtained public occurrence records for 18 taxa, including one plant family, nine plant genera, one genus of fungi, three families and one genus of arthropods, one family of snakes, one family of skates, and one genus of bony fish (Table 1).

Table 1:

The study groups and their taxonomy.

This study includes three marine and 15 terrestrial taxa, seven of them animals, one group of fungi and ten plants, belonging to 16 different orders. The outlines illustrate the broad taxonomic group (i.e., an evolutionary relative if relative if no icon of the direct study group was available). Icons from http://www.phylopic.org if available under a Public Domain license otherwise created by the authors (Heath, 2020; Hillewaert, 2006; Hough, 2008; Menchetti, 2020; McNair, 2020a; McNair, 2020b; Müller, 1885; Nimphel, 2020; Petar, 2020; Pohl, 1827; Reinke, 2020; PhyloPic, 2020a; PhyloPic, 2020b; Welter-Schultes, 2017; Xgirouxb, 2020; Veronidae, 2012).

	Taxon	Taxon rank	Realm	Common name	‘Phylum’	Family
	Diogenidae	Family	Marine	Hermit crabs	Arthropoda	Diogenidae
	Entomobryidae	Family	Terrestrial	Springtails	Arthropoda	Entomobryidae
	Neanuridae	Family	Terrestrial	Springtails	Arthropoda	Neanuridae
	Tityus	Genus	Terrestrial	Scorpions	Arthropoda	Buthidae
	Arhynchobatidae	Family	Marine	Skates	Chordata	Arhynchobatidae
	Dipsadidae	Family	Terrestrial	Snakes	Chordata	Dipsadidae
	Harengula	Genus	Marine	Herrings	Chordata	Clupeidae
	Thozetella	Genus	Terrestrial	Sac fungi	Ascomycota	Chaetosphaeriaceae
	Conchocarpus	Genus	Terrestrial	NA	Angiosperms	Rutaceae
	Gaylussacia	Genus	Terrestrial	Huckleberries	Angiosperms	Ericaceae
	Harpalyce	Genus	Terrestrial	NA	Angiosperms	Fabaceae
	Iridaceae	Family	Terrestrial	NA	Angiosperms	Iridaceae
	Lepismium	Genus	Terrestrial	Cacti	Angiosperms	Cactaceae
	Oocephalus	Genus	Terrestrial	NA	Angiosperms	Lamiaceae
	Pilosocereus	Genus	Terrestrial	Cacti	Angiosperms	Cactaceae
	Prosthechea	Genus	Terrestrial	Orchids	Angiosperms	Orchidaceae
	Tillandsia	Genus	Terrestrial	Bromeliads	Angiosperms	Bromeliaceae
	Tocoyena	Genus	Terrestrial	NA	Angiosperms	Rubiaceae

DOI: 10.7717/peerj.9916/table-1

Species occurrence data

We downloaded occurrence information for all study groups from http://www.gbif.org using the rgbif v1.4.0 package (Chamberlain, 2017) in R (GBIF.org, 2019a; GBIF.org, 2019b; GBIF.org, 2019c; GBIF.org, 2019d; GBIF.org, 2019e; GBIF.org, 2019f; GBIF.org, 2019g; GBIF.org, 2019h; GBIF.org, 2019i; GBIF.org, 2019j; GBIF.org, 2019k; GBIF.org, 2019l; GBIF.org, 2019m; GBIF.org, 2019n; GBIF.org, 2019o; GBIF.org, 2019p; GBIF.org, 2020a; GBIF.org, 2020b). We downloaded GBIF-interpreted data including only records with geographic coordinates and limited the study area to a rectangle between 90°S–33°N and 35°W–120°W reflecting the Neotropics (Morrone, 2014), our main area of expertise. The natural distributions of all included taxa are confined to the Neotropics except for Arhynchobatidae, Diogenidae, Dipsadidae, Entomobryidae, Gaylussacia, Iridaceae, Neanuridae, and Tillandsia, for which we only obtained the Neotropical occurrences. We consider GBIF data generally of high quality and use them as a case study because GBIF is the largest, most widely used and taxonomically most comprehensive data source for species occurrence records; however many more exist (e.g., https://bien.nceas.ucsb.edu/bien/, http://www.fishbase.de or Guedes et al., 2018). GBIF provides information on the internal consistency of records, among others including information on decimal rounding of coordinates, geographic projection, date validity and geospatial issues. Since we specifically aimed to test the effect of user-level filtering we included records flagged with issues by GBIF (this was also the default option). Geospatial issues flagged by GBIF only concerned 0.4% of the records used in this study and including them had the added benefit to make our results directly comparable to other databases, which may use different internal consistency checks or none at all.

