Specimen-based analysis of morphology and the environment in ecologically dominant grasses: the power of the herbarium

Abstract

Herbaria contain a cumulative sample of the world's flora, assembled by thousands of people over centuries. To capitalize on this resource, we conducted a specimen-based analysis of a major clade in the grass tribe Andropogoneae, including the dominant species of the world's grasslands in the genera Andropogon, Schizachyrium, Hyparrhenia and several others. We imaged 186 of the 250 named species of the clade, georeferenced the specimens and extracted climatic variables for each. Using semi- and fully automated image analysis techniques, we extracted spikelet morphological characters and correlated these with environmental variables. We generated chloroplast genome sequences to correct for phylogenetic covariance and here present a new phylogeny for 81 of the species. We confirm and extend earlier studies to show that Andropogon and Schizachyrium are not monophyletic. In addition, we find all morphological and ecological characters are homoplasious but variable among clades. For example, sessile spikelet length is positively correlated with awn length when all accessions are considered, but when separated by clade, the relationship is positive for three sub-clades and negative for three others. Climate variables showed no correlation with morphological variation in the spikelet pair; only very weak effects of temperature and precipitation were detected on macrohair density.

This article is part of the theme issue ‘Biological collections for understanding biodiversity in the Anthropocene'.

1. Introduction

Herbarium specimens constitute a massive but underused resource for the study of biodiversity. They represent the combined efforts of thousands of collectors working over the course of centuries throughout the world; together, they represent an enormous sampling of the world's flora. While the specimens have unquestionable value, recent advances in data acquisition and manipulation have enhanced their use even further. Imaging and image analysis provide new ways to capture, analyse and preserve morphological data from herbarium specimens, even from very tiny structures. In addition, advances in DNA sequencing technology and phylogenetic analysis permit construction of molecular phylogenies from most herbarium material. Ready availability of environmental data and digitization of label data makes linking environmental data to georeferenced herbarium specimens possible and amenable to automation.

In this study, we focus on the use of herbarium specimens to examine patterns of biodiversity and evolution in grasses in the tribe Andropogoneae, a group of species that includes the ecologically dominant grasses of the tallgrass prairie of North America, the savannahs of Africa, the grasslands of Australia and the tropical cerrado of Brazil [1]. The tribe also includes the economically dominant species maize, sorghum and sugarcane [2] The morphological-taxonomic pattern in the 1200+ species of Andropogoneae is complex [3–5]. Like most grasses, Andropogoneae bear their flowers in tiny clusters known as spikelets (literally, little spikes). In Andropogoneae, the spikelets are arranged in pairs, one sitting (sessile) on the inflorescence axis (the sessile spikelet, hereafter SS), and the other on a short stalk (pedicel, the pedicellate spikelet, hereafter PS) (figure 1) [6,7]. The SS and PS generally differ in morphology and function. Most notably, only one (usually the SS) produces a seed. In addition, the SS often has a slender outgrowth, or awn, extending from one of its bracts, although some species have awns on the PS as well [6,7].

Figure 1.

(a) Cartoon showing two spikelet pairs, slightly separated by abscission at the node. (b) Photograph of a herbarium specimen of Andropogon ligulatus (Miller 7053, MO). ss, sessile spikelet; ps, pedicellate spikelet; awn, awn; int, inflorescence internode; ped, pedicel.

Several centuries of grass taxonomy have focused on flower and spikelet morphology, and species are recognized and named based on the size, shape, hairiness (pubescence), texture and other attributes of their component parts, characters that appear fixed within species but strikingly variable among them [8–13]. Despite this long history and richly detailed descriptive literature, the origins, genetics and functional role of the component parts of the flowers and spikelets are largely unknown. One key question surrounding the evolution of the spikelet pair is whether all parts of the spikelet pair co-evolve. The spikelet pair is usually described as an integrated structure that originates from a single large meristem [14–16], but it is unclear whether the structures that make up the spikelet pair experience different selective pressures. Moreover, taxonomic definitions of species are also largely disconnected from ecological data. For instance, within Andropogoneae, there is much diversity in the structure of the SS and PS (e.g. differences in size and weight of the diaspores, number of anthers per ovule, types and amount of hairs), in the awns (e.g. amount of twisting, degree of bending, type and amount of hairs), and in the amount of pubescence on the various structures of the diaspore [7,10,13]. However, hardly any data address whether aspects of the abiotic environment might correlate with one species having hairy spikelets and its sister species having glabrous ones, or one species having long awns and the other having awns that are short or absent.

Taxonomic descriptions and morphometric analyses have traditionally focused on linear measurements, particularly lengths of structures, because they are easy to generate and to interpret. However, such measurements may not capture the variation in the shape of complex organs. We hypothesize that natural selection might be acting on the shape of an organ, rather than just its length, and thus that a metric which captures shape could exhibit a stronger correlation with aspects of the abiotic environment than length does. Many methods have been developed for capturing and analysing shape, and this area has received increasing attention as machine learning is applied to analysis of images. Examples include outline-based approaches such as elliptical Fourier analysis [17] or geometric approaches, including geomorphometrics (e.g. [18]) and persistent homology [19]. Each of these is based on different underlying assumptions and functions but all have the same goal of providing a metric that permits comparison of complex shapes (see references in [20]).

