Identifying biodiversity knowledge gaps for conserving South Africa’s endemic flora

As a megadiverse country with a rapidly growing population, South Africa is experiencing a biodiversity crisis: natural habitats are being degraded and species are becoming threatened with extinction. In an era of big biodiversity data and limited conservation resources, conservation biologists are challenged to use such data for cost-effective conservation planning. However, while extensive, key genomic and distributional databases remain incomplete and contain biases. Here, we compiled data on the distribution of South Africa’s > 10,000 endemic plant species, and used species distribution modelling to identify regions with climate suitable for supporting high diversity, but which have been poorly sampled. By comparing the match between projected species richness from climate to observed sampling effort, we identify priority areas and taxa for future biodiversity sampling. We reveal evidence for strong geographical and taxonomic sampling biases, indicating that we have still not fully captured the extraordinary diversity of South Africa’s endemic flora. We suggest that these knowledge gaps contribute to the insufficient protection of plant biodiversity within the country—which reflect part of a broader Leopoldean shortfall in conservation data.


Introduction
Life on earth is unevenly distributed and so too is our knowledge about it. In the past two decades, efforts to identify biodiversity knowledge gaps have attracted considerable interest among scientist and conservation practitioners (e.g. Scarascia-Mugnozza et al. 2000;Jarnevich et al. 2006;Costion et al. 2015;Oliveira et al. 2019). This interest has been sparked by the increasing evidence suggesting losses of species due to habitat destruction, overexploitation of natural resources, pollution, the spread of invasive species and climate change (Ceballos et al. 2015). Despite concerted attempts to curb biodiversity loss, global biodiversity is thought to be declining at unprecedented rates (Pereira et al. 2012;Ceballos et al. 2015). Due to limited time, financial resources and human capacity, it is essential that conservation efforts are prioritized effectively to maximize conservation returns. However, efficient conservation planning is hampered by key knowledge gaps (Scott et al. 1993;Jennings 2000;Meyer et al. 2015): the Linnaean shortfall describes the mismatch between the number of described species and the true number of species, the Wallacean shortfall reflects gaps in our knowledge of species distributions, and the Darwinian shortfall refers to our lack of knowledge on the evolutionary relationship among species (Hortal et al. 2015). We here refer to the lack of data to make informed conservation decisions as the Leopoldean shortfall, after Aldo Leopold, in recognition of his contributions to the modern conservation movement.
In many regions, we lack knowledge of species taxonomy (Paton et al. 2008), distributions (Meyer et al. 2016;Sporbert et al. 2019) and genetic diversity (Hoban et al. 2013) due to biases in species collection, lack of funding and research infrastructure, the declining number of taxonomists and complexity in identifying and describing species, and the inadequate training of conservation practitioners to collect and incorporate new types of bioinformatics data into biodiversity databases. As a result, efficient biodiversity conservation is frequently hampered by incomplete and unrepresentative data. Identifying and addressing such biodiversity knowledge gaps is a particular challenge for megadiverse, developing countries, were the economic commitment of governments to biodiversity conservation is frequently far less than what is required (He 2009;Yu 2010;Silveira et al. 2018). South Africa, the third most biologically diverse country globally, and with more than 50 percent of the population living below the poverty line (Statistics South Africa 2019), provides an object lesson.
South Africa's biological diversity is estimated at between 250,000 to 1,000,000 plant and animal species combined (Wynberg 2002). This astounding richness is thought to be due, in part, to the country's climatic diversity, ranging from desert to humid forest climates, and from cool temperate to warm subtropical climates, and varied topography (Cowling and Hilton-Taylor 1994). Over 13,000 of the plant species occurring within South Africa are found nowhere else (Raimondo et al. 2009). Much of this endemic diversity is restricted to the country's three global biodiversity hotspots: the Cape Floristic Region (CFR), the Succulent Karoo, and Maputaland-Pondoland-Albany (Myers et al. 2000). The CFR is the only plant kingdom confined within the borders of a single country, it is also a centre of diversity for endemic mammals, reptile and amphibian (Cowling et al. 2003). This hotspot contains approximately 8200 plant species, 6210 of which are endemic to the region (Raimondo et al. 2009). The Succulent Karoo includes a third of the world's succulent plant species (Brownlie and Wynberg 2001) and is also home to over 2400 endemic plants. The Maputaland-Pondoland-Albany biodiversity hotspot, with approximately 600 tree species, has the highest tree richness of any temperate forest on the planet (Tarrant 2012), and is home to approximately 1900 endemic plants (Raimondo et al. 2009). To safeguard this rich biodiversity, taxonomic, genomic and spatial knowledge gaps need to be identified, and filled.