Automated cleaning

We followed the cleaning pipeline outlined by Zizka et al. (2019) and first filtered the data as downloaded from GBIF (“raw”, hereafter) using meta-data for those records for which they were available (although meta-data were often missing, Peterson et al., 2018), removing: (1) records with a coordinate precision below 100 km (as this represents the grain size of many macro-ecological analyses); (2) fossil records and records of unknown source; (3) records collected before 1945 (before the end of the Second World War, since coordinates of old records are often imprecise); and (4) records with an individual count of less than one and more than 99. Furthermore, we rounded the geographic coordinates to four decimal places and retained only one record per species per location (i.e., test for duplicated records). In a second step, we used the clean_coordinates function of the CoordinateCleaner v2.0-11 package (Zizka et al., 2019) with default options to flag errors that are common to biological data sets (“filtered”, hereafter). These include: coordinates in the sea for terrestrial taxa and on land for marine taxa, coordinates containing only zeros, coordinates assigned to country and province centroids, coordinates within urban areas, and coordinates assigned to biodiversity institutions. See Table 2 for a summary of all filters we used and their classification into “erroneous” and “unfit”.

Downstream analyses

We first generated species richness maps using 100x100 km grid cells for the raw and filtered datasets respectively, using the package speciesgeocodeR v2.0-10 (Töpel et al., 2017). We then performed an automated conservation assessment for all study groups based on both datasets using the ConR v1.2.4 package (Dauby et al., 2017). ConR estimates the EOO, AOO, and the number of locations, and then suggests a preliminary conservation status based on Criterion B of the global IUCN Red List. While these assessments are preliminary (see IUCN Standards and Petitions Subcommittee, 2017), they can be a proxy used by the IUCN to speed up full assessments. We then benchmarked the preliminary conservation assessments against the global IUCN Red List assessments for the same taxa (where available), which we obtained from http://www.iucn.org via the rredlist v.0.5.0 package (Chamberlain, 2018).

Evaluation of results

Each author provided an informed comment on the performance of the raw and cleaned datasets, concerning the number of removed records and the accuracy of the overall species richness maps. We then compared the agreement between automated conservation assessments based on raw and filtered occurrences with the global IUCN Red List for those taxa where IUCN assessments were available (http://www.iucn.org).

We carried out all analyses in the R computing environment (R Core Team, 2019), using standard libraries for data handling and visualization (Wickham, 2018; Garnier, 2018; Ooms, 2014; Ooms, 2019; Hijmans, 2019). All scripts are available from a Zenodo repository (doi:10.5281/zenodo.3695102).

The relevance of individual filters

The aim of automated filtering is to identify possibly problematic records that are unsuitable for particular downstream analyses. While those records filtered as “erroneous” will likely cause problems for most biodiversity research, those filtered as “unfit” might have varying impact, depending on the type and spatial resolution of the downstream analyses. Unwanted effects include an unnecessary computational burden, which can be a bottleneck for large-scale analyses (i.e., duplicates, Antonelli et al., 2018), and increased uncertainty (due to low precision), or completely compromising results. For instance, records assigned to country centroids might be acceptable for inter-continental comparisons, but are likely to be erroneous for species distribution modelling on a local scale. The importance of each test and the linked thresholds must be judged based on the specific downstream analyses. As our results show, it may be advisable to adapt automated tests to the geographic study area or the taxonomic study group. For instance, the high number of records flagged for centroids on the Lesser Antilles (Fig. 3) might be overly strict (https://data-blog.gbif.org/post/country-centroids/), although we chose a conservative distance for the Political centroid test (1 km).

Several factors may explain the high proportion of records flagged as duplicates. First, the deposition of duplicates from the same specimen at different institutions is common practice, especially for plants, where a specimen duplication is entirely feasible. Second, independent collections at similar localities may occur, in particular for local endemics. Third, low coordinate precision, for instance based on automated geo-referencing from locality descriptions, may lump records from nearby localities. Fourth, different data contributors might add the same record to GBIF, if their sources overlap, as can for instance be the case for the Barcode of Life and Plazi databases.