Here we examine morphological variation in a set of 632 accessions representing 186 nominal species in a major clade of Andropogoneae, including Diheteropogon, Andropogon, Schizachyrium and Hyparrhenia (DASH) plus several smaller genera. We first correlated classically used morphological measurements (length values) with each other and with aspects of the environment for all 186 species. We then extracted a subset of 81 accessions, for which we sequenced full plastomes and generated additional morphometric measurements using persistent homology (PH) (a flexible and robust mathematical approach to comprehensively describe morphology of two-dimensional (2D) and three-dimensional structures; see Li et al. [19,21,22,26]) as well as other aspects of overall shape. Importantly, these analyses are specimen-based, with completely parallel datasets, not based on extrapolations from one dataset to the other. The specimen-based approach also makes no assumptions about species limits, which is particularly appropriate in taxa for which monographs do not exist and keys may be provisional. The objectives of this study were to use this specimen-based, phylogenetic context to: (i) test whether taxonomic characteristics within the DASH clade are evolutionarily correlated, and (ii) to test for associations between climate and morphological variation within the clade.

2. Methods

(a). Taxon sampling, imaging, morphological data collection and environmental data

To sample as many members of the DASH clade as possible, we used the extensive herbarium collections at the Missouri Botanical Garden (MO) and the Royal Botanic Gardens, Kew (K). We were able to sample 186 nominal species (87 Andropogon, 47 Hyparrhenia, 39 Schizachyrium, five Elymandra, four Diheteropogon, three Monocymbium and one Exotheca species) representing approximately 75% of the clade (figure 2; electronic supplementary material, table S1). We chose three specimens per species that encompassed the broadest possible geographical range of the species, given available herbarium material. Species assignment was verified insofar as possible, but a comprehensive monograph is available only for Hyparrhenia [23], so other determinations should be considered provisional. If no specimens of a given species included complete, undamaged spikelets, that species was not included in the analysis. Our full sample included 632 accessions (the 632 specimen dataset). A small amount of leaf material was removed from each for DNA extraction (see below).

Figure 2.

Distribution of all 632 specimens included in this study, overlaid on a map showing mean annual precipitation. Red symbols, specimens from the 81 specimen dataset; black symbols, remaining specimens.

Digital photographs were taken of three spikelet pairs per herbarium specimen, using a Canon EOS Digital Rebel XTi 10.1MP Digital SLR camera with a standard macro lens and ring flash, mounted on a copy stand. Spikelet pairs were arranged on a platform of black velvet with a scale bar and specimen identification tag (see example in [24]). A photograph of the unaltered diaspores was taken, followed by one with the SS awns (if present) removed and arranged beside the spikelet pair for ease of later digital measurement. Photographs of the complete herbarium sheet and the herbarium label were also captured for each specimen. All photographs are available online via the Missouri Botanical Garden's Tropicos database (www.tropicos.org).

Morphometric data were collected digitally using TPSDig (v. 2.0; [25]). These included measurements of the lengths of the SS and PS, pedicel, internode and awns of the SS and PS (if present) (table 1). Macrohair density (pubescence) on the diaspore was estimated by assigning a score of zero (absent), 1 (sparse) or 2 (dense) (figure 3 and table 1) for each part of the diaspore, then averaged for each specimen to create a continuous variable. All measurements were averaged across the three spikelets per specimen to produce an average value for each trait for each specimen.

Table 1.

Summary of morphological characters and their character states. (All length measurements in mm. Method of measurement was either digital (dig), using TPSDig [18], by manual inspection (man) or by fully automated methods (auto) in Matlab. Prickle hairs are microscopic projections along the edges of the spikelets and awn, and were recorded according to their location. Macrohairs are pubescence that was visible without a microscope.)

trait	method of measurement	comments
SS length	dig
SS awn length	dig	removed from spikelet, measured from top of lemma to tip of awn
PS length	dig
PS awn length	dig
pedicel length	dig
internode length	dig
SS awn	man	geniculate (bent) or straight
SS awn	man	twisted or straight
prickle hairs	man	hair-like, tooth-like or absent
macrohairs	man	absent, sparse or dense
mean spikelet pair area	auto	measured on reduced dataset only
spikelet pair aspect ratio	auto	measured on reduced dataset only
spikelet pair circularity	auto	measured on reduced dataset only
spikelet pair solidity	auto	measured on reduced dataset only
persistent homology values of spikelet pair shape	auto	measured on reduced dataset only

Figure 3.