Various studies have shown that sampling bias and the decline in taxonomic studies present fundamental challenges to biodiversity conservation in South Africa (e.g. Robertson and Barker 2006;Thuiller et al. 2006;Von Staden et al. 2013;Tolley et al. 2019). The perilous state of plant taxonomic research in South Africa has been wellrecognised for over a decade (Victor and Smith 2011). Only 62% of the South African flora has been revised since 1970, 13% of the flora has no revision, and taxonomic data for 25% of the flora is outdated (Von Staden et al. 2013)-a Linnaean shortfall. Locality data from herbarium specimens is curated within the Botanical Database of South Africa (BOD-ATSA; Ranwashe 2015)-the single largest database of species georeferenced occurrences and herbarium voucher data in South Africa, with more than a million specimens (Williams and Crouch 2017). Nonetheless, even basic data on the geographic distributions of many species is lacking. More than a 1000 species do not have any recorded occurrence data, while [ 2600 species have less than five occurrence records-a Wallacean shortfall. Despite rapid advances in sequencing technology, and major sequencing initiatives, such as the DNA barcoding effort (Hebert et al. 2003) led by the International Barcode of Life (iBOL) consortium (https://ibol.org/), supported by the African Centre for DNA Barcoding in South Africa (Bezeng et al. 2017), most species still lack sequence data that would allow them to be placed on the plant tree-of-life-a Darwinian shortfall.
Here, we examine biodiversity data gaps for the endemic flora of South Africa. First, we use species distribution models (SDMs) to identify areas with broad climatic conditions suitable for (encompass the climate niche envelope of) many species. Importantly, we do not expect our projected occurrences to necessarily reflect species realised distributions. Indeed, we would expect species not to occur within much of their area of projected occurrence, even within highly suitable environments, as fine scale edaphic landscape features and other environmental requirements likely occur in only a small fraction of suitable climate space, and our grid cells additionally encompass urban areas, farmland, and other highly transformed habitat. Our models should thus be viewed as defining the coarse grained area of extent with suitable climate for species, and not their actual area of occupancy (for a conceptually similar approach see Engelbrecht et al. 2016). Second, we contrast our SDM projections with the geographic distribution of biodiversity sampling effort, and identify locations and taxa that are poorly represented in existing biodiversity database. Third, we merge taxonomic and spatial data with information from GenBank/ NCBI (https://www.ncbi.nlm.nih.gov/genbank) to explore unevenness in the representation of genetic data across space and species. We hope our study will be useful in helping guide future sampling efforts, and contribute to addressing the shortfall in critical biodiversity data necessary for informed conservation decision making.

Endemic flora and occurrence data
We compiled a database of South Africa's endemic flora representing 175 families, 1061 genera, 10,965 species, and 762,655 distribution records (413,491 unique species 9 location occurrence records) of bryophytes, pteridophytes, gymnosperms, and angiosperms, using the Checklist of South African plants (Germishuizen et al. 2006) to crosscheck species endemism. Species taxonomy was standardized using The Plant List (www. plantlist.org) and the Angiosperm Phylogeny Group (APG IV: Chase et al. 2016) as taxonomic authorities for plant names and families respectively, and then matched to geographic occurrence records from the Botanical Database of South Africa (BODATSA). Records for subspecies and varieties were merged into single species.