Similarities and differences among taxa

The number of records flagged by individual tests was similar across study groups, suggesting that similar problems might be relevant for collections of plants and animals. Therefore, the same filters can be used across taxonomic groups. Some notable exceptions stress the need to adapt each filter to the taxonomic study group to balance data quality and data availability. The high fraction of records filtered by the “basis of record” filter for Tityus, Dipsadidae, Prosthechea and Tillandsia, were mostly caused by a high number of records in these groups based on unknown collection methods, which might be caused by the contribution of specific datasets lacking this information for these groups. The high fraction of records flagged by the “collection year” filter for Dispadidae was caused by a high collection effort in the late 1880s and early 1900s, as can be expected for a charismatic group of reptiles, but also by 500 records dated to the year 1700. The latter records likely represent a data entry error: they are all contributed to GBIF from the same institution, and the institution’s code for unavailable collection dates is 1700-01-01–2014-01-01, which has likely erroneously been converted to 1700. The high number of species flagged at capital coordinates and within urban areas for the plant groups Iridaceae and Oocephalus might be related to horticulture, since at least some species in those groups are commonly cultivated as ornamentals. This was supported by the detailed examination of the data for Iridaceae, which showed that after filtering 1605 records from 69 exotic species remained in the dataset, stressing the importance to address these species in certain taxonomic groups.

The general agreement between the species richness maps based on raw and filtered data was encouraging, in terms of the use of this data for large-scale biogeographic research (Fig. 4, Appendix S2). The filter based on political centroids had an important impact on species richness patterns, which is congruent with the results from a previous study in the coffee family (Maldonado et al., 2015). Records assigned to country or province centroids are often old records, which are geo-referenced at a later point based on vague locality descriptions. These records are at the same time more likely to represent dubious species names, since they might be old synonyms or type specimens of species that have only been collected and described once, which are erroneously increasing species numbers.

Overall, we consider the effect of the automated filters as positive since they identified the above-mentioned issues and increased the data precision and reduced computational burden (Table 3, Appendix S2). However, in some cases filters failed to remove major issues, often due to incomplete meta-data. For instance, for Diogenidae we found at least two records of an species known only from Eocene fossils (Paguristes mexicanus) which slipped the “basis of record” test because they were marked as “preserved specimen” rather than “fossil specimen”. Furthermore, for Entomobryidae we found that for 1,996 records the meta-data on taxonomic rank was “UNRANKED” despite all of them being identified to species level, leading to a high fraction of records removed by the “Identification level” filter. Additionally automated filters might be overly strict or unsuitable for certain taxa. For instance, in Entomobryidae, 2,004 samples were marked as material samples and therefore removed by our global filter retaining only specimen and observation data, which in this case was overly strict.

The filters we included in this study address a set of important but relatively easy to identify problems. In fact, the internal quality control of GBIF does flag some of the problems we tested for (i.e., zero coordinates, equal lat/lon) while others might be implemented in the near future (country centroids, https://data-blog.gbif.org/post/country-centroids/). While this internal quality control is very helpful, we see a huge potential to overcome issues with data quality in a user-feedback system that allows users to provide expert assessments, i.e., a meta-annotation of records being challenged (and why). Such a system would not need to change the original data and could include multiple levels to account for differing opinions.

As next steps for automated filtering, tests for intrinsic consistency and support by external data (if available) can help to detect additional problematic records. For instance, testing if records’ coordinates fall within the state or province of collection noted for a record (intrinsic) or testing if they agree with external species distribution information, for example from http://www.iucn.org (vertebrates; extrinsic) or https://wcsp.science.kew.org/ (selected seed plant families; extrinsic) can further corroborate the accuracy of a record’s geographic referencing. If such tests are included, it is essential to account for the sampling year, in particular for older records, since the names of political entities may change and the ranges of species may shift. Furthermore, while in this study we focused on meta-data and geographic filtering, taxonomic cleaning—the resolution of synonymies and identification of accepted names—is another important part of data curation, but depends on taxon-specific taxonomic backbones and synonymy lists which are not readily available for many groups and often are contradictory within individual taxa.

The impact of filtering on the accuracy of automated conservation assessments

The accuracy of the automated conservation assessment was in the same range as found by previous studies (Nic Lughadha et al., 2019; Zizka et al., 2020). The similar accuracy of the raw and filtered dataset for the automated conservation assessment was surprising, in particular given the EOO and AOO reduction observed in the filtered dataset (Table 4) and the impact of errors on spatial analyses observed in previous studies (Gueta & Carmel, 2016). The robustness of the automated assessment was likely due to the fact that the EOO for most species was large, even after the considerable reduction caused by filtering. This might be caused by the structure of our comparison, which only included species that were evaluated by the IUCN Red List (and not considered as Data Deficient) and at the same time had occurrences recorded in GBIF. Those inclusion criteria are likely to have biased the datasets towards species with large ranges, since generally more data are available for them. The robustness of automated conservation assessments to data quality is encouraging, although these methods are only an approximation (and not replacements) of full IUCN Red List assessments, especially for species with few collection records (Rivers et al., 2011).