Spikelet pairs with different density of macrohairs. (a) No macrohairs: Schizachyrium claudopus (Astle 906; K); (b) sparse macrohairs: Schizachyrium gracilipes (Rosengurtt 8413; K); and (c) dense macrohairs: Andropogon kelleri (Beckett 1334; K). Scale bar in (a) applies to all three panels and is 2 mm long.

For most accessions for which we had DNA samples (an 81 specimen subset, described below), spikelet images were also processed digitally using automated methods in Matlab to extract the outline of each spikelet pair and the dissected awn. Using the extracted outlines for the spikelet pairs, overall area of the structure, aspect ratio, solidity and circularity were computed. Awn area was also estimated independently.

In addition, overall spikelet shape was characterized using methods of PH described by Li et al. [19,21,22,26]. Here, we used a 2D application that is robust to noise, and orientation-invariant [19,26]. The outline of each spikelet pair was extracted as a 2D point cloud, and the point cloud centred and normalized to its centroid size (electronic supplementary material, figure S1A). A Gaussian density estimator assigned each pixel a value based on the density of its neighbouring pixels (electronic supplementary material, figure S1B). This function can reveal some of the shape structure as it normally has high value in the sinuses of an outline or intersection. Studying the function that falls in a sequence of expanding rings allows capture of spatial geometric information (electronic supplementary material, figure S1C). Within each ring, the function can be treated as ridges with the value as height (electronic supplementary material, figure S1D). Next, it is possible to assume there is a plane that intersects the ridges at some height level. For each plane/height level, we then counted the number of ‘blobs' and ‘holes' above the plane and recorded a number known as the Euler characteristic (number of ‘blobs' minus number of ‘holes') for each plane level. As the plane was moved from the top to bottom of the density function, we accumulated a sequence of Euler characteristic numbers which forms a curve to describe the shape (electronic supplementary material, figure S1E).

The end product of the PH analysis is a dataset of very high dimensionality that is hard to visualize. To reduce the dimensionality and to aid in visualization, we performed a principal components analysis (PCA) on the Euler characteristic curves. From the PCA plot, in general, the shape with the most negative value for principal component 1 (PC1) has many long hairs and roundish overall shape, while the shape with most positive PC1 has single long awn and long overall shape (electronic supplementary material, figure S1F).

Herbarium specimens were georeferenced as necessary, using herbarium label locality information and Google Earth (www.google.com/earth/). Nineteen bioclimatic variables derived from the WorldClim global climate database [27] associated with each specimen locality were extracted at 30 arc-second resolution using the raster package in R (v. 3.4.3; [28]).

(b) Plastome sequences and phylogeny

We constructed a plastome phylogeny for an 81 specimen dataset, focusing on Andropogon, Schizachyrium and Hyparrhenia, as well as several smaller genera that appeared likely to be related. DNA extraction followed a modified cetyl trimethyl ammonium bromide method [29]. Illumina library prep followed efficient methods developed in the Kellogg laboratory for dried material [30]. Chloroplast genomes were assembled from raw reads using Fast-Plast (https://github.com/mrmckain/Fast-Plast) and were annotated using Verdant [31]. Annotations were verified using scripts leveraging GenBank validation (https://github.com/mrmckain/Verdant_Utilities/), and errors corrected manually. In all cases of potential errors, reads were searched to identify misassemblies that may have resulted in incorrect annotations as described in McKain et al. [32]. Final annotations are available in Verdant (verdant.iplantcollaborative.org) and GenBank (MH181163-MH181237, MH660709, MH660710 and MH185800; see also the electronic supplementary material, table S1).

Assembled chloroplast genomes were split into their subcomponent regions: large single copy (LSC), small single copy (SSC) and inverted repeat (IR) using a custom perl script (https://github.com/mrmckain/Genome_Skimming_Utilities). Each region was aligned individually using MAFFT v. 7.245 [33] with high-speed default parameters, and aligned regions were concatenated using a custom perl script (https://github.com/mrmckain/Genome_Skimming_Utilities). The best base pair substitution model was identified as TVM + gamma + I based on the Aikake information criterion (AIC) in jModelTest2 v. 2.1.5 [34]. RAxML does not have a TVM equivalent, so GTR was used as a replacement. A maximum-likelihood (ML) phylogeny was reconstructed using RAxML v. 8.0.22 [35] using the model GTR + gamma + I with 500 bootstrap replicates.

To be sure the number of samples per species was not influencing our results, we did a second phylogenetic analysis reducing the tree to only one specimen per species for a total of 66 terminal taxa. Results were identical to those for the 81 specimen dataset.

(c). Statistical analysis

We analysed both the full morphological-environmental dataset (632 specimens) and the set of 81 species for which we had phylogenetic data.

(i). 632 specimen dataset

Environmental data were first analysed by PCA using XLSTAT (v. 2018.1, Paris, France) to reduce the 19 original variables to a smaller set. The first two PCs (PC1 and PC2) were used as approximations of temperature and precipitation, respectively (see Results). Relationships between average values of individuals' morphological traits (table 1) and between those traits and the environmental PCs were tested using linear regression implemented in R (R Core Team [36]). Because we do not yet have a phylogeny that includes all species, we were unable to correct for phylogenetic covariance in these analyses.