Herbarium specimens in BODATSA (specimens for endemic species, n = 292,393) are georeferenced to quarter degrees squares (QDS), approximately 25 km 9 25 km, representing the approximate spatial precision of historical records. Recent and more precise plant occurrence records were sourced from the Protea Atlas Protect (n = 245,407), the Custodians of Endangered Wildflowers (CREW) programme, (n = 15,244) and the National Vegetation Map project (VEGMAP; n = 135,165), ACOCKS database (n = 66,068) and smaller projects (n = 138,957). Data were obtained from the South African National Biodiversity Institute (SANBI), with the assistance of LW Powrie. Several species not included in these databases (n = 1054) were manually added by georeferencing location data from protologues, Floras, and revisions, and extracting latitude-longitude coordinates from Google Earth (https://www.earth.google.com).

Genetic data
To assess taxonomic bias in DNA sequence data, we queried GenBank (https://www.ncbi. nlm.nih.gov)-a comprehensive public database containing nucleotides for approximately 260,000 formally described species (Benson et al. 2012)-for records for each of South Africa's endemic species using the package SeqinR (Charif and Lobry 2007), in R version 3.5.2 (R Development Core Team 2006), returning the number of DNA sequence available for each species.

Species Distribution Modelling
Nineteen raster-based bioclimatic variables were sourced from the WorldClim database (https://worldclim.org; Hijmans et al. 2005) at a spatial resolution of 10 arc minutes, approximating the resolution of the species occurrence data. These bioclimatic variables were then used as environmental predictors to generate species distribution models (SDMs) using an ensemble forecast (Hijmans and Elith 2013) of three models: generalized linear models (GLMs; Guisan et al. 2002), random forests (RFs; Breiman 2001), and the gradient boosting machines (GBM; Friedman et al. 2000) fitted in R version 3.5.2 (R Development Core Team 2006), using the gbm, lmtest, boot, dismo, and randomForest packages (Hothorn et al. 2019;Ridgeway 2006;Chang and Hanna 2005). These standard modelling approaches use presence-absence data for predicting habitat suitability for species. Since the dataset lacks true absences, pseudo-absences were generated from background data, bounded to the country borders of South Africa. Twenty-five percent of occurrence records were used for testing the model and 75% of occurrence records were used for training the model. Duplicate records were removed to prevent model over-fitting. Ensemble predictions were generated by combining the individual model outputs weighted by the average of their AUC (Area Under the Curve), which gives more weight to the algorithm that gives a better estimation (Ranjitkar et al. 2014;Breiner et al. 2015). Predictions with AUC values \ 0.5 were given a weight of zero. Average predictions were transformed into binary presence-absence maps by applying the threshold that maximizes the sum of actualpositive rate and actual-negative rate (Manel et al. 2001).
Spatial predictions from SDMs are influenced by the number of occurrence points, with accuracy decreasing as they are informed by less data; we, therefore, fit SDMs only to species with five or more occurrences (n = 8295). For species with three or four occurrence points (n = 691), the species range was defined using a convex hull in ArcGIS 10.5 (Esri, CA, USA), following Goldsmith et al. (2016). For species with one or two occurrences (n = 1961), the QDS in which they occurred was considered as their range.

Areas of high climate suitability
To identify areas with broad climate suitability (within the climate envelope) for many species, rasters from each of the output SDMs were stacked in R, using the Mass library (Ripley et al. 2012). Raster values for each summary output were then extracted onto an equal-area 25 km 9 25 km lattice in ArcGIS 10.6 (Esri, CA, USA) to match the resolution of the underlying species occurrence records (QDS) in the BODATSA database. High values cells represent areas with mean climate matching to that within a high number of projected species' distributions, although we would not expect species to fill their climate niche as additional niche factors varying over finer spatial scales, species interactions, and historical contingency all likely influence species' realised distributions. Coastal grid cells with \ 50% of land were excluded because the projected richness for such cells would likely overestimate their true richness.
To generate maps of observed species richness and sampling density, we conducted a spatial join in ArcMap v.10.6 (Esri, CA, USA) to summarise the number of species from occurrence records and total number of occurrence records within each cell of the polygon lattice, used above.

