Afromontane Forest Diversity and the Role of Grassland-Forest Transition in Tree Species Distribution

Local factors can play an important role in defining tree species distributions in species rich tropical forests. To what extent the same applies to relatively small, species poor West African montane forests is unknown. Here, forests survive in a grassland matrix and fire has played a key role in their spatial and temporal dynamics since the Miocene. To what extent these dynamics influence local species distributions, as compared with other environmental variables such as altitude and moisture remain unknown. Here, we use data from the 20.28 ha montane forest plot in Ngel Nyaki Forest Reserve, South-East Nigeria to explore these questions. The plot features a gradient from grassland to core forest, with significant edges. Within the plot, we determined tree stand structure and species diversity and identified all trees ≥1 cm in diameter. We recorded species guild (pioneer vs. shade tolerant), seed size, and dispersal mode. We analyzed and identified to what extent species showed a preference for forest edges/grasslands or core forest. Similarly, we looked for associations with elevation, distance to streams and forest versus grassland. We recorded 41,031 individuals belonging to 105 morphospecies in 87 genera and 47 families. Around 40% of all tree species, and 50% of the abundant species, showed a clear preference for either the edge/grassland habitat or the forest core. However, we found no obvious association between species guild, seed size or dispersal mode, and distance to edge, so what leads to this sorting remains unclear. Few species distributions were influenced by distance to streams or altitude.


Introduction
Afromontane forests are widespread across the African highlands [1][2][3]. Typically occurring above 1500 m in elevation, they extend from the Arabian Peninsula south along the rift to the Drakensberg Mountains in the east. In Western Africa, they are restricted to the Cameroon volcanic line and the Guinea highlands [1,4] (Figure 1). This disjunct distribution is ancient, pre-dating the Miocene doming that resulted in the Rift Valleys and the aridification of the East African interior [5]. Despite this distribution, Afromontane forests have a distinctive fauna and flora and harbor a high proportion of endemic species [6][7][8]. They have substantial carbon stocks [9], provide watershed protection [10]

