Available via license: CC BY
Content may be subject to copyright.
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
96Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Internaonal Journal of Odonatology
2022, Vol. 25, pp. 96–106
doi:10.48156/1388.2022.1917184
Research Arcle
OPEN ACCESS
This arcle is distributed
under the terms of the
Creave Commons
Aribuon License,
which permits unrestricted use,
distribuon, and reproducon in
any medium, provided the original
author and source are credited.
Published: 8 December 2022
Received: 11 June 2022
Accepted: 7 November 2022
Citaon:
Sáenz Oviedo, R. Kuhn, Rondon
Sepulveda, Abbo, Ware &
Sanchez-Herrera (2022):
Are wing contours good classiers
for automac idencaon in
Odonata? A view from the
Targeted Odonata Wing
Digizaon (TOWD) project.
Internaonal Journal of
Odonatology, 25,
96–106
doi:10.48156/1388.2022.1917184
Data Availability Statement:
All relevant data are
within the paper and its
Supporng Informaon les.
Are wing contours good classiers
for automac idencaon in Odonata?
A view from the Targeted Odonata
Wing Digizaon (TOWD) project
Mayra A. Sáenz Oviedo 1, William R. Kuhn 2, Marn A. Rondon Sepulveda 1,
John Abbo 3, Jessica L. Ware 4 & Melissa Sanchez-Herrera 4,5*
1 Department of Epidemiology and Biostascs. Poncia Universidad Javeriana, Ak. 7 # 40-62,
10231, Bogotá, Colombia
2 Discover life in America, Gatlinburg, TN, USA
3 Department of Museum Research & Collecons. University of Alabama Museums. Tuscaloosa,
AL 35487, USA
4 Division of Invertebrate Zoology. American Museum of Natural History. 200 Central Park West
at West 79th Street, New York, NY 10024, USA
5 Faculty of Natural Sciences, Biology Department, Universidad del Rosario,
Sede Quinta Mus. Ak. 26 # 63C-48, 11122, Bogotá, Colombia
* Corresponding author. Email: melsanc@gmail.com, melissa.sanchezh@urosario.edu.co
Abstract. In recent decades, a lack of available knowledge about the magnitude, identy
and distribuon of biodiversity has given way to a taxonomic impediment where species
are not being described as fast as the rate of exncon. Using Machine Learning methods
based on seven dierent algorithms (LR, CART, KNN, GNB, LDA, SVM and RFC) we have
created an automac idencaon approach for odonate genera, through images of wing
contours. The training populaon is composed of the collected specimens that have been
digized in the framework of the NSF funded Odomac and TOWD projects. Each contour
was pre-processed, and 80 coecients were extracted for each specimen. These form a
database with 4656 rows and 80 columns, which was divided into 70% for training and
30% for tesng the classiers. The classier with the best performance was a Linear Dis-
criminant Analysis (LDA), which discriminated the highest number of classes (100) with
an accuracy value of 0.7337, precision of 0.75, recall of 0.73 and a F1 score of 0.73. Ad-
dionally, two main confusion groups are reported, among genera within the suborders of
Anisoptera and Zygoptera. These confusion groups suggest a need to include other mor-
phological characters that complement the wing informaon used for the classicaon
of these groups thereby improving accuracy of classicaon. Likewise, the ndings of this
work open the door to the applicaon of machine learning methods for the idencaon
of species in Odonata and in insects more broadly which would potenally reduce the
impact of the taxonomic impediment.
Key words. Classicaon, Machine Learning, supervised, wings
Introducon
Dragonies and damselies (Odonata) are one of the most charismac insect
groups, due to their relavely big size, ight paerns and beauful coloraons.