Spatial and taxonomic gaps
We identified potential geographical gaps in taxon sampling effort by quantifying the fraction of observed species from occurrence records relative to projected richness from SDMs-which we refer to as the sampling fraction. As a proxy for sampling intensity, we also generated a density map of documented species coordinates and calculated the fraction of occurrence records relative to projected richness from SDMs-which we refer to as the sampling density. To test for the ''road effect''-the tendency for collections to be greater in sites that are easily accessible-we overlaid a map of national roads on the sampling density map (https://mapcruzin.com/free-south-africa-arcgis-maps-shapefiles.htm).
We further quantified the fraction of the richness of species with genetic data relative to projected richness from SDMs to identify geographical areas which represent potential targets for genetic sampling. We then classified those species with sequence data by their IUCN Red List threat status (https://redlist.sanbi.org; Raimondo et al. 2009) to explore whether threatened species were more likely to lack sequence information.
Last, we examined whether biodiversity data gaps along one axis aligned with data gaps among other axes across space and taxa. First, we evaluated spatial correlation strengths using Pearson's correlation coefficients, adjusting degrees of freedom to account for spatial non-independence among grids cells using Moran's I, as implemented in SAM V.40 (Rangel et al. 2010). Second, we evaluated congruence across taxa using linear regression models constructed in R. For the spatial analyses we evaluated correlations between: (1) sampling fraction-ratio of observed species richness to projected richness-and sampling density-ratio of documented plant records to projected richness; (2) sampling fraction and ratio of species with genetic data; and (3) the residuals from (1) and (2) above. For the taxonomic analysis we evaluated correlations between: (1) the number of endemic species per family and the number of georeferenced occurrence records for endemic species per family; (2) the number of endemic species per family and the number of sequences for endemic species per family; and (3) the number of georeferenced occurrence records for endemic species per family and the number of sequences for endemic species per family.

Results
Species richness, sampling effort and the spatial distribution of biodiversity data gaps Spatial patterns of observed and projected endemic richness were mapped onto 1790 grid cells (25 km 9 25 km) by (1) stacking projections from SDMs, and (2) recording observed species richness from occurrence records (Fig. 1a, b). We show that 4% of the country has not been sampled for endemics-70 grid cells with no endemic species recorded-while our SDMs indicate that all cells enclose climate space that falls within the climatic niche space of at least 69 species (although we would not necessarily expect this to translate into realised richness). The overall spatial structure in relative richness is similar for SDMs and observed records (Fig. S1 Supplementary Information: r 2 = 0.7, p \ 0.05), but total grid cell richness is much lower in the latter (the highest number of observed species per grid cell is 1838, while the equivalent projected richness from SDMs is 5303 species per grid cell).
The projected endemic richness from SDMs matches well to current understanding of species diversity across South African biomes (Fig. 1). The Maputo-Pondoland-Albany biodiversity hotspot, the Soutpansberg, and the Wolkberg centres of endemism coincide with areas of high projected endemic-richness. The Savanna Biome is the largest and one of the most species-rich biomes in South Africa; here we show projected endemic richness is greatest in the northeast and lower in the northwest (Fig. 1a). The Succulent Karoo Biome, is synonymous with the Succulent Karoo biodiversity hotspot, and has projected grid cell endemic richness peaking at 4908 species (mean projected endemic richness: 1896, range = 136-4908; Fig. 1a). The Albany-Thicket and the Indian Ocean Coastal Belt biomes encompass the grid cells with the highest projected endemic richness within the Maputo-Pondaland-Albany biodiversity hotspot (mean projected endemic richness: 952, range 296-4661; Fig. 1a). The biome with the highest overall mean projected richness is the Fynbos. This biome falls within the Cape Floristic Region biodiversity hotspot, where projected grid cell endemic richness peaks at 5,303 species (mean projected endemic richness: 2765 range 526-5303; Fig. 1a). The Grassland and Nama Karoo biomes have the lowest mean projected endemic richness, however, even within these biomes some cells have high projected richness, for example, those that coincide with the with the Maputo-Pondoland-Albany hotspots, the Drakensberg escarpment, and the Sekhukhuneland and Barberton regional centres of endemism.