Study Area and Study Site
The study was conducted in the 260 × 780 m (20.28 hectares) Ngel Nyaki Forest Dynamics Plot, part of the Smithsonian Forest Global Earth Observatory network (Forest-GEO). The plot coordinates are 07 • 04 05" N in latitude and 11 • 03 24" E in longitude, with an elevation range of 1588 m to 1690 m. The plot is within the 5.7 km 2 Ngel Nyaki forest, located within the 46 km 2 Ngel Nyaki Forest Reserve on the Mambilla Plateau, Taraba State [27,28] (Figure 1). The Nigerian Montane Forest Project (NMFP) [29] is based on the edge of the forest and collects regular weather data. The mean annual rainfall is Diversity 2020, 12, 30 3 of 19 1800 mm, with most rain falling between April and October and a six-month dry season. The mean annual temperature is 19 • C and the monthly mean maximum and minimum temperatures for the wet and dry seasons are 25.6 and 15.4 • C, and 28.1 and 15.5 • C, respectively (NMFP weather data). The soil in Ngel Nyaki forest is volcanic and has a high clay content and a pH ranging between 6 and 6.5 [27]. Despite its relatively high nutrient content, the forest soil is easily leached.
Ngel Nyaki forest is floristically diverse and contains six International Union for the Conservation of Nature (IUCN) Red-Listed tree species including Entandophragma angolense (Meliaceae), Lovoa trichiliodes (Meliaceae), Eugenia gilgii (Myrtaceae), Dombeya ledermannii (Malvaceae), Prunus africana (Rosaceae), and Khaya grandifoliola (Meliaceae) [30]. It is a Birdlife International Important Bird Area with several species of large-gaped birds such as hornbills (Tockus alboterminatus) and turacos (Tauraco spp.). Primate species include the endangered Nigeria-Cameroon chimpanzee (Pan troglodytes ssp. ellioti), the putty-nosed monkey (Cercopithecus nictitans), and the tantalus monkey (Chlorocebus tantalus). Smaller mammals include civet cats (Civettictis civetta) and rodents such as the African giant pouched rat (Cricetomys spp.). While the forest is a State Forest Reserve and therefore protected from hunting and grazing, in practice there is very little protection. To what extent past human disturbance has influenced species distribution is unclear. Forest edges have been farmed on the lower slopes of the forest and cattle have damaged a substantial proportion of the reserve. Within the plot however, there is little evidence of disturbance other than past cattle incursions at the edge.
Diversity 2020, 12, x FOR PEER REVIEW 3 of 19 dry season. The mean annual temperature is 19 °C and the monthly mean maximum and minimum temperatures for the wet and dry seasons are 25.6 and 15.4 °C, and 28.1 and 15.5 °C, respectively (NMFP weather data). The soil in Ngel Nyaki forest is volcanic and has a high clay content and a ranging between 6 and 6.5 [27]. Despite its relatively high nutrient content, the forest soil is easily leached. Ngel Nyaki forest is floristically diverse and contains six International Union for the Conservation of Nature (IUCN) Red-Listed tree species including Entandophragma angolense (Meliaceae), Lovoa trichiliodes (Meliaceae), Eugenia gilgii (Myrtaceae), Dombeya ledermannii (Malvaceae), Prunus africana (Rosaceae), and Khaya grandifoliola (Meliaceae) [30]. It is a Birdlife International Important Bird Area with several species of large-gaped birds such as hornbills (Tockus alboterminatus) and turacos (Tauraco spp.). Primate species include the endangered Nigeria-Cameroon chimpanzee (Pan troglodytes ssp. ellioti), the putty-nosed monkey (Cercopithecus nictitans), and the tantalus monkey (Chlorocebus tantalus). Smaller mammals include civet cats (Civettictis civetta) and rodents such as the African giant pouched rat (Cricetomys spp.). While the forest is a State Forest Reserve and therefore protected from hunting and grazing, in practice there is very little protection. To what extent past human disturbance has influenced species distribution is unclear. Forest edges have been farmed on the lower slopes of the forest and cattle have damaged a substantial proportion of the reserve. Within the plot however, there is little evidence of disturbance other than past cattle incursions at the edge.

Census
The plot was established and measured following the standard protocol of the Smithsonian Center for Tropical Forest Science [32]. We measured, tagged and identified to morphospecies, free-standing woody plant with diameter at breast height (dbh) ≥ 1 cm between January 2015 and January 2016. Stem diameter was recorded at 1.3 m above the ground, except when buttressed or swollen. Measurements were taken 50 cm above large buttresses and 2 cm below the lowest point of swellings. Trees were mapped with the highest possible precision in the field, typically at 0.1 m resolution within 5 × 5 m quadrats. Species identification was carried out mostly in the field and supplemented with DNA barcoding. Voucher specimens were collected for each morphospecies

Census
The plot was established and measured following the standard protocol of the Smithsonian Center for Tropical Forest Science [32]. We measured, tagged and identified to morphospecies, every free-standing woody plant with diameter at breast height (dbh) ≥1 cm between January 2015 and January 2016. Stem diameter was recorded at 1.3 m above the ground, except when buttressed or swollen. Measurements were taken 50 cm above large buttresses and 2 cm below the lowest point of swellings. Trees were mapped with the highest possible precision in the field, typically at 0.1 m resolution within 5 × 5 m quadrats. Species identification was carried out mostly in the field and supplemented with DNA barcoding. Voucher specimens were collected for each morphospecies

