Big data of tree species distributions: how big and how good?

Species selection and occurrence records

We selected species from the world tree species checklist (GlobalTreeSearch; GTS; Beech et al. 2017) for use in our analysis. In this list, a species is considered to have a tree growth habit when it is “a woody plant with usually a single stem growing to a height of at least two metres, or if multi-stemmed, then at least one vertical stem five centimetres in diameter at breast height”— a definition from the Global Tree Specialist Group of the International Union for Conservation of Nature. The tree species checklist is available through an online searching engine, which provide tree species list by country (http://www.bgci.org/global_tree_search.php, accessed June 2017). To avoid potential taxonomic issues, we used the Taxonomic Name Resolution Service (TNRS) online tool (Boyle et al. 2013). This tool provides a method for species name standardization, resolving issues related to taxonomic semantic heterogeneity. From the initial species checklist, we selected species that TNRS rendered an ‘accepted name’ taxonomic status. In order to maximize the inclusion of species, we also included species which name was classified as ‘no opinon’ in the TNRS. The final set includes 58,100 species.

We collected tree species occurrence data from five major sources of widely used, easy to access and publicly available data of species occurrences: the Global Biodiversity Information Facility (GBIF; http://www.gbif.org), the public domain Botanical Information and Ecological Network v.3 (BIEN; http://bien.nceas.ucsb.edu/bien/), Latin American Seasonally Dry Tropical Forest Floristic Network (DRYFLOR; http://www.dryflor.info/; Banda et al. 2016), RAINBIO database (http://rainbio.cesab.org/; Gilles et al. 2016) and the Atlas of Living Australia (ALA; http://www.ala.org.au/). These databases were selected because they aggregate many different sources of species occurrence records, cover a wide range of ecosystems, and constitute the main sources of information used in biodiversity studies. GBIF was accessed using the rgbif package (Chamberlain 2017), BIEN data was accessed using the BIEN package (Maitner 2017), and ALA data was accessed using the ALA4R package (Raymond et al. 2017) developed in R v3.4.1(R Core Team 2017). DRYFLOR and RAINBIO were accessed and downloaded online.

Quality assessment of occurrence records

We used a workflow to assess the quality, quantity and coverage of occurrence data for each species (Table 1; Fig. 1). The quality control workflow follows a hierarchical procedure and labels each occurrence record with an alphabetic code from A to H (Fig. 1). These labels broadly categorize positional as well as biological characteristics of the record and are fit for the purpose of big data macro-ecological analysis. That is, some labels reflect the precision of the occurrence data with respect to the resolution of the environmental spatial layers used for analysis. Therefore, a label for a given record may vary if input environmental data used for the workflow is different. In this study, we used worldclim v. 1.4, a collection of climate environmental layers at 0.5 arcsec spatial resolution (Hijmans et al. 2005).

Label	Label name	Label information (I) and potential actions to be developed (A)
H	Missing coordinates	(I) No detected coordinates.
		(A) Trace back record and assign coordinates.
G	Duplicated records	(I) If record is duplicated within the environmental grid cell, it may give information of sampling effort.
F	Unknown range	(I) Not known range.
		(A) Double check country-level databases and invasive registries of countries where the record occurs. If present, update database.
		(A) Re-check common coordinate errors (Yesson et al. 2007).
E	Missing environmental information or unlikely environment (botanic garden)	(A) Check suitability of spatial layers.
		(A) Confirm botanic garden location.
		(A) Re-check common coordinate errors (Yesson et al. 2007).
D	Geographic coordinate issues and environment issues	(A) Re-check common coordinate errors (Yesson et al. 2007).
		(A) Check values in environmental layers.
C	Geographic coordinate issues	(A) Re-check common coordinate errors.
B	Environmental space issues	(A) Check values in environmental layers.
A	No issues detected	(A) Send a ‘thank you’ email to the database custodian.
AA	High precision	(A) Send a ‘thank you’ email to the database custodian.
AAA	High precision and low environment uncertainty	(A) Send a ‘thank you’ email to the database custodian.