Their associaon with aquac environments means they serve as excellent bioin-
dicators of water quality, given their high suscepbility to environmental changes
(Córdoba-Aguilar, 2008; Moore, 1997; Samways & Steytler, 1996). Global odonate
richness is esmated to comprise around 6.323 species (Paulson et al., 2021). This
is a relavely small number of species in comparison with other insect orders like
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
97Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Coleoptera, which includes approximately 300,000 spe-
cies (Lorenzo-Carballa & Cordero Rivera, 2012; Paulson
personal communicaon, May 14, 2021). However, re-
searchers suspect that ~20% of species remain to be
discovered (Kalkman et al., 2008). The tangible and rel-
avely low diversity in odonates makes them an ideal
scenario to address the “taxonomic impediment”—a
lack of available knowledge about and trained exper-
se to determine the magnitude, identy, and distribu-
on of biodiversity (González, 2009). This phenomenon
is of parcular interest given the current worldwide
“biodiversity crisis” (i.e., rapid declining of populaons
as a result of massive habitat destrucon and climate
change), in which, there are esmates that ~50% of the
living species will face exncon in the next 50 years
(Koh et al., 2004). Maximizing eorts to gather and
learn the taxonomy and biology of species is more rel-
evant now than ever (Ceballos et al., 2015; Kuhn, 2016;
La Salle et al., 2016).
There are several morphological characteriscs that
dene the odonate suborders Zygoptera and Aniso-
ptera, including: characters regarding the shape of the
wings, head, thorax, abdomen and genitalia (Garrison
et al., 2006). In parcular, wing shape and venaon
paerns are one of the most commonly used traits to
classify dragonies and damselies to family or genus
level. For example, without the aid of a microscope
one can easily dierenate them because anisopterans
have dierent shapes of the fore- and hindwings, which
remain perpendicular to the body when in rest, while
zygopterans’ fore- and hindwings are similar in shape
and are usually folded in line with the body (Heckman,
2006, 2008). Recent contribuons by Appel & Gorb
(2014) proposed detailed micro-morphological charac-
teriscs of the wings such as the types of vein joints
and combinaons among them (i.e., four types of vein
joints and ve combinaons), spine distribuon across
the wings (i.e., located on transversal veins, possibly in-
volved in movement limitaon), and the distribuon of
patches of the exible protein resilin in the wings (e.g.,
on the joints, and/or along the veins). These new mor-
phological traits have been discussed in the classica-
on for both suborders, and are used to infer funcon
and ight behavior.
Recently, Kuhn (2016) developed an automac clas-
sicaon system for 26 dragony genera, using stan-
dardized image scans of specimen wings. He trained
and classied them using a random forest algorithm by
extracng feature vectors to describe texture and pat-
terning through Gabor Wavelet Filters and a color as-
sessment with a chromacity standardizaon sampling
within the images. Here we assessed the classicaon
power to genera of a novel classier trait for wings—
their contour. By using standardized wing images from
the Targeted Odonata Wing Digizaon project, we test-
ed mulple Machine Learning classicaon algorithms
(e.g., Linear Discriminant Analysis—LDA, Logisc Regres-
sion—LR, Classicaon and Regression Trees— CART,
K-Nearest Neighbors—KNN, Naive Bayes—NB, Support
Vector Machines—SVM, and Random Forest Classier—
RFC) to establish the potenal use of the wing contour
within automated classicaon systems for odonates.
Materials and methods
We analyzed data from the Targeted Odonata Wing
Digizaon Project (TOWD; hps://digizingdragon-
ies.org), which aims to digize the wings of all North
America species of Odonata and to develop tools for
automacally extracng useful characters from odo-
nate wings to facilitate comparave studies and au-
tomac species classicaon. We analyzed a dataset
comprising 2,328 dragony and damselies specimens
from 111 genera, which were digized through the
TOWD Project. The dataset consisted of the contour
(outline) of the fore- and hindwings of each specimen.
These data were extracted from digital scans of the
specimens using an edge-nding algorithm to recover
a series of points (x,y-coordinates) represenng the lo-
caon of each pixel along the edge of a wing. In most
cases, the contours represented the right wings, which
were excised from the specimen’s body and scanned
on a atbed scanner, except in some cases where the
le-side wings were scanned when the right ones were
damaged (see Supplementary Table 1 for a list of speci-
mens). In the laer case, wing contours were reected
le-to-right to match up with right-side wing contours.
As part of the TOWD preprocessing, each contour was
rotated so that the upper side (costal margin) is approx-
imately horizontal, translated so that the upper-le cor-
ner is at (0,0) and scaled to millimeters. The edges of
some wings were damaged, which was also apparent in
their respecve contours; such damage was used as an
exclusion criterion.