Environmental Variables and Gradients Evaluated
We calculated the location of each individual tree along three "gradients"; grassland edge to core forest, distance to water, and elevation. The elevation was based on detailed topographic maps obtained during the installation of the plot. We also mapped the two streams that run through the plot. Whether a tree was located in "forest vs. grass" was based on a geo-located Google Earth [34] image. The image comprised channels in the RGB space. An area was considered to be "grass" when the red content in a 1 × 1 m pixel in a jpg image was 37% or more. This 37% cutoff in the red channel resulted in a binary habitat map for the area with a very precise concordance with the grass vs. forest distinction on the ground (Figure 2). To avoid spurious results when calculating the minimum distance to grass from each individual tree, we removed isolated grass-looking pixels within the forested area. We forced all pixels with known presence of trees to be "forest" pixels, so all the trees were, by definition, at a distance >0 from the grass.
After defining the environment, we calculated four variables for each individual tree: (1) The proportion of grass in a 20 × 20 m area surrounding the individual tree. Rather than count the grass vs. forest pixels in a 20 m diameter circle around each tree, we used a 20 × 20 m square window around each tree because it was computationally more efficient than defining a circle around each tree. Within this square we counted the number of red (grass) vs. other 1 × 1 m pixels and from this calculated the proportion of "grass" vs. forest pixels. We used this metric to indicate whether a species was a forest or grassland species. Our 20 × 20 m squares overlapped and were not necessarily independent of each other. (2) The distance to the closest grass pixel. This was calculated from the coordinates of each individual and the grass vs. forest binary map. This metric indicates if a species prefers the forest core (=far from grass) or the forest edge or the grassland itself (=close to the grass). (3) The shortest distance to any of two streams in the plot. (4) The local elevation.

Species Guilds
Our five-guild classification is modified from Hawthorne [35,36]: (1) pioneer; (2) non-pioneer light demanding-needs shade to germinate and establish but larger individuals need light; (3) cryptic pioneer-needs light to germinate and establish but can survive for many years as adult in the understory; (4) shade-tolerant; and (5) grassland. For species not included in Hawthorne's 1995 Ghanaian study [35], we identified species guild from information on the species in the JSTOR online plant database [37].
Our seed weight data came from data collected on fruit and seed traits as part of the Nigerian Montane Forest Project [29] ongoing research. Twenty seeds from fruits of at least five individual trees across the forest were weighed to the nearest g and the average seed weight taken. Dispersal modes were based on fruit and/or seed morphology, as described in [38]. Fleshy fruits were considered animal dispersed and hard seeds in pods, ballistically dispersed. Winged seeds are wind dispersed and very heavy, large seeds embedded in a thick husk, gravity dispersed. Growth form, i.e., tree, shrub or strangler, and position in canopy (emergent, canopy, or understory) was based on [27]. In addition, trees growing in the grassland with obvious adaptations to fire were classified as savannah trees. For the purposes of this study, "grass" was defined as pixels with red content 37% or more in a jpg image of the area taken from Google Earth, presented on the right (B). This 37% threshold was optimized to obtain the best possible correspondence between the aerial image and the binary map of the grass/forest boundary. Pixels are approximately 1 × 1 m, allowing for an individual tree-level resolution. This allowed the fine-tuning of the geolocation of the plot to obtain an almost perfect match between our individual tree map (trees within the plot) and the larger grass vs. forest map.