The workflow starts with a table of species presence records in latitude-longitude WGS84 geographic coordinate system. This table results from the integration of the sources listed above with the following fields: ‘x’ (longitude), ‘y’ (latitude), ‘elevation’ , ‘country’ , and ‘locality’. In step 1, we identify those species for which we have no spatial data, and we categorize them as quality label H. In step 2, we identified duplicate records. This step is important because data aggregators may use the same sources of information and data integration and subsequent analysis may lead to pseudo-replication or overestimation of sampling intensity. Two types of duplicates were identified and flagged: records with identical coordinates –true duplicates—and those records that fall in the same grid cell of the environmental spatial layers – geographical duplicates. These records where assigned the quality label G. Subsequently, in step 3, the workflow performs an environmental congruence analysis: it checks that occurrence records are not located in the sea or lakes. Due to accuracy of presence records or the environmental data used, some tree species occurrences located in shorelines may wrongly be assigned to such environments. To avoid such spurious effects we use the function nearestcell in package biogeo in R (Robertson et al. 2016), which checks whether neighboring environmental cells have congruent (e.g. terrestrial) environments. If that is the case, the function assigns new coordinates to the record and creates a new field with the names ‘x.orginal’ and ‘y.original’ to keep track of the transformation. Records with transformed coordinates are evaluated again for duplicates, as outlined previously in step 2.

In step 4, we identified potential geolocation issues based on known species ranges. We used the GTS database (Beech et al. 2017) to identify countries where the target species is considered native. Additionally, we identified countries where the species is considered introduced and naturalized through a database merging the global invasive species database (GISD; Invasive Species Specialist Group ISSG 2017) and the global register of introduced and invasive species (GRIIS; Invasive Species Specialist Group ISSG 2017). The latter provide a conservative assessment of the naturalized range, as areas where the species only has sparse naturalized occurrences may not be registered. If the location of the record did not fall in the combined list of countries in which a species is known to occur, the record was assigned a quality label E. Country spatial layer was obtained from the Global Administrative database v.2.8 (GADM 2015).

The six independent tests outlined above allow classification into five quality categories (A to E; highest to lowest). Label E is assigned to records for which there was not environmental information (analysis 1) or the occurrence was likely to be in a botanical garden (analysis 2). Label D is assigned to records with at least one issue in the environmental (analysis 3–5) and at least one issue in the geographical space (analysis 6), thus compromising potential macroecological analysis. Label B indicates records with issues only in geographical space (analysis 3–5), whereas label C was assigned to records with potential issues in the environmental space (analysis 6). Label A was assigned for records with no apparent issues.

Finally, we further differentiated several categories among the highest quality records (Label A). Specifically, we add two more high-quality categories: AA and AAA. These labels determine the geographic precision of the data and may be especially useful in climate-distribution analysis. Precision was determined by the location of the record with respect to the gridded environmental data. Low precision records are identified when presence record was located exactly at the top-left corner or center of the spatial resolution of a grid cell (Robertson et al. 2016). AA label is assigned to already classified A records with good environmental precision Subsequently, AAA label was assigned to already classified AA records which had suitable altitudinal precision (100 m). This is important in locations with topographic heterogeneity, where climate may differ greatly within a coarse-resolution grid cell of a climate spatial layer. We compared the recorded altitude of the presence to the elevation derived from digital elevation model used to obtain the environmental variables (Hijmans et al. 2005). If mismatch was less than 250 m, we reassigned AA record to an AAA record.

From scrubbing to profiling and corrections

The filters applied in our workflow reflect that, for trees, major barriers to classifying high-quality records were due to the geographical space analysis (label B) as well as the range of the species (label F). These two labels held 1,212,822 and 859,433 records, respectively.

Label B reflects issues related to the geographical coordinates. That is, records within the environmental space of the species, but found in either the country centroid, the capital centroid, highly urbanized areas, or areas outside a typical alpha-shape extent of occupancy. Whether to include these records in an SDM or not may require further scrutiny and will largely depend on the question being addressed. These records, even if not correct, are unlikely to strongly bias estimates of the environmental space, although that will depend on total species sample size and model complexity (Merow et al. 2014). For example, such records would have little to no effect on delimiting range boundaries (depending on the modeling algorithm used) although they may have stronger effects on continuous estimates of habitat suitability. Sample biases may inflate the number of records under label B, despite the record being correct. For example, for a given species, sampling in two different countries may be very different and records in countries with lower sample intensity may appear as a geographical outlier. Therefore, we suggest that records under such profile (label B) should be considered in species with low sample sizes when the record is identified to be outside the alpha hull. In such cases, the risk of inclusion may be lower than the benefit of being able to fit an SDM with higher number of occurrences.

Issues with coordinates have long been discussed in occurrences (Yesson et al. 2007; Anderson et al. 2016) and new tools have been recently developed to correct them. This is of utmost importance for rare species or for species in scarcely sampled locations. Robertson et al. (2016) developed an R package capable of implementing alternative geographic coordinates for a sample based on common errors encountered in GBIF. Errors like substitution between latitude and longitude or wrong hemisphere recording are some of these common errors. An automated and scalable form of this would qualitatively advance the use of such records, currently being filtered (or scrubbed) from analysis.