Contours data were preprocessed and analyzed in
Python (van Rossum & Drake Jr, 2009; v. 3.9.2) using
the Spyder Integrated Development Environment (IDE)
Spyder (Raybaut, 2009; v. 4.2.1), which is part of the
Anaconda Soware Distribuon (2016). Data treat-
ment was divided into four main steps (Available code:
hps://doi.org/10.5281/zenodo.6614239):
(i) Preprocessing and Fourier’s descriptors extracon:
Standardizaon was performed on every contour, to
ensure the comparability of data and improve the
classicaon accuracy (Pal & Sudeep, 2016). This
was accomplished by following a series of funcons
that returned a slightly modied set of coordinates
that fulll common main characteriscs: The con-
tour was closed by appending the rst coordinate to
the last one, in case these didn’t coincide; the direc-
on of the coordinates of every contour was veri-
ed and changed to be on a clockwise orientaon;
in case the contour contained less than 200 points,
some points were interpolated. Next, the apex of
the wing is located, and the coordinates are rotated
to make it the starng point. Finally, the contour
was checked again to ensure it had been closed.
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
98Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Aer the preprocessing, the Fourier descriptor’s
coecients were extracted using the Python im-
plementaon for approximang contours with a
Fourier series, PyEFD (Blidh, 2016). This process al-
lowed the extracon of the same number of coef-
cients for each wing, regardless of their size. The
normalized coecients were kept in a separate
database, with each specimen’s unique idener
(uniq-id).
(ii) Database division in training and test datasets: A
train-test division was performed, following a 70/30
proporon: 70% to train the model and 30% for
tesng/validaon.
(iii) Denion, training, and tesng of classier algo-
rithms: Seven classiers were chosen to be trained
and tested for classicaon from the Scikit-learn
distribuon (Pedregosa et al., 2011):
• Logisc Regression (LR) is a binary linear classier,
which is the simplest and is used as a baseline mod-
el. To adjust LR to a mulclass problem, where the
classicaon is done with a one vs rest method, the
opon mul_class = 'ovr' was set.
• Classicaon and Regression Trees (CART) is a mul-
class classier that uses recursive paroning fol-
lowing the Gini Impurity Index to build a decision
tree.
• K-Nearest Neighbors (KNN) is a mulclass classier
that assumes similarity depending on class proxim-
ity, calculated as an Euclidean distance.
• Naïve Bayes (NB) is a mulclass classier that as-
sumes condional independence between every
pair of classes.
• Linear Discriminant Analysis (LDA): is a linear clas-
sier for a mulclass problem. It ensures the maxi-
mum separability of classes by reinforcing the pro-
poron of intra and inter class variance (Narayan,
2020; Tharwat et al., 2017).
• Support Vector Machines (SVM) build a hyperplane
or group of hyperplanes on a higher dimensionality
space that allow the separaon of nonlinear prob-
lems (Gandhi, 2018). The opon StandardScaler
was used to normalize and scale the data; and the
opon SVC, is used to specify the classicaon task.
• Random Forest Classier (RFC) ts several decision
trees on dierent sub-samples of the data. To set
the number of trees in the forest, the opon n_es-
mators = 200 was set.
In addion, for each classier, a cross validaon
score and a classicaon report was obtained with
ve items: Accuracy (number of correct predic-
ons from the total number of predicons), Preci-
sion (number of true posives from all the posive
predicons), Recall (number of posive predicons
from the total number of posive classes), F1 score
(following equaon:
2 (True Posives (TP) × False Posives (FP) ÷ 2 TP +
FP + False Negaves (FN))
and Support (number of individuals in each class).
(iv) Confusion matrices: Confusion matrices were plot-
ted for each classier to obtain a detailed visual-
izaon of the classicaon errors: On them, the
predicted and real classes are found on the x- and
y-axis, respecvely. The correct predicons of the
classier are found on the diagonal where the pre-
dicted and true labels coincide. In consequence, the
predicons that lay outside of this diagonal, corre-
spond to classicaon errors that inform about the
performance of the classiers, as well as possible
confusion paerns.
Finally, we performed ANOVA and Tukey tests in order
to compare the accuracy and F1 scores from the clas-
sicaon report, along with box plots calculated from
the data.
Figure 1. (A) Accuracy (number of correct predicons from
the total number of predicons) and (B) F1 Score (a mea-
sure of a model’s accuracy on a dataset that follows the for-
mula: ((2 × Precision × Recall) ÷ (Precision + Recall)) boxplots
of 3-fold cross validaon for each of the seven classiers
tested. A total of 1397 individuals for each tesng dataset
per classier was used; same leers indicate non-signicant
comparisons, p-values are shown for the CART—NB and
SVM—RFC comparisons which were non-signicant for both
scores.