Data Analyses
As part of the first assessment on species composition and diversity conducted in the Ngel Nyaki Forest Dynamics Plot, we calculated abundances, basal area and Fisher's alpha diversity on a per hectare basis (i.e., within squares of 100 × 100 m, each a subset of the entire plot) for three diameter classes: ≥1 cm, ≥10 cm, and ≥30 cm. We used 13 table combinations of our plot for our calculations so that we had a total of 182 1-ha plots from which we calculated our per hectare diversity indices. We analyzed diversity at three taxonomic levels: species, genus, and family.
The four observed environmental preferences (% of surrounding grass, distance to grass, distance to stream, elevation) were compared with null expectations. The null expectations for each species were obtained after 999 constrained torus translations, an adaptation of the standard or unconstrained torus translation (see Appendix A for more details on standard (unconstrained) and constrained torus translations). During a standard torus translation, all individuals are translated together as a rigid cloud of points [39][40][41]. The spatial pattern of the species (the spatial aggregation) is respected, but the relationship with the habitat map (the grass vs. forest map, or the streams, or the elevation) is broken. All of our translations were constrained to only happen within the forest, never in the grass ( Figure 3). We used the same constrained translations to assess significant associations with the grass, the streams and certain elevations. We developed constrained torus translations instead of unconstrained or standard torus translations across the entire plot because the latter predicted many individuals within the grass itself, just because a significant proportion of the Ngel Nyaki plot is covered in grass. This unrealistic expectation is a meaningless null hypothesis to test; if we did test it, it would result in all species being significantly far from the grass, even those with strong preferences for the forest edge.
After generating the null expectations, we evaluated the significance of the associations. If the observed x for the species was greater or lower than 95% of the expected (null) x, then the species Diversity 2020, 12, 30 6 of 19 was considered significantly associated to x. For example, if the observed median distance to grass of the individuals of a species was greater than 95% of the expected median distance after 999 torus translations, the species was deemed "significantly far from grass", i.e., a forest core specialist. In contrast, if it was lower than 95% of the expected median distances, the species was deemed "significantly close to grass". We followed the same approach with each of the other environmental variables considered. We used a Chi-Squared test to examine the relationship between species position on the gradient and seed size and dispersal mode. We grouped seeds into four size classes according to their weight: small (<0.6 g), medium (0.6-1.5 g), large (>1.5-3.0 g), and very large (>3.0 g). All the analyses were performed in R, version 3.3.3 [42] using customized functions or the CTFS R package [43].
1 Figure 3. Comparison between the standard (unconstrained) torus translation (A) and the constrained version that we used in our study (B). The square polygon within the plot corresponds to the original position of a subset of stems, used as an example. Three random translations of this subset of stems are presented for comparison purposes. When the stems land within forest after the translation, both methods behave similarly (whitish points). When the stems land within grass, the constrained translation makes them "jump" again until they find forest, distorting to some degree the translated cloud of points (black points). The standard (unconstrained) translation does nothing in particular, since it is blind to the forest vs. grass distinction. This non-adaptive behavior of the standard torus translation contributes to the unrealistic null expectation of "many stems within the grass", that we wanted to avoid.

Diversity and Abundance
A total of 41,031 trees were recorded with an average density of 2062 individuals/ha. The number of individuals ≥10 cm dbh and ≥30 cm dbh were 6931 (352/ha) and 1686 (85/ha), respectively (Table 1). We recorded 105 species belonging to 47 families and 87 genera in the 20.28-ha plot. Rubiaceae, with nine species, was the most diverse family, followed by Euphobiaceae and Fabaceae with six species each, and Meliaceae with five species. Fisher's alpha at the 1-ha scale was 12.42 for all stems, 12.03 for stems ≥10 cm, and 9.55 for the largest diameter class (≥30 cm dbh) ( Table 1). The most abundant species was Garcinia smeathmannii, an understory tree of the Clusiaceae family with 11,960 individuals (607.63/ha), representing approximately 30% of individuals ≥1 cm dbh (Table 2). This was followed by Deinbollia pinnata (Sapindaceae) with 3077 individuals (158.68/ha) and Pleiocarpa pycnantha (Apocynaceae) with 2142 individuals (107.51/ha).

Species Distribution and Preferences
The four gradients (distance from edge, distance to stream, elevation, % grass) evaluated in the plot were weakly correlated with each other (Figure 4). Individual trees at higher elevations were closer to the grass (Pearson's r = −0.46). The rest of the correlations were weaker than |r| = 0.30 ( Figure 4).