(ii). 81 specimen dataset

To verify that the results from the 81 specimen dataset were representative of those of the 632 specimen dataset, we repeated the linear regressions on non-phylogenetically corrected morphological and environmental variables. Although R² values for these correlations were consistently lower than those obtained using the large dataset, as expected given the much smaller sample size, the rank order of R² values was identical.

Having confirmed that the 81 specimens were representative of the whole dataset, we corrected for the effects of relatedness by doing the same correlation analyses as for the full dataset but used a phylogenetic generalized least-squares regression (PGLS), implemented in the R package caper (v. 0.5.2; [37]). We also used the R package phytools (v. 06–044; [38]) to map character traits and climate variables onto the phylogeny. Also using PGLS, we tested whether the PCs of the Bioclim variables were able to predict any of the standard length measurements, mean spikelet pair area, aspect ratio, circularity, solidity, awn area, or the first three PCs of the PH values. The latter eight analyses were done on a slightly smaller dataset and tree because we lacked the appropriate images for several specimens.

3. Results

(a). Morphological trends in the DASH clade

In analyses of the 632 specimen dataset, length measurements of morphological features varied widely within and between clades. The length of the SS varied from 2.2–13 mm, with a median value around 4.5 mm (figure 4). Length of the PS was more variable than length of the SS. In 20 species, the PS was absent or reduced to nearly non-existent (less than 1 mm), whereas in Diheteropogon grandiflorus, specimens had an average PS length of over 18 mm long, or approximately twice the average length of the SS. Analyses were done first using average values per specimen (specimen-based, our preferred approach) and second using average values per species (species-based, sometimes used in the literature). R² values were virtually indistinguishable (not shown), so all results presented here are specimen-based. Residuals were examined and no significant relationships between fitted values and residuals were observed.

Figure 4.

Spindle plot of morphological variables for all specimens. Each green dot is the value of the relevant character for one specimen. Magenta horizontal lines indicate median values.

For this large set of specimens, the strongest correlation between two morphological features was the relationship between PS and SS length (R² = 0.52). Correlations between other pairs of quantitative variables were generally lower (table 2).

Table 2.

Results of linear regression analyses between lengths of given structures, using the full dataset of 632 individual specimens.

length comparison	R2	estimate	F1,630	p-value
SS – SS awn	0.26	5.4	226.1	<0.001
SS – PS	0.52	1.1	675.1	<0.001
SS – pedicel	0.41	0.4	441.5	<0.001
SS – internode	0.14	0.3	107	<0.001
SS awn – PS	0.44	0.1	506	<0.001
SS awn – pedicel	0.04	0.01	29.5	<0.001
SS awn – internode	0.01	0.01	6.4	0.01
PS – pedicel	0.09	0.1	60.5	<0.001
PS – internode	0.03	0.1	17.6	<0.001
Pedicel – internode	0.37	0.7	362.9	<0.001

The length of the SS awn varied far more than any other morphological trait in this study, with awns up to nearly 110 mm in Elymandra subulata, and completely absent in eight species (A. arenarium, A. bicornis, A. bourgaei, A. festuciformis, A. leucostachyus, A. sanlorenzanus, A. selloanus and A. virgatus). In five other species (A. durifolius, A. hallii, A. insolitus and A. salzmannii), awns were present on some specimens and absent on others. It is unclear whether this represents polymorphism within the species or inaccurate species delimitation. In most specimens (84%), awns on the SS were twisted and geniculate, traits that could be associated with hygroscopically active awns [39]. Approximately 41% of species also had awns on the PS. The length of the SS is not strongly correlated with the length of the SS awn; SS length predicted approximately 26% of the variation in SS awn length (table 2).

Pedicel length varied from 0.72 mm to 9.5 mm, with a median of 3.2 mm, but pedicel length predicted only about 9% of the variation in PS length (table 2). Internode length showed similar patterns of variation; internodes were nearly absent in some Hyparrhenia species, but ranged from close to zero to 8.9 mm overall, with a median of 2.8 mm.

In a PCA of the five length characters, PC1 (50.9% of the variance) largely reflected variation in length of the SS. SS awn length and PS length loaded negatively on PC2 (28.5%), whereas pedicel and internode length loaded positively, consistent with the pairwise plots. In general, plants lacking awns also had smaller spikelets, as did those in which the awns were straight rather than twisted.

Macrohair density varied considerably within the 632 specimen dataset. Hairs were most commonly found on the pedicel, internode and callus (the base of the SS). More rarely, hairs were found along the edges of the outer bracts (glumes) of the SS and PS. Occasionally, hairs were also noted along the twisted grooves of the awn. Hairs on awns were most common in Hyparrhenia.