Despite the overall strong correlation between observed (Fig. 1b) and projected endemic richness (Fig. 1a), there is spatial structure in the residuals of the relationship (Fig. 1c)-the sampling fraction. For example, much of the Nama-Karoo and Savanna biomes have low sampling fraction (see also a conceptually similar analysis by Robertson and Barker 2006), whereas the sampling faction is much greater in the generally speciesrich Fynbos and Succulent-Karroo, and species-poor Grassland biomes.
The Fynbos Biome and Gauteng Province are the most intensively sampled regions (Fig. 2a)-estimated from the total number of occurrence records. However, we find that there is a generally higher sampling density-ratio of documented plant occurrence records to predicted endemic species richness-in areas near roads (mean sampling density = 0.16 and 0.04 for grid cells with road and grid cells without roads, respectively; t = 20.27, p \ 0.01; Fig. 2b). There is also a strong correlation between sampling density and sampling fraction-the ratio of observed endemic species richness from occurrence records to projected endemic richness from SDMs (Pearson's r = 0.862, d.f. = 269, p \ 0.001, after adjusting degrees of freedom to correct for spatial non-independence). For example, we again identify much of the Nama-Karoo and part of the Savanna Biome as undersampled (low sampling density), whereas the Fynbos and much of the Grassland Biome have higher sampling density, with a greater number of documented occurrence records per species. As an important biodiversity hotspot, the relatively low sampling density in the Maputo-Pondoland-Albany region, which peaks at around only 0.2, is notable. . We do not necessarily expect true richness to match to predicted richness as it is likely that many fine scale process not captured in our SDMs limit species realised distributions; nonetheless, there is an obviously high correlation between observed and predicted richness (compare maps a and b), and we suggest model predictions are informative for identifying potential sampling gaps. b Observed endemic species richness from occurrence records (sum of the unique species records in each cell). Cells are shaded using a graduated colour scheme: red = high species richness, blue = low species richness. c Sampling fraction: ratio of observed endemic species richness to predicted endemic species richness -cells with low sampling fraction indicate areas with climate suitable for supporting high endemic richness, but for which there are relatively few occurrence records. Red = high species sampling, blue = low species sampling. Grid cell resolution a-c 25 km 9 25 km. d Map of the biomes of South Africa, after Mucina and Rutherford (2006); data from https://bgis.sanbi.org/ SpatialDataset In general, the sampling of genetic data for endemic species is poor relative to the projected richness; only 5% (80 of 1790) of grid cells have more than 50% of projected species with sequence data (Fig. 3a), with the interior of the country a notable 'coldspot' of genetic sampling (Fig. 3b). Several areas along the South African border have been relatively well sampled for genetic data, and these might represent lower genetic sampling priorities. Notably, there is no correlation between areas in need of better taxonomic sampling (low sampling fraction) and areas in need of genetic sampling (Pearson's r = -0.004, d.f. = 111, p = 0.963, adjusted degrees of freedom). Fynbos biodiversity hotspot has been relatively well sampled, while proportional sampling density in the species-rich Maputo-Pondoland-Albany hotspot peaks at around 0.2. Red = high sampling density richness, blue = low sampling density Fig. 3 Shortfalls in our knowledge of the sampling of genetic data for endemic plants: a projected richness of endemic species with genetic data (species with at least one sequence in GenBank) and b Sampling of DNA sequences (proportion of species with at least one sequence in GenBank relative to total endemic species richness per cell). Cells are shaded using a graduated colour scheme: blue cells indicate poorer genetic sampling of taxa, while red cells indicate higher genetic sampling. There is high sampling effort needed in the interior and northern regions of the country, while species-poor, these regions have been largely overlooked by past genetic sampling efforts. The Fynbos and parts of the Karoo and Albany thicket appear to be better sampled

Taxonomic and phylogenetic distribution of biodiversity data gaps