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
99Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Results
We extracted a dataset of 4656 rows and 81 columns of
Fourier coecients, aer the preprocessing images and
the Fourier extracon loop we dened. Each row of this
dataset belongs to an individual organism and each col-
umn to one of the coecients. In total, we obtained 39
descriptors for each wing (hindwing and forewing) per
individual, to be used later in the classicaon.
To the laer database, we tagged the genus label to
each individual (row) in order to create a training and a
tesng set following a 70:30 proporon, respecvely.
As a result we generated 3259 individuals (70%) for the
training, and 1397 individuals (30%) for the tesng sets.
The accuracy scores were similar enough in all seven
classiers dened (LDA, SVC, LR, CART, NB, RFC, KNN)
between the two sets, which rules out possible over-
ng of the classicaon models (see Supplementary
Tables 2, 3). Furthermore, using the tesng set the clas-
sicaon report obtained showed that the LDA classi-
er had the best performance in terms of: (1) accuracy
(0.7337); (2) precision (0.75); (3) recall (0.73) and F1
score (0.73); in comparison with the other six classi-
ers tested (Supplementary Table 3). The ANOVAs per-
formed for the F1 score and accuracy were signicant
(Fig. 1; Supplementary Tables 4 + 6), across the models.
Figure 2. Confusion matrix. (A) LDA Confusion Matrix. next page. (B) Confusion Matrix showing misclassicaon zones distrib-
uted mainly on four families: Gomphidae (blue), Libellulidae (red), Coenagrionidae (green), Lesdae (orange). Each cell of the
matrix corresponds to every possible true label and predicted label pairing. The color bar on the side of each plot, shows the
code for the number of coincidences on each cell (from 0 = white, to 80 = dark blue).
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
100Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
The post-hoc Tukey mulple comparisons test showed
dierences for the accuracy and the F1 score compari-
son among all the classiers, with the excepon of the
CART and NB comparison, and the SVM and RFC com-
parison (Fig. 1; Supplementary Tables 5 + 7). The lat-
ter performance metrics relies on the calculated con-
fusion matrix per model, the LDA classier confusion
matrix shows the highest number of individuals on the
diagonal, meaning these are true posives (Fig. 2A).
Despite its beer performance we noced consistency
in parcular taxa that create misclassicaon in almost
all classiers, that we call confusion groups (Fig. 2B,
Supplementary Figures 1–6). Parcular genera within
the following four families—Gomphidae, Libellulidae,
Coen agrionidae and Lesdae—seem to be responsible
for the misclassicaon observed (Fig. 3).
Discussion
Image preprocessing funcons allowed a standardiza-
on of the coordinates on the contour dataset. This
process has been found to guarantee data comparabil-
ity and improve classicaon accuracy when compared
with non-preprocessed images (Pal & Sudeep, 2016;
Shahriar & Li, 2020; Sharma et al., 2020). The similarity
of accuracy scores for all the classiers in both training
and tesng sets, suggest that there were not overt-
ng issues in the models tested (Brownlee, 2017). Fur-
thermore, we detected dierent numbers of classes
(genera) for each of the seven classiers. For example,
the classier with the best performance, LDA, created
and recognized a total of 100 (classes proxy of genera)
from 111 genera we included in the taxon sampling.
Figure 2. Connued (see page before).
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
101Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Figure 3. Confusion groups.
(A) Ani soptera: True (real) label
on le column and Predicted
label on the right column. Sur-
rounded by a red square (top):
the Gomphidae genus Arigom-
phus (True label) was predicted
as Hylogomphus, Progomphus,
Ophiogomphus, Stylurus and
Gomphurus (also from the Gom-
phidae family). At the boom
of the gure, surrounded by a
blue square: the genus Libel-
lula (Libellulidae), was confused
with Gomphurus (Gomphidae),
Erythemis (Libellulidae), Aeshna
(Aeshnidae) and Coryphaeshna
(Aeshnidae). (B) Zygoptera: True
(real) label on le and right col-
umn and Predicted label on the
center column. Surrounded by a
red square (top le) Coenagrio-
nidae genus Enallagma was pre-
dicted as Argia, Acanthagrion
and Cyanallagma (Also Coen-
agrionidae genera). Surrounded
by a green square,the genus
Lestes (Lesdae) was predicted
as Coenagrionidae genera Argia
and Enallagma. At the boom of
the gure, surrounded by a blue
square, genus Ischnura (Coen-
agrionidae), was predicted as
Acanthagrion, Argia, Enallagma
(all Coenagrionidae) and Lestes
(Lesdae). Illustraons from
Amanda Whispell.