Species Distribution and Preferences
The four gradients (distance from edge, distance to stream, elevation, % grass) evaluated in the plot were weakly correlated with each other (Figure 4). Individual trees at higher elevations were closer to the grass (Pearson's r = −0.46). The rest of the correlations were weaker than |r| = 0.30 ( Figure 4). Individual trees occur across all gradients, therefore our analysis reflects the underlying correlation between individuals and measured local environmental variables. For computational reasons, the distance to the grass was measured with high resolution at short distances and coarser resolution at larger distances (the maximum peripheral error was constrained to < 15%). The The average individual values at the species level, however, showed stronger correlations, with some obvious elevational trends ( Figure 5). Many of the average preferences for the species did not differ from the expected by the random translations, suggesting that many of the species in Ngel Nyaki do not respond to the measured environmental gradients (See Table S2). The more common species however, showed stronger preferences ( Table 3). All of the 65 abundant species (those with at least 20 individuals in the plot) were surrounded by significantly more or less grass than the expected by the null model. Four of these were clearly savannah species (Dombeya ledermannii, Entada abyssinica, Psorospermum aurantiacum, Rytigynia sp.) and two species (Clausena anisata and Guarea cedrata) mostly occurred in areas with a high percentage of grass (> 50%) even though they are not savannah species. More than half of the common species (35) were significantly close (23) or far (12) from the grass. Around 20% of the common species (14) were significantly associated with high (8) or low (6) elevations. However, in contrast, few common species (6) were significantly close (5) or far (1, Chionanthus) from the stream.
Our analysis investigating the association between species position on the gradient and seed size or dispersal mode showed no relationship between position on the gradient and seed size (χ = 4.36, df = 6, p = 0.63) or dispersal mode (χ = 7.57, df = 6, p = 0.27). For example, species significantly closer to the grassland matrix ("edge" species in Table 3) had a median seed size of 0.25 g, twice the median size of those species significantly far from the grassland matrix (0.13 g; "core" species in Table 3). Twenty five percent and 75% quantiles for species close and far from the grass are 0.13-0.34 and 0.04-0.27 g respectively. Many of the average preferences for the species did not differ from the expected by the random translations, suggesting that many of the species in Ngel Nyaki do not respond to the measured environmental gradients (See Table S2). The more common species however, showed stronger preferences ( Table 3). All of the 65 abundant species (those with at least 20 individuals in the plot) were surrounded by significantly more or less grass than the expected by the null model. Four of these were clearly savannah species (Dombeya ledermannii, Entada abyssinica, Psorospermum aurantiacum, Rytigynia sp.) and two species (Clausena anisata and Guarea cedrata) mostly occurred in areas with a high percentage of grass (>50%) even though they are not savannah species. More than half of the common species (35) were significantly close (23) or far (12) from the grass. Around 20% of the common species (14) were significantly associated with high (8) or low (6) elevations. However, in contrast, few common species (6) were significantly close (5) or far (1, Chionanthus) from the stream.
Our analysis investigating the association between species position on the gradient and seed size or dispersal mode showed no relationship between position on the gradient and seed size (χ = 4.36, df = 6, p = 0.63) or dispersal mode (χ = 7.57, df = 6, p = 0.27). For example, species significantly closer to the grassland matrix ("edge" species in Table 3) had a median seed size of 0.25 g, twice the median size of those species significantly far from the grassland matrix (0.13 g; "core" species in Table 3). Twenty five percent and 75% quantiles for species close and far from the grass are 0.13-0.34 and 0.04-0.27 g respectively. Table 3. Associations between the distribution of the common species and four environmental gradients in the Ngel Nyaki plot (Nigeria). The numeric values correspond to the proportion of null values (medians across individuals, except in the case of % of surrounding grass, which are means) that are lower than the observed value. The columns "association" contain a short label for the ecological interpretation of the results: forest vs. savannah species, edge vs. core species (core = species that are significantly far from the grass), species that prefer low vs. high elevations, and species that prefer being close to the streams (wet) or far from them (dry). NA's mean that a given association does not deviate significantly from the expected by the null model. We include species guild, growth form, seed size, and dispersal mode (Growth-form categories of species were ET = emergent tree, CT = canopy tree, UT = understory tree, USh = understory shrub, ST = savannah tree, Sh = shrub (modified from [27]). Species shade tolerance guilds included P = pioneer, S = shade-tolerant, CP = Cryptic pioneer, NPLD= Non-pioneer light demanding, G = Grassland (modified from [35]). Dashes (-) signify insufficient data).