Our database includes plants from 175 families and 1061 genera, with large variation in the taxonomic and phylogenetic distribution of biodiversity data (Fig. 4). The ten families with the highest number of endemic species within South Africa are listed in Table S1 (Supplementary Information). These ten families comprise 61% of all endemic species in the database. Forty families are represented by just one endemic species. Families with the highest number of unique species 9 location occurrence records include Proteaceae, Asteraceae, and Fabaceae (Table S2: Supplementary Information). The top ten families by sampling (Table S2) comprise 69% of all the occurrence records in the database. Three families are represented by a single record in our analyses, all are monotypic (Ditrichaceae, Potamogetonaceae and Thelypteridaceae). In general, more species rich families have been better sampled than less species rich families (Fig. S2 Supplementary Information: r 2 = 0.78; slope = 1.11; p-value \ 0.05), as would be expected if all species had an equal probability of being sampled. However, there is some notable variation in sampling intensity across families. For example, Anemiaceae has only one endemic species but is represented by 508 records; perhaps of more conservation concern are the several families that are relatively under-sampled. There is large taxonomic variance in the availability of DNA sequence data, indicating a bias in the species targeted for sequencing (Fig. 4). Only 36% of endemic species have DNA sequences available on GenBank, and less species-rich families tend to be sampled less, as might be expected, although the strength of the correlation is not particularly high (Fig. S3 Supplementary Information: r 2 = 0.60; slope = 1.11; p-value \ 0.05). Families without a GenBank record for endemic species include Fissidentaceae (3 endemic species), Lythraceae (4 endemic species), and Pylaisiadelphaceae (4 endemic species), as well as some families with higher endemic richness (e.g. Ricciaceae [26 endemic species] and Celestraceae [33 endemic species]). Nonetheless, some moderately species poor families have also been intensively sampled for genetic data. For example, Zamiaceae, with an endemic richness of 29 species has, at the time of writing, 995 sequences on GenBank. For the set of species with DNA sequences, 77% are of low conservation concern, 20% are threatened and 3% are data deficient. In contrast, for the set of species lacking DNA sequences, 68% are of low conservation concern, 20% are threatened and 12% are data deficient. There is no significant relationship between families that lack sequence data and families that lack georeferenced occurrence records ( Fig S4 Supplementary Information: r 2 = 0.004; slope = 0; p-value = 0.84, from the linear regression of the number of sequence per family against and number of coordinates per family), hence these two gaps also need to be targeted separately. A list of top families ranked by number of sequences is provided in Table S3.

Discussion
South Africa is characterized by an interior with wide-ranging plains and plateaux, bounded by remarkable mountain landscapes and undulating coastal plains. Endemic-rich areas are found in a virtually continuous arc around the Great Escarpment, mainly within the three global biodiversity hotspots-the Succulent Karoo, Cape Floristic Region (CFR) and Maputo-Pondoland-Albany-and the Barberton, Sekhukhune, Soutpansberg and Wolkberg, centres of endemism. The southwestern part of the Fynbos is the most speciesrich area in the country. Past climatic conditions are thought to be one of the main factors contributing to the high species richness in this region. During the Pleistocene, an epoch of high glacial-interglacial climate variability, the southwestern Cape remained relatively stable (Sniderman et al. 2013), while the eastern part of the country experienced greater climatic fluctuations. This resulted in higher speciation and lower extinction rates in the southwest relative to the southeast, leading to a greater accumulation of species over time in the former (Cowling and Lombard 2002;Cowling et al. 2004). As a consequence of these evolutionary dynamics, a high proportion of the native flora is composed of range restricted endemics (Goldblatt 1997), many of which are vulnerable to extinction (https:// redlist.sanbi.org/) yet remain under-researched and poorly represented in biodiversity databases. To adequately protect this rich diversity, we must address these biodiversity data gaps. Here, focussing on endemic plants, we have explored the distribution of biodiversity data across space and phylogeny to identify regions and taxa that have been under-sampled as a guide to help future data gathering efforts.