Great care should be taken when correcting or identifying such geographical issues. We use alpha-shapes because they have been used to capture extent of occurrences and may be preferred over other methods to delineate species range boundaries (García-Roselló et al. 2015). However, the method may be misleading for rare species or may tend to identify geographical outliers for rare species, even though these may be widely distributed (Zizka et al. 2017).

Unknown status of a species in a country (native, invasive, naturalized) was another label (F) that led to profiling a large amount of occurrences (859,433). Generally, this filter is imposed to account for the fact that, in some cases, occurrence records may have coordinates in countries where the collection is located, rather than where the specimen was collected. We acknowledge that this filter may be overly conservative, however it remains useful as a globally comprehensive registry of native, naturalized and invaded country-level ranges are lacking. We integrated global registries and species checklists (GTS, GIISD, GRIIS), but these are unlikely to be complete, and are under constant updates and may not cover other important sources, like alien species used in plantations. In addition, while checklists are key political instruments for conservation, they may be of limited use for quality control, especially in the case of large countries or countries with complex geographies. For instance, archipelago-countries may be composed of species endemic to an island and not necessarily to others; or large countries covering large variation in climates (e.g. Brasil, USA, China) may not be able to capture erroneous records that are endemic to some biomes, but not others. Current biodiversity informatics is performing large efforts and key developments given the need for plant range status information to increase the quality of the data. For instance, developing curated lists of native ranges of species (http://bien.nceas.ucsb.edu/bien/tools/nsr/).

The workflow developed in this study is not only aimed at filtering data, but rather at emphasizing and flagging errors and potentially increasing discoverability – the degree to find information associated with the occurrence data (Table 2). For instance, a record with a profile F may help cross-check species-country checklists. We expect that profiling may enhance the quality of species occurrence data but may also be useful when implementing workflows for geo-correction.

Collaboration, networks and mobilization in the forest big data era

The data aggregators used here include many different sources of information and researchers that have participated in this large collaborative effort. Yet, our geographical coverage analysis shows that geographical data gaps are yet to be filled for tree species (Fig. 4). This was somewhat surprising because trees are among the best studied groups of organisms in the world. Issues related to funding for data production and mobilization (Bradley et al. 2014; Nowogrodzki 2016) and proper acknowledgement of investigators still need to be fully resolved to ‘uncover’ data for species distributions (Costello et al. 2014; Franklin et al. 2017). We encourage scientists to acknowledge the effort of those scientists and data infrastructures that support high-quality data (Table 2), but we are aware that a full system recognition of data work recognition is still to be developed. Scientific networks have increased the availability of some regions and to some extent under-sampled forests, while allowing recognition (co-authorship) of the network participants. However, these tend to be restricted to forest types of biomes of the research focused group (e.g., Amazon tree Network, etc.).

In the case of trees, national forest inventory programs offer good opportunities for expansion of tree occurrence data. Assessment of wood resources has historically been performed in many countries (Vidal et al. 2016) and may offer good temporal and geographical coverage for many forest species. In addition, they may lack some of the typical spatial biases present in museum records collected rather opportunistically (Pyke and Ehrlich 2010). Some forest inventories are readily incorporated in the big data aggregators. For instance, BIEN3 incorporates the United States of America Forest Inventory and Analysis (FIA) data, and the Spanish Forest Inventory dataset is incorporated in GBIF. However, a larger mobilization (e.g. incorporation into digital format) of forest inventory data into biodiversity databases is lacking, probably due to difference in the tradition of specimen collections versus assessment of wood resources. Initiatives like the Global Forest Biodiversity Initiative (http://www.gfbinitiative.org/) world forest plot data will be key for improving tree coverage.

It is important to note that mobilization of data may increase geographical biases, sometimes in surprising ways. For instance, Yang et al. (2016) found that vascular plants were relatively well sampled in mountains in China, but densely-populated areas were under-sampled. It is likely that forest inventories could increase geographical bias, towards countries with certain economic status. That said, new techniques to overcome such biases are in place for overcoming such unintended consequences of data mobilization (Phillips et al. 2009; Merow et al. 2016). Therefore, the issue of data sparsity may be more critical than the biases incurred when mobilizing data. Under this context, forest inventories and capacity building performed by the FAO-forestry in certain regions is an important effort that could significantly contribute to global biodiversity datasets.