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
102Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
(see Table 1, Supplementary Table 3). These dierences
may be due to class imbalance, meaning that there is
unequal representaon of genera in the dataset, with
some of them having only one individual in the data-
set (Table 1). Therefore, it is possible that during the
data paroning some groups were not included in the
training dataset, which prevents the label from being
created and in consequence, it would then not be in-
cluded in the classicaon report. Likewise, if any of
the groups were not represented in the tesng dataset
then its label would sll be created, but the values for
the metrics would be zero.
Furthermore, since machine learning algorithms de-
pend on the distribuon of classes in the training set to
esmate the probability of observing examples in each
class, class imbalance causes algorithms to learn that
less well represented classes are not as important as
the majority classes, so the performance will be bet-
ter in the laer (Brownlee, 2017). To solve this incon-
venience, an alternave could be to paron the data
set in a straed way, to ensure that all classes are bal-
anced in the training and test sets. Moreover, it is nec-
essary to increase the number of individuals in the less
represented genera.
According to Kuhn (2016), the accuracy values are
not strongly aected by the number of classes. In his
study, a comparison was made between models with
dierent numbers of classes, which ranged from three
to 26. The results of the research suggest that a greater
number of classes does not have a signicant eect on
accuracy, which slightly decreased its variaon, as the
number of classes increased, staying around 80%. Thus,
it is possible to infer that the inuence on the accuracy
of the number of classes in the present study is also
lo w.
The qualitave assessment of the confusion matrix
(Fig. 2, Supplementary Figures 1–6), reveals classica-
on mistakes in parcular taxa just by looking at wing
contours. The diagonal of the matrices shows the co-
incidences between the real and the predicted labels:
if there individuals appear along this diagonal, that
means the performance of the classier is beer, since
on this diagonal the coincidences between the true la-
bels and the predicons (true posives) will be found
(we expect a 1:1 relaonship if so). Consequently, on ei-
ther side of these true posives diagonals, classicaon
errors (false posives and false negaves) are found.
True negaves, meanwhile, correspond to all the true
instances found on the diagonal, dierent from the one
of interest (Harrington, 2012).
Unlike the present invesgaon, on which the shape
of the contour of the wings from 111 genera of drag-
onies was exclusively evaluated, and seven classiers
tested, Kuhn (2016) made a classicaon of 26 genera
of dragonies, in which characteriscs such as color,
texture and shape of the wings were included, reaching
a maximum accuracy of 91%, using only the Random
Forest algorithm classier. We suggest a possible ex-
planaon for the dierence found in accuracy between
Kuhn’s (2016) and our data is due to addional charac-
ters assessed for the dierenaon of species (texture,
coloraon and wing proporons). Our data suggests
that the contour used here by itself does not provide
enough informaon to obtain the accuracy found in
Kuhn (2016). Thus, we suggest that the combinaon of
the wing contours and the wing aributes previously
assessed by Kuhn (2016) (including a morphometric
analysis using 15 measurements, a chromac analysis
and, nally, the use of the Gabor wavelet transforma-
on on the images with dierent rotaons and scales)
might increase the accuracy of the automac idenca-
on for these taxa. In addion, we did nd that the LDA
classier has beer performance, suggesng the need
to assess other classiers than RF, which include all
the possible wing aributes to test their performance
in the classicaon. We expect to combine our results
with the previous wing aributes tested by Kuhn (2016)
for the automac idencaon to keep decreasing the
taxonomic impediment in the current biodiversity cri-
sis.