(b). Morphology and climate in the DASH clade

Two PCs summarizing variation in nineteen bioclim variables explained 52.16% of the total variation in climate associated with the 632 individuals sampled. Factor loadings indicated that PC1 (30.66%) represents a gradient of increasing temperature, while PC2 (21.5%) represents a gradient of increasing precipitation and decreasing precipitation seasonality. In general, the climate PCs were poor predictors of variation in morphological features of the spikelet pair (table 3). The strongest correlation was between PC2 (precipitation) and SS awn length, but PC2 explained only 8% of the variation in the length of the awn.

Table 3.

Results of linear regression analyses between length of indicated structures or average macrohair density and environmental data using the full 632 specimen dataset. (Bioclim PC1 and bioclim PC2 represent the first and second principal components of the PCA of 19 bioclim variables associated with the locality of each specimen. PC1 represents an axis of increasing temperature and PC2 represents and axis of increasing precipitation and decreasing seasonality of precipitation.)

	bioclim PC1				bioclim PC2
response variable	R2	estimate	F1,630	p-value	R2	estimate	F1,630	p-value
SS	0.01	−0.06	4.7	0.03	0.03	−0.14	17.1	<0.001
SS awn	0.01	0.71	5.2	0.02	0.08	−2.6	52.4	<0.001
PS	0.003	−0.06	1.7	0.19	0.03	−0.2	19.5	<0.001
pedicel	0.001	−0.02	0.8	0.36	0.004	−0.03	2.3	0.13
internode	0.0001	−0.05	0.08	0.78	0.004	−0.05	2.9	0.09
average macrohair	0.02	−0.03	14.2	<0.001	0.01	−0.03	10.5	0.001

(c). Phylogeny

All sequences generated for this study were first included in a large plastome phylogeny for the entire tribe Andropogoneae to verify their placement in the DASH clade [5]. We confirmed previous data placing Andropogon burmanicus well outside the clade [40]. Andropogon brachystachyus and one specimen of A. laxatus (Kayombo 2, MO) were placed sister to Sorghastrum (tree not shown); specimens appeared to be correctly identified. Andropogon crossotos and A. munroi were placed in a clade with Cymbopogon and sister to Heteropogon triticeus (not shown). Schizachyrium delavayi was placed near Bothriochloa; S. delavayi has also been placed in Eremopogon (as E. delavayi (Hack.) A. Camus), and our data hint that that placement could be accurate. Accordingly, all these taxa were included as outgroups for the present analysis, with the tree rooted at A. burmanicus.

The plastome phylogeny of the 81 specimen dataset is generally well-supported (figure 5) but does not correlate well with current taxonomy. Elymandra appears to be monophyletic but its sister relationship with Monocymbium is unexpected. Hyparrhenia could be monophyletic if Exotheca abyssinica and A. chrysostachyus are transferred to the genus. Schizachyrium s.s. is a clade that we have shown elsewhere [40] to include the type species S. brevifolium. Other Schizachyrium species fall into a clade that could correspond to a greatly expanded Diectomis. Currently, the generic name Diectomis has only been applied to D. fastigiatus, a synonym of A. fastigiatus. For ease of discussion, we are using the generic name in quotes for the entire clade, but we are not suggesting any nomenclatural changes at this stage.

Figure 5.

Maximum-likelihood (ML) phylogeny of complete plastomes for the 81 specimen dataset. Numbers above branches are ML bootstrap values. Highlighted clades are discussed in the text.

Andropogon is polyphyletic as shown previously ([40]; see Methods). The type species is A. distachyos, which falls in a clade (Andropogon s.s.) with A. mannii, A. pusillus and A. abyssinicus. Other species of Andropogon form a clade made up largely of members of sect. Leptopogon, many of which are easily recognized by having slender awns, and long silky hairs on their internodes and slender pedicels. (One of our specimens of A. laxatus (Kayombo 2, MO) appears in the outgroup, although the specimen appears to be correctly identified; this may represent a laboratory error or a long-distance transfer of a chloroplast.) Other species of Andropogon form a clade including species assigned to sects. Notosolen and Piestium (see Nagahama & Norrmann [41] for a summary of the current classification of the genus). Six species of Andropogon—A. brazzae, A. schirensis, A. heteranthus, A. gerardi, A. greenwayi, A. leprodes—are not obviously linked to any of the other clades.

(d). Phylogenetically informed analyses of morphology

Morphological analyses of the 81 specimen dataset showed that individual linear measurements were highly homoplasious on the phylogeny (not shown), with considerable variation between even closely related species. After accounting for phylogenetic covariation, significant positive correlations were found between most pairs of variables except for sessile spikelet awn and internode length (figure 6). Lambda values for the comparisons of linear measurements ranged from less than 0.001–0.93 (data not shown), indicating appreciable phylogenetic signal in some traits.

Figure 6.

Pairwise correlations of quantitative variables, with phylogenetically corrected regression lines and R² values.