Discussion
Our study is the first detailed inventory of a West African montane forest that explores species distributions in relation to local environmental gradients. Our data are from the first census of the Ngel Nyaki Smithsonian Forest Global Earth Observatory plot.

Diversity and Composition of the Woody Plant Species
Within our total plant count of 41,031 trees with dbh ≥1 cm we recorded 105 species from 87 genera and 47 families. This is relatively low diversity compared with the results from previous studies in other tropical Afromontane forests. For example, Hamilton and Bensted-Smith [44], who recorded only trees with dbh >10 cm counted 104 species from the sub-montane forest of the East Usambara Mountains in East Africa. Dowsett-Lemaire, et al. [45], similarly including only larger sized trees recorded 232 species over 140 km 2 in northern Malawi, with less diversity south of 14 • S; e.g., 94 species in a forest of 11.3 km 2 in Central Malawi. This drop in species diversity in Southern Malawi was due to range limits and several species 'skipping' southern Malawi and reappearing farther south [45]. However, in concordance with some Afromontane forests, e.g., forests in the Bonga region in East Africa [46], a high proportion (90%) of the genera at Ngel Nyaki were each represented by a single species and the most common families were Rubiaceae and Fabaceae [47,48].
The number of individuals of species in the plot varied considerably, from 11 singletons (e.g., Terminalia ivorensis and Olax sp.) to the extremely abundant Garcinia smeathmannii (Clusiaceae) with about 12,000 individuals (607 trees/ha). Garcinia smeathmannii, which makes up about one-third of the total tree abundance in the plot, is also the third most abundant species in the Rabi plot (Gabon), with 304 trees/ha [48]. Anthonotha noldeae (Fabaceae), an Afromontane endemic and one of the most common canopy species in Ngel Nyaki forest, showed the highest dominance if measured by basal area. This is of interest because locally, A. noldeae is only found in Ngel Nyaki forest and Cabbal Shirgu, and is absent from the rest of the Gotel mountains in Nigeria [51] and Mt Oku in neighbouring Cameroon [52]. We recorded 607 trees/ha of our most abundant species G. smeathmannii which is comparable to the abundance of 684 trees/ha of Dichostemma glaucescens in Rabi (Gabon) [48] and 535 trees/ha of Phyllobotryon spathulatum in Korup (Cameroon) [47]. About 39% of the species in the plot had <1 tree/ha, which when compared to other ForestGEO plots is similar to the 45% in Korup (Cameroon) and BCI (Panama), 34% in Pasoh (Malysia) and 40% in Yasuni (Equador), but lower than in the monodominant Ituri forest (Democratic Republic of Congo) where 72% of the species there are represented by <1 individual per hectare [47,50]. Among large diameter trees (≥10 cm dbh), Strombosia schefleri (Erythropalaceae) was the most abundant species with 35.3 trees/ha which is very similar to the abundance of Trichilia tuberculata (Meliaceae), the commonest large tree in BCI (Panama) with a reported abundance of 35 trees/ha. Although Ngel Nyaki is less species diverse when compared with many tropical forests of Africa, it comprised species both widespread across Africa and Afromontane endemics or near-endemics sensu White [2].