The Wallacean and Linnean shortfalls
The observed richness of species obtained from occurrence records differs importantly to the richness estimated from species distribution models. Species distribution models trained on observed occurrence date and broad-scale climate variables will likely overestimate species realised distributions, which are shaped by various additional processes and more fine scale niche partitioning (e.g. see Dubois et al. 2013). Our estimates of projected richness should thus be viewed as defining the coarse grained area of extent with suitable climate for species, and not their actual area of occupancy (see Elith and Leathwick 2009, for related discussion). Nonetheless, variation in the ratio of observed and projected richness highlights potential geographical biases in sampling effort. For example, the number of recorded species from occurrence records in the Nama-Karoo and Savanna biomes appears to be lower than that expected from projected species distribution models relative to observations across the Fynbos and Grassland biomes. These discrepancies are informative as they allow us to identify potential sampling gaps-areas where increased sampling effort is needed to fully characterise species geographic distributions-and thus help address the Wallacean shortfall.
We suggest important areas for future sampling include much of the Nama-Karoo, and some of the Savanna Biome, as highlighted above, and also the Maputo-Pondoland-Albany biodiversity hotspots. One reason for apparent under-sampling in these regions may be that there are fewer roads and centres of research nearby (Reddy and Dávalos 2003). Our results show that areas near roads are better sampled, likely because they are more accessible (Daru et al. 2018;Meyer et al. 2016). For example, the province of Gauteng has been relatively well-sampled, perhaps reflecting its status as the economic hub of South Africa, with a high density of roads and research institutes.
In comparison with the Nama-Karoo and the Maputo-Pondoland-Albany hotspots, the CFR has been relatively well sampled, and it is one of the regions with the greatest density of species records in the country. The CFR is recognised as a distinct floristic kingdom within the Mediterranean biome-the most threatened biome in the world (Cox and Underwood 2011)-and has thus attracted national and international research attention. Several non-government conservation agencies, including the World Wildlife Fund (WWF), Wildlife Protection Society of South Africa (WESSA), Earth Life, and CAPE, have offices located in the region, and support research on and conservation of the Fynbos flora. In addition, government programs, such as the Millennium Seed Bank (MSB) and the Custodian of Rare and Endangered Wildflowers (CREW), make use of volunteers and citizen scientists to sample remnants of natural vegetation in the region. While the considerable research effort focussed on the CFR is, of course, very welcome, other speciesrich regions require equal attention.
In the past decade, plant collection efforts have decreased substantially in the country, reflected by the 14,000 plant collection records between 2006 and 2010 in comparison to the 94,000 records between 1976 and 1980 (Williams and Crouch 2017). We show that there is positive spatial correlation between areas of low sampling fraction-ratio of observed species richness to predicted species richness-and areas of low sampling density-ratio of documented plant records to predicted species richness-indicating that we are missing records for much of the diversity in areas that have been poorly sampled taxonomically, and raising the possibility that we may also be missing undescribed species in these areas (there is no evidence that the rate of new species description is declining over time; Victor et al. 2015)-the Linnaean shortfall.
Bias in plant collection has not only been spatial, but also taxonomic. Large families have been better sampled than smaller ones. Societal interest also plays a role in the sampling of taxa: more charismatic species are more likely to attract funds and research attention (Wilson et al. 2007;Troudet et al. 2017). For example, Proteaceae-a large family of significant agricultural and horticultural value-is the most intensively sampled family in our database, and has been the target of large-scale ecological research through the Protea Atlas Project (Rebelo 1993). It is also possible that smaller families occur in regions that have been less well sampled, or are more likely to be comprised of narrow ranged endemics, thus making them less likely to be included in general biodiversity surveys (Eberhard et al. 2009;Hemp 2006). However, there is large variation in sampling intensity across families independent from species richness, and other idiosyncratic or historical explanations likely contribute to taxonomic differences in sampling representation.