The largest number of misclassicaons of our data
are centered on the tested genera within the aniso-
pteran families Gomphidae and Libellulidae and the
zygo pteran Coenagrionidae and Lesdae families. This
is interesng as Gomphidae, Libellulidae and Coenagrio-
nidae are the most species rich families in the Odona-
ta. Our results suggest that most of the confusion and
classicaon errors are distributed among parcular
groups within families belonging to the same suborder
(Fig. 3). In parcular, there are two confusion groups
that belong to the Anisoptera suborder (Fig. 3). In the
rst group (Fig. 3A, red square), six genera of the Gom-
phidae family are included, while in the second group
(Fig. 3A, blue square), there are two genera that belong
to the Libellulidae family, two genera of the Aeshnid-
ae family and one of the Gomphidae family. Likewise,
Kuhn’s (2016) confusion matrix has similar classicaon
mistakes to the ones we observed here (Fig. 3). For ex-
ample, the genus Erythemis with the classier and ari-
butes tested by Kuhn (2016) was confused with species
of the genera Pachydiplax and Libellula; in our results it
was also confused with Libellula and a couple of aesh-
nids (Fig. 3A). For Zygoptera, our observed confusion
occurs mainly between the Lesdae and Coenagrioni-
dae families (Fig. 3B). The occurrence of greater confu-
sion within this suborder may be a consequence of the
low level of variaon in their shape between families.
This fact, in turn, underscores the need for idenca-
on of the Zygoptera facilitated by characteriscs such
as coloraon, types of joints of the veins in the wings,
paerns of venaon, presence of spines and distribu-
on of resilin patches (Appel & Gorb, 2014; Hassall,
2014).
Interesngly, our data suggest that these confusion
groups have similar wing contours, which can lead us to
look for possible hypotheses that explain these similari-
es among these taxa. Some explanaon can be due to
their ecology: for example, within the Anisoptera there
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
103Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Table 1. Genera found by each classier and number of individuals in the dataset. The rst column (“Genus”) has the names of
the 111 genera included in the dataset. The Xs mark where the class was found and the dark gray empty cells show the classes
(genera) that were absent in the classicaon report, for each of the classiers.
Genus Count LDA LR NB CART KNN SVM RFC
Acanthagrion 16 xxxxxxx
Aeshna 131 xxxxxxx
Amphiagrion 1x
Amphipteryx 1x
Anax 39 x x x x x x x
Anisagrion 2xxxxxxx
Aphylla 33 xxxxxxx
Archilestes 20 x x x x x x x
Argia 85 x x x x x x x
Arigomphus 81 xxxxxxx
Basiaeschna 27 xxxxxxx
Boyeria 11 xxxxxxx
Brachymesia 22 x x x x x x x
Brechmorhoga 4x x x x x x x
Calopteryx 149 xxxxxxx
Cannaphila 4x x
Castoraeschna 1x
Celithemis 149 x x x x x x x
Cordulegaster 90 xxxxxxx
Cordulia 8xxxxxxx
Coryphaeschna 25 xxxxxxx
Crocothemis 2x x
Cyanallagma 3x x x x x x x
Diastatops 13 xxxxxxx
Didymops 6xxxxxxx
Dorocordulia 17 x x x x x x x
Drepanoneura 1x x x x x x x
Dromogomphus 45 xxxxxxx
Dythemis 14 xxxxxxx
Enallagma 264 xxxxxxx
Epiaeschna 22 x x x x x x x
Epipleoneura 2x
Epitheca 130 xxxxxxx
Erpetogomphus 57 xxxxxxx
Erythemis 174 x x x x x x x
Erythrodiplax 107 x x x x x x x
Euthore 1xxxxxxx
Fluminagrion 1xxxxxxx
Gomphaeschna 17 xxxxxxx
Gomphurus 160 x x x x x x x
Gynacantha 29 x x x x x x x
Hagenius 17 xxxxxxx
Helocordulia 13 xxxxxxx
Hesperagrion 15 x x x x x x x
Hetaerina 75 x x x x x x x
Heteragrion 2
Hylogomphus 62 xxxxxxx
Idiataphe 1xxxxxxx
Iridictyon 3x x x x x x x
Ischnura 83 x x x x x x x
Ladona 55 xxxxxxx
Lanthus 21 xxxxxxx
Leptobasis 2x
Lestes 187 x x
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
104Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Genus Count LDA LR NB CART KNN SVM RFC
Leucorrhinia 70 xxxxxxx
Libellula 342 xxxxxxx
Macrodiplax 9x x x x x x x
Macromia 40 x x x x x x x
Macrothemis 43 xxxxxxx
Mecistogaster 1
Mesamphiagrion 4xxxxxxx
Miathyria 17 x x x x x x x
Micrathyria 84 x x x x x x x
Misagria 2xxxxxxx
Mnesarete 5xxxxxxx
Nannothemis 18 x x x x x x x
Nasiaeschna 14 x x x x x x x
Nehalennia 1x
Neoerythromma 1
Neoneura 1x
Nephepela 6x x x x x x x
Neurocordulia 7x x x x x x x
Octogomphus 9xxxxxxx
Oligoclada 1x
Ophiogomphus 88 x x x x x x x
Oplonaeschna 1
Orthemis 35 xxxxxxx
Pachydiplax 68 xxxxxxx
Palaemnema 3x
Paltothemis 14 x x x x x x x
Pantala 85 x x x x x x x
Perithemis 81 xxxxxxx
Phanogomphus 233 xxxxxxx
Phyllocycla 17 x x x x x x x
Phyllogomphoides 29 x x x x x x x
Plathemis 88 xxxxxxx
Polythore 220 xxxxxxx
Progomphus 62 xxxxxxx
Protoneura 4x x x x x x x
Pseudoleon 18 x x x x x x x
Remarnia 1xxxxxxx
Rhionaeschna 3
Rhodopygia 2x x x x x x x
Rimanella 1x x x x x x x
Somatochlora 55 x x x x x x x
Staurophlebia 1xxxxxxx
Stenocora 1x x
Stenogomphurus 7x x x x x x x
Stylogomphus 18 x x x x x x x
Stylurus 38 xxxxxxx
Sympetrum 146 xxxxxxx
Tachopteryx 10 x x x x x x x
Tanypteryx 3x x x x x x x
Tauriphila 21 x x x x x x x
Telebasis 6xxxxxxx
Tholymis 9xxxxxxx
Tramea 86 x x x x x x x
Triacanthagyna 13 x x x x x x x
Tuberculobasis 1
Uracis 9xxxxxxx
Zenithoptera 4xxxxxxx
Total Not found 11 19 19 12 17 19 19
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
105Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
are an array of ight behaviors (iers, gliders and perch-
ers; Corbe & May, 2008) and these ight styles can
be reected in the similaries found in wing contours
within both our observed confusion groups. For exam-
ple, in migratory species of libellulids’ hindwings can
show convergence towards a wing planform that favors
the gliding ight as an energy saving strategy (Suarez-
Tovar & Sarmiento; 2016). For zygopterans, their ight
is more passive, and their ability to disperse might be
associated with slow ight or overight (Bomphrey et
al., 2016), which would explain any similaries in wing
contours for coenagrionids and lesds. Comparisons
of the damping raos and natural frequencies of two
dragony and two damsely species, shows that for the
anisopterans damping properes between fore- and
hindwings were signicantly dierent, while in zygo-
pterans there were no or very weak dierences in the
damping raos between both wings, suggesng that
the structural design and wing shape can inuence
the aerodynamics of their ight behaviors (Lietz et al;
2021). In addion, funconal morphology traits of the
wings, like types of joints of the wing veins, spines and
presence of resilin, a protein that gives certain exibility
to the wings of insects can be evaluated in this groups,
like previously done by Appel and Gorb (2014) to un-
derstand the wing contour similaries in these taxa.
Overall, our results suggest that the wing contours by
themselves can discriminate with a moderate accuracy
and precision, in comparison with other wing aributes
obtained using high resoluon images. In addion, we
tested mulple classifying algorithms for the contours,
where LDA had the best performance.
Acknowledgements
The authors would like to acknowledge the funding from NSF Grant
#1564386: ODOMATIC: Automac Species Idencaon, Funcon-
al Morphology, and Feature and NSF DBI WK Postdoctoral Grant
#16116642: Leveraging face-detecon methods to idenfy insects
from eld photos, automacally.
References
Anaconda Soware Distribuon. (2016). hps://anaconda.com
Appel, E. & Gorb, S. N. (2014). Zoologica Comparave funconal
morphology of vein joints in Odonata. Zoologica, 159.