The Role of Environmental Variables
The main aim of this study was to look for evidence of local environmental factors influencing species distributions. In terms of core to edge gradients, we tested the hypothesis that species sorting from the core forest to the forest edge should result from the progressive encroachment of forest into the grassland matrix via dispersal of small seeded, light tolerant species [26]. We found partial support for this hypothesis in that 40% of all tree species and 50% of the abundant species did show a clear preference for either the edge/grassland habitat or the forest core. However, we found no association between position on the gradient and seed size or dispersal mode, which suggests that factors other than dispersal are influencing species distributions (see below).
Many of the species strongly associated with edge were expected; e.g., Entada abyssinica, Psorospermum aurantiacum, Rytiginia sp., and Dombeya ledermannii are all species restricted to grassland or the very edge of the forest, where light is always abundant. Other edge associated species such as Bridelia speciosa, Rauvolvia vomitoria, Millettia conraui, and Eugenia gilgii are also fire resistant, small to medium-sized pioneers, but occur in the understory as well as on the forest edge. Jackson [53] noted a similar guild in the Imatong Mountains of Sudan, and Adie, Kotze, and Lawes [26] in the Drakensberg montane forest of South Africa. That these light demanding species occur in the forest understory indicates their 'cryptic pioneer' [35,54] status. That is, they need light to germinate and establish but as they increase in size, they become more shade tolerant and can survive in the understory as adults for many years. Hawthorne [35] suggested that in Ugandan forests these shade tolerant traits may be adaptive and not necessarily present throughout a species range, which would be worth investigating. Of note is that in our study site the very common canopy legume tree, Anthonotha noldeae also fits the description of a cryptic pioneer. Anthonotha's abundance on the forest edge and its likely nitrogen-fixing properties [55] suggests it is an important component of the forest flora, potentially playing a key role in forest expansion.
Species strongly associated with the core forest were all shade tolerant, forest species and included the emergent wind dispersed, Newtonia buchannii, and animal dispersed canopy species such as Drypetes gossweileri and Chrysophyllum albidum. Understory shade species included Voacanga africana and Dicranolepis grandifolia. Most of these core forest species are widespread across Africa and in Nigeria are a components of forests up to at least 1800 m in elevation [27].
Against our hypothesis for a core to edge gradient we found that almost 50% of species with ≥20 individuals in the plot (seven of which were in the 20 most abundant species in the plot), occurred throughout the forest, with no statistical preference for edge or core. Together, these species included almost the entire range of species guilds. For example, the understory tree species Carapa oreophila, Celtis gomphophylla, and Mallotus oppositifolius are 'cryptic pioneers', present deep within the forest as adults but showing no evidence of regeneration [56]. In contrast, species such as Parkia filicoidea, Entandrophragma angolense and Pouteria altissima are 'Non-Pioneer Light Demanding' [35] in that they need shade to germinate and establish, but as adults are light demanding. Shade tolerant understory species with no associations with either core or edge include Dasylepis racemosa, Diospyros monbuttensis, and Santiria trimera.
In contrast to our expectations, we found that species closer to the edge had, on average, larger seed sizes than the core species. So that unlike the situation in the South African temperate montane forests where small seeds are obviously key to forest success, we found that in West Africa, larger-seeded and fire-adapted species seem to dominate the edges of each forest patch. Other adaptations, like nitrogen-fixing nodules in Milletia [57] and likely A. noldeae [55] may help in seedling establishment and subsequent survival in the grassland matrix. In general, how these forest patches grow and function as species refugia during times of forest expansion [26] are exciting questions and opportunities for further research.
While the grassland to forest gradient was the steepest in our analysis, we also explored other potentially relevant gradients to further explain the distribution of tree species at Ngel Nyaki. For example, elevation gradients often play a key role in driving functional diversity [58]. We found 14 species were significantly associated with either low or high elevation within the plot. Among those associated with high elevation, Anthonotha noldeae, Clausena anisata, Entada abysinnica, Ritchea albersii, Rothmannia urcelliformis, and Tabanaemontana contorta were also associated with grass/forest edge and distance to grass, which makes sense considering that the grasslands are at high elevation in our plot.
Another gradient explored was distance to stream (a proxy for soil moisture) but this was only weakly associated to species distribution in Ngel Nyaki, with only 6 of our 65 abundant species showing significant associations to distance from stream ("wet" or "dry" species in Table 3). This was unexpected given that water availability is well known to play a key role in governing local patterns of plant distributions in many ecosystems [59]. For example, in the Barro Colorado Island plot, Comita and Engelbrecht [60] reported significant associations between species distributions and water availability. In contrast however, and more like us, Scalley, et al. [61] in the Luquillo wet montane forest plot in Puerto Rico found no association between the spatial distribution of tree species and stream locations, but rather, that stem density and species richness was highest in areas closest to streams.
Our findings may, to some extent, corroborate those of Adie, Kotze, and Lawes [26] who suggest that in the South African Drakensberg mountains, fire, coupled with harsh abiotic conditions have filtered Afrotemperate forest tree composition resulting in a prevalence of species traits such as small, bird-dispersed fruits and fire resistant bark across the forest. Such filtering could perhaps explain the presence of the long lived 'cryptic' pioneers Polyscias fulva and Trema orientalis throughout Ngel Nyaki forest. However, in Nigerian montane forests there is a range of species guilds and seed sizes; while some may reflect species filtering through the matrix, others are shade tolerant, late succession old forest species such as Santiria trimera [61]. Regarding S. trimera, both the stilt root and non-stilt root morphs occur at our study site. The separation between the morphs is deep rooted in the Santiria phylogenetic tree [62], therefore their presence at Ngel Nyaki perhaps suggests that Ngel Nyaki forest is ancient and was a forest refugia during periods of Pleistocene drying. Population genetic analysis following that of Koffi, Hardy, Doumenge, Cruaud and Heuertz [62] on Ngel Nyaki S. trimera and other forest tree species would help shed light on this. However, that Ngel Nyaki forest is the most species rich forest on the Mambilla Plateau [45] with several rare species [25] and some even new to science [63], and that A. noldeae occurs here in such abundance but almost nowhere else in West Africa, may again support the idea that Ngel Nyaki was a Pleistocene refuge. Moreover Brailsford [64] recorded high levels of chloroplast haplotype diversity at Ngel Nyaki as compared with other forest fragments on the Mambilla Plateau. Of course alternative explanations exist for the high diversity and again, molecular analyses would help shed light on this.