However, these overall positive correlations obscure patterns in subsets of the data. We conducted exploratory PGLS analyses of morphological relationships within the clades identified in figure 5. The number of species within most clades was too low for robust statistics, but we found that many relationships varied between clades (figure 7). For example, length of the SS correlates positively with the length of its awn when all 81 specimens are analysed together (R² = 0.2237, p < 0.001). However, when trend lines are drawn for individual clades (figure 7), only three have a positive slope; in three clades (Andropogon s.s., Andropogon sects. Notosolen/Piestium and the Diectomis clade) the correlation is negative. Likewise, the length of the PS is negatively correlated with the length of the pedicel in Andropogon sect. Leptopogon, positively correlated in Hyparrhenia, and uncorrelated in Schizachyrium s.s.

Figure 7.

Pairwise correlations of quantitative variables showing linear regression lines for each of the major clades. Underlying pairwise plots the same as in figure 6, but with the points removed. Clades as in figure 5.

A handful of species have unusually large spikelets and awns, and appear as outliers in the bivariate plots. Most notable are the two species of Elymandra, Exotheca abyssinica and Hyparrhenia anemopaegma, which have independently developed awns that are several centimetres long.

Measurement of mean area of the spikelet pair using the automated, digital approach produced a different, more holistic view of the structure. Multiple regression of all length measurements as predictors of mean area found that the length measurements together explained 40% of the variation in area, with SS awn length, pedicel length and macrohair density being significant. These are the features that most directly impact the overall outline of the spikelet pair, and thus could serve as proxies for spikelet pair area. When the length measurements were used independently to predict mean area, SS length, SS awn length, PS length, pedicel length and average macrohair density were significant but not internode length (table 4), suggesting that estimates of mean area might capture much of the variation of the length measurements. The other fully automated measurements of spikelet morphology–aspect ratio, circularity and solidity—were not strongly correlated with any of the length measurements (table 4).

Table 4.

PGLS analyses of correlation between length measurements of individual structures and shape measurements (mean spikelet area, aspect ratio, solidity and circularity) on reduced dataset of 75 specimens.

response variable		SS	SS awn	PS	pedicel	internode	macrohair density
mean spikelet area	R2	0.16	0.21	0.06	0.19	0.04	0.1
	estimate	0.00003	0.0003	0.00002	0.00002	0.00001	0.000004
	F1,73	14.5	20.1	4.9	16.6	2.8	8.1
	p	p < 0.001	p < 0.001	0.03	p < 0.001	0.10	0.01
aspect ratio	R2	0.00007	0.04	0.03	0.02	0.001	0.06
	estimate	−0.0008	0.002	−0.01	0.02	0.005	0.1
	F1,73	0.005	2.8	2.6	1.3	0.1	4.8
	p	0.94	0.10	0.11	0.26	0.75	0.03
solidity	R2	0.007	0.3	0.02	0.0004	0.09	0.009
	estimate	0.007	−0.005	0.001	0.003	0.04	–0.1
	F1,73	0.5	39.9	0.02	0.03	8.0	7.9
	p	0.5	<0.001	0.89	0.87	0.005	0.006
circularity	R2	0.01	0.1	0.006	0.002	0.01	0.26
	estimate	0.002	0.0007	0.002	0.004	0.005	–0.05
	F1,73	0.8	10.0	1.5	1.1	1.8	28.1
	p	0.4	0.002	0.23	0.29	0.18	<0.001

(e). Phylogenetically informed analyses of environmental correlates of morphological diversity

Climate PCs generated from the complete 632 specimen dataset define the environmental parameter space for the entire dataset, and the relevant values were extracted for the 81 accessions for which we have phylogenetic data. Overall, phylogenetic correction of the regression of morphological variation and climate did not reveal many significant relationships (table 5). Only macrohair density was significantly correlated with climate PC1 (temperature) and PC2 (precipitation). However, as with the environmental correlations with the complete dataset, these correlations were very weak, explaining only 5% of the variation in each relationship. Likewise, mean spikelet pair area and PH PCs were poorly predicted by the environmental PCs, with only PC3 showing a significant correlation with temperature, although temperature explained only 5% of the variation in the relationship (table 5).

Table 5.