The Darwinian shortfall
Species richness has been used as an index for classifying important areas of biodiversity for decades (Pimm et al. 2014;Veach et al. 2017). However, a narrow focus on species may fail to capture genetic and functional diversity. There have been numerous calls to incorporate phylogenetic diversity, as a surrogate for functional or feature diversity, more directly into conservation planning (e.g. Cadotte and Davies 2010;Rolland et al. 2011;Winter et al. 2013;Faith 2015). Phylogenies are important for understanding structural and functional aspects of biodiversity in an evolutionary context, and allow us to assess how the tree of life will be affected by global change (Rolland et al. 2011). However, the use of phylogenetic data in conservation decision making remains a challenge, particularly in developing countries, where genetic data is often scarce or incomplete, and DNA sequencing is costly (Rodrigues and Gaston 2002)-the Darwinian shortfall.
While there is a strong need to gather more genomic data, it must be done efficiently to avoid escalating costs. Optimization strategies for data collection include the targeting of regions for which there is a high probability that data-poor species occur, and the selection of localities were many target species can be found (Parra-Quijano et al. 2012). In this study, we find that only a third of South Africa's endemic species have DNA sequences available, and that IUCN data deficient species are disproportionately under-represented, which makes the incorporation of genetic data into systematic conservation planning in South Africa even more of a challenge. We identify locations with climates suited to supporting high diversity but for which only a small fraction of projected species have sequence data, and suggest these as priority areas for tissue sampling. Species distribution models have been previously used for guiding the collecting genetic data to good effect (Ramírez-Villegas et al. 2010;van Zonneveld et al. 2014;Khoury et al. 2015). Here we show that many locations in the interior of the country have not been well-sampled for genetic data, whereas the exterior of the country has been better sampled, partly reflecting the success of DNA barcoding initiatives across the three biodiversity hotspots (e.g. see Lahaye et al. 2008;Bezeng et al. 2017;Powell et al. 2018).
On average, species-rich families have been better sampled for genetic data than species-poor families. Zamiaceae (a relatively small family) is an exception, with a high number of sequences per species. This family has been the subject of intense research, and its deep evolutionary history has made it a model taxon for studies on plant evolution and biogeography (e.g. Gregory and Chemnick 2004;Calonje et al. 2019). In addition, several species within the family are valuable medicinal, ornamental and commercial plants, attracting increased research effort (e.g. Ndawonde et al. 2007;Ravele and Makhado 2010;Cousins et al. 2011).
Genetic data is not only important for ecological and evolution studies, but is increasingly a fundamental component of taxonomy. Currently 611 endemic species are listed as data deficient by the IUCN as a consequence of taxonomic uncertainty. DNA sequencing and phylogenetic studies could assist in addressing this issue, and thus facilitate appropriate IUCN Red Listing, which might provide increased conservation protection. A further 291 species are data deficient due to lack of ecological information, and DNA sequence data could help here also. Genetic data can be used to predict the conservation status of a species, for example, via phylogenetic imputation of traits or extinction risk (Bland et al. 2015;González-del-Pliego et al. 2019). Targeted sequencing efforts could thus help address both the Linnean and Darwinian shortfalls. However, there is a no significant correlation between areas that need sampling for occurrence data (Wallacean shortfall) and areas that need sampling for genetic data.

The Leopoldean shortfall
In this study, we have identified important biodiversity knowledge gaps. Strong geographical and taxonomic sampling biases indicate that we have not fully captured the extraordinary diversity of South Africa's endemic Flora in biodiversity databases. We suggest that these conservation data gaps represent a Leopoldean shortfall-contributing to the insufficient protection of plant biodiversity within the country. We identify areas and taxa that are in need of increased research attention. However, we show that the Wallacean and the Darwinian shortfalls need to be targeted separately, as gaps in our knowledge of species' distributions do not overlap with gaps in our knowledge of species' genomes. One way to help address these shortfalls is for scientist to reach out to non-professional to assist in data collection, as exemplified by the Protea Atlas Project (Rebelo 1993). Most importantly, there is a renewed call for scientists across the globe to make use of emerging and new technologies such as artificial intelligence, image-recognition algorithms, remote sensing, metagenomics etc. to collect data, identify, locate, and track species (see Pimm et al. 2015). By making use of these innovative and non-invasive approaches, the research community will be able to better address the data shortfalls we highlight here, and contribute to protecting and conserving biodiversity.