Conclusions
Our study provides the first detailed insight on patterns of species and trait guild composition in a West African montane forest. The Ngel Nyaki forest is similar to many other tropical and Afrotropical forests in terms of species diversity and dominance patterns. The gradient from forest edge to core is a key dimension of the niche as perceived by plants. Overall,~40% of the individuals belong to species with significant preferences towards edges or forest core areas, which in turn represent~45% of the local woody diversity. While dispersal abilities play a role in forest expansion, our data did not show any association between species position on the gradient and dispersal mode. Further research is needed to disentangle the effect of changes in forest patch size and environmental preferences of the species to make more accurate predictions of how Afromontane forests will behave under changing fire regimes and human disturbance in the region. grass within it. If we translated points to forest or grass indistinctly, we would end with many trees within the grass. This would generate the (erroneous) expectation of all the species being strongly associated with grass. By comparing that expectation with the observations, we would conclude that all the species are significantly far from the grass. Even edge specialists.
To avoid that we run constrained translations in the torus. Like standard torus translations, a constrained translation makes the trees jump together (as a rigid cloud of points) randomly in the torus. Unlike the standard torus translations, we added the constraint of never arriving in grass. Jumps were forced to be sufficiently long that they always arrived in forest rather than grass, while at the same time respecting as much as possible the relative position of the translated points. One way of thinking about this is as a non-perfect torus, narrower in places with more grass, and wider in places with more forest. The torus is still a torus, topologically speaking. Another way to think about this is to imagine a map with invaginations in the forbidden areas, on top of which a regular grid of coordinates is placed (with no invaginations), to serve as a reference for the translations.
In summary, a constrained torus translation is a translation of a cloud of points that respects, as much as possible, the relative position of the individuals. Species that are strongly aggregated have expectations composed of tight clouds of points, and vice versa. We inspected visually many translations and we concluded that the result was satisfactory for our purposes.