Results of PGLS analyses using bioclim PC1 and PC2 as the predictors. (*F-values (1,73), based on a slightly smaller phylogeny.)

	bioclim PC1				bioclim PC2
response variable	R2	estimate	F1,79	p-value	R2	estimate	F1,79	p-value
SS	0.02	−0.08	1.6	0.21	0.04	−0.2	3.1	0.08
SS awn	0.05	1.3	4.1	0.05	0.007	−0.8	0.6	0.46
PS	0.01	−0.09	0.9	0.34	0.04	−0.3	2.4	0.07
pedicel	0.0002	−0.005	0.02	0.89	0.06	−0.1	4.8	0.07
internode	0.004	0.03	0.3	0.58	0.04	−0.1	3.5	0.06
average macrohair	0.06	0.03	5.1	0.03	0.06	−0.05	5.1	0.03
mean spikelet pair area*	0.02	1586.9	2.9	0.09	0.004	−1143.1	0.5	0.49
persistent homology PC1*	0.02	0.6	0.6	0.42	0.07	−1.4	1.2	0.48
persistent homology PC2*	0.01	0.6	1.9	0.17	0.05	1.5	4.1	0.05
persistent homology PC3*	0.06	−0.1	0.1	0.72	0.02	0.1	2.0	0.16
circularity*	0.001	0.0004	0.1	0.75	0.04	0.004	2.9	0.09
solidity*	0.01	−0.006	1.0	0.32	0.01	0.01	1.2	0.32
aspect ratio*	0.02	0.007	1.3	0.25	0.008	0.009	0.6	0.44

4. Discussion

Herbarium specimens provide a wealth of data as shown by this paper and those in this special issue. Although taxonomists have always aimed to capture and use this information, many tools can now be deployed to expand the value of the herbarium. Additionally, data such as those generated in the present study can be returned to the herbarium to enrich the collections and digital holdings. For this project, approximately 1800 images and associated measurement data were captured and have been returned to the herbaria where they can be digitally linked to the specimens for easy public access.

In this study, we used two distinct approaches to capturing variation in spikelet pair morphology. First, we used a standard, semi-automated, landmark-based approach to record lengths of the various components of the pair. This approach fits well with traditional taxonomic practice and production of formal taxonomic descriptions, in which the spikelet pair is atomized into its component parts. It also captures structures that form at different times in development. In general, spikelets approach their mature form before macrohairs begin to develop; pedicels and internodes then elongate later. Second, we used fully automated approaches—calculation of mean area, aspect ratio, circularity, solidity and PH—that capture overall shape, which reflects the structure as it interacts with the environment and might be expected to correlate better with the environmental variables. Automated, digital measurements of overall spikelet area were only moderately correlated with the more labour-intensive, semi-automated length measurements, suggesting that they capture different aspects of morphology. Automated extraction of data from images rather than measurement of landmarks could substantially speed up data collection, but the parameters measured here are not fully commensurate with standard length measurements. Decomposing such images to capture more traditional measurements is not easy using current methods, but is an area of active research.

(a). Phylogeny and classification

The phylogeny presented here includes about one third of the taxa in the DASH clade. Consistent with previous work [3,40], the major genera are not monophyletic and some realignment will be necessary; taxonomic changes will follow completion of the phylogeny (in progress). The tree used here is a plastome tree in a group known for hybridization and polyploidy and thus reflects only the maternal phylogenetic history. However, previous data have shown that hybridization occurs largely between closely, rather than distantly, related species [3], and thus is unlikely to affect the major clades of the tree. Of the sampled taxa, Andropogon gerardi is known to be an allopolyploid of at least two distinct lineages (Schizachyrium s.s. and Andropogon sect. Leptopogon) [3]. The maternal history reflected in the chloroplast phylogeny is consistent with these previous findings and even hints that the Schizachyrium s.s. lineage may have contributed a chloroplast in the hybridization.

(b). Evolution of morphology

One key question surrounding the evolution of complex morphological structures is the ability of the component parts to change independently over time. This type of evolutionary change has been documented in vertebrate morphology [42–44]. The evolution of the diaspore in grasses raises similar questions. The spikelet pair (figure 1) characterizes all Andropogoneae including the DASH clade studied here, and although the spikelet pair is usually described as an integrated structure that originates from a single large meristem [14–16], our data show that the components of the pair change independently over evolutionary time, such that change in one part of the structure does not require change in another. For example, length of the internode is not strongly correlated with length of the pedicel. The two are on different axes (orders of branching) in the inflorescence and thus are developmentally distinct. Also, they elongate late in inflorescence development, well after the spikelets themselves are largely differentiated (see for example Hodge & Kellogg [45]). Because the different organs do not appear to change in lock step, there is opportunity for diversification within the clade. Indeed, since species are defined by relative sizes of the organs, size variation is diversification.

Individual clades differ in the relative sizes of the parts of the spikelet pair. In other words, changes in the size of one part may not drive changes in the size of other parts. This suggests that clades are marked not by changes in individual organs but rather changes in overall shape and proportions. For instance, the negative relationship between SS size and SS awn length in Andropogon sect. Piestium implies that in this lineage, selection has favoured an inverse relationship between spikelets and awns. Conversely, in Hyparrhenia, length of the sessile spikelets and their awns are positively correlated.

The approach of PH captures overall shape change and indeed appears to correlate well with the phylogeny. While all morphological characters are homoplasious on the tree, as we had expected, the first PC of the PH values appears to be more uniform within clades than any individual morphological measurement (electronic supplementary material, figure S2). Additionally, the lambda estimation for this PC (0.86) is comparable to that of several of the length measurements (data not shown).