Blidh, H. (2016). Python implementaon of “Ellipc Fourier Features
of a Closed Contour.” hps://pyefd.readthedocs.io/en/latest/
Bomphrey, R. J., Nakata, T., Henningsson, P. & Lin, H. T. (2016). Flight
of the dragonies and damselies. Philosophical Transacons of
the Royal Society B: Biological Sciences, 371(1704). doi:10.1098/
rstb.2015.0389
Brownlee, J. (2017). Master Machine Learning Algorithms (1.12).
Machine Learning Mastery. hps://machinelearningmastery.
com/master-machine-learning-algorithms/
Ceballos, G., Ehrlich, P. R., Barnosky, A. D., García, A., Pringle, R. M.
& Palmer, T. M. (2015). Accelerated modern human-induced spe-
cies losses: Entering the sixth mass exncon. Science Advances,
1, 5. doi:10.1126/sciadv.1400253
Corbe, P. S. & May, M. L. (2008). Fliers and perchers among Odona-
ta: dichotomy or muldimensional connuum? A provisional re-
appraisal. Internaonal Journal of Odonatology, 11(2), 155–171.
doi:10.1080/13887890.2008.9748320
Córdoba-Aguilar, A. (2008). Dragonies and Damselies: Model Or-
ganisms for Ecological and Evoluonary Research. In Dragonies
and Damselies: Model Organisms for Ecological and Evoluonary
Research. doi:10.1093/acprof:oso/ 9780199230693.001.0001
Gandhi, R. (2018). Support Vector Machine—Introducon to Ma-
chine Learning Algorithms. Towards Data Science. hps://to-
wardsdatascience.com/support-vector-machine-introduction-
to-machine-learning-algorithms-934a444fca47
Garrison, Rosser W., von Ellenrieder, N. & Louton, J. A. (2006). Drag-
ony genera of the New World: an illustrated and annotated key
to the Anisoptera. In Choice Reviews Online. Johns Hopkins Uni-
versity Press.
González, A. (2009). El conocimiento sistemáco impedimento ta-
xonómico la biodiversidad y. Revista de La Sociedad Española de
Biologia Evoluva, 4(1), 19–32.
Harrington, P. (2012). Machine Learning in Acon Ill MANNING Shel-
ter Island. Manning Publicaons Co.
Hassall, C. (2014). Connental variaon in wing pigmentaon in ca-
lopteryx damselies is related to the presence of heterospecics.
PeerJ, 2014(1), e438. doi:10.7717/peerj.438
Heckman, C. W. (2006). Encyclopedia of South American Aquac In-
sects: Odonata – Anisoptera. In Encyclopedia of South American
Aquac Insects: Odonata – Anisoptera. The Netherlands: Spring-
er. doi:10.1007/978-1-4020-4802-5
Heckman, C. W. (2008). Encyclopedia of South American Aquac In-
sects: Odonata – Zygoptera. In Encyclopedia of South American
Aquac Insects: Odonata – Zygoptera. The Netherlands: Spring-
er. doi:10.1007/978-1-4020-8176-7
Kalkman, V. J., Clausnitzer, V., Dijkstra, K. D. B., Orr, A. G., Paulson,
D. R. & van Tol, J. (2008). Global diversity of dragonies (Odona-
ta) in freshwater. Hydrobiologia, 595(1), 351–363. doi:10.1007/
s10750-007-9029-x
Koh, L. P., Dunn, R. R., Sodhi, N. S., Colwell, R. K., Proctor, H. C. &
Smith, V. S. (2004). Species Coexncons and the Biodiversity
Crisis. Science, 305(September), 1632–1635. doi:10.1126/sci-
ence.1101101
Kuhn, W. R. (2016). Three approaches to automang taxonomy,
with emphasis on the Odonata (dragonies and damselies).
(Thesis). Rutgers, The State University of New Jersey.
la Salle, J., Williams, K. J. & Moritz, C. (2016). Biodiversity analysis in
the digital era. Philosophical Transacons of the Royal Society B:
Biological Sciences, 371(1702). doi:10.1098/rstb.2015.0337
Lietz, C., Schaber, C.F., Gorb, S.N. et al. (2021) The damping and
structural properes of dragony and damsely wings during
dynamic movement. Commun Biol 4, 737. doi:10.1038/s42003-
021-02263-2
Lorenzo-Carballa, M. O. & Cordero Rivera, A. (2012). Odonatos. In
P. Vargas & R