Our data also show that multivariate measures of climate do not correlate with the morphology of the spikelet pair in the DASH clade. We initially predicted that the overall area of the spikelet pair or the PH description of spikelet pair shape would be better predicted by environmental parameters than the individual linear measurements were, however none of our morphological metrics correlated strongly with climate. This suggests that either the environment is not selecting for particular structures, or that the bioclim variables we employed are not capturing the biologically relevant aspects of the abiotic environment. The literature does suggest linkages between some features of the spikelets and the abiotic environment. For example, studies of awn structure and function have shown that hygroscopically active awns increase reproductive success in some grass species. Peart [39] documented the role of twisted, hygroscopic awns in pushing diaspores into soil microsites, resulting in higher rates of seed germination and seedling establishment. While the prevalence of hygroscopic movement of awns across the DASH clade is unknown, approximately 84% of specimens in our dataset showed the twisted, geniculate morphology associated with awn movement in other species. Awn length can also play a role in seed burial in the soil. Awn length is heritable in Hyparrhenia diplandra and, owing to the role of the awn in burying seeds deeper in soil, variation in awn length within the species may be maintained by interannual variation in fire frequency and intensity [46]. In H. diplandra, long-awned diaspores were found to be buried deeper in the soil, where high temperature associated with intense fires would have less impact on seedling germination and establishment.

Likewise, patterns of macrohair density on diaspores in the DASH clade might also be shaped by responses to climate. Climate could impact the evolution of hairiness on diaspores in a number of ways. Peart [39] demonstrated that hairs on the diaspores of some grass species aided in securing diaspores into microsites in the soil. Long hairs collapsed in high humidity but resumed their original texture and orientation when humidity dropped, which resulted in a ‘creeping' movement across soil surfaces. Peart speculated that these hairs increased seed lodging in soil, but also probably increased buoyancy during transport of the seed by wind and helped reduce water loss from the diaspore via evaporation. Mamut et al. [47] also found that higher levels of pubescence on seeds were associated with higher amounts of water imbibed and retained by seeds in the soil. While our data showed significant correlations between average macrohair density and the first two PCs of the environmental data, when corrected for phylogenetic covariation, the climate PCs explained only 6% of the variation in macrohair density (table 5). However, in some lineages within the DASH clade, our data suggest a connection between macrohair density and bioclim PC2, particularly in the Hyparrhenia and Schizachyrium lineages (figure 8). As we continue to add to our phylogeny and increase our sample size, we plan to continue to investigate this connection in the future. We also hope to refine our examination of the potential impact of abiotic factors on features of the diaspore in the future with finer-resolution soil type, soil moisture and fire frequency data.

Figure 8.

(a) Macrohair density mapped on the phylogeny. (b) PC2 of the bioclim variables (precipitation) mapped on the phylogeny.

5. Conclusion

Our data show the power of herbarium specimens for exploratory analyses of morphology and the environment. The images enhance the value of the collections, and once captured permit extensive data exploration. Methods of capturing entire shapes of structures, such as estimates of spikelet pair area or a PH-based shape descriptor also evaluate the collection of organs in a holistic way. Because it is the set of organs together (spikelets, pedicels, internodes and awns) that interact with the environment, rather than the individual organs themselves, the holistic measurements may be most appropriate for studies such as this that attempt to infer correlations between morphology and the abiotic environment. However, as noted in this study, reliance on fully automated methods would not have shown the clade-specific differences in morphological correlations that were uncovered by semi-automated methods. While both the automated and semi-automated approaches for morphological data collection described here are powerful tools that could help to better leverage the enormous amounts of data in herbarium collections, specific choices of data collection approaches will be driven by available time and resources, as well as the nature of the questions being explored.

Data accessibility

All images have been submitted to Missouri Botanical Garden for inclusion in the Tropicos database (www.tropicos.org), where they are linked to specimen records and will be available online. Sequence data have been deposited at GenBank (Appendix), and full data matrices are available from the Dryad Digital Repository at: http://dx.doi.org/10.5061/dryad.70f538g [48].

Authors' contributions

C.A.M. developed methods for specimen imaging, led imaging and efforts to extract morphometric data and led data analysis methods in consultation with M.R.M., M.L. and E.A.K. M.R.M. led DNA extraction, library construction, plastome assembly, plastome annotation and phylogenetic analysis. B.B. georeferenced specimens, generated digital measurements, extracted and analysed bioclim variables and provided GIS maps. E.A.K. designed, structured and oversaw the study. M.L. conducted automated image analysis and assessment of persistent homology. E.A.K. and C.A.M. wrote the manuscript, with input from M.R.M., M.L. and B.B. All authors read and approved the final manuscript.

Competing interests

We have no competing interests.

Funding

This work was supported by NSF grant nos DEB-1145884 and DEB-1457748 to E.A.K., and by the Summer Research Assistantship Fund from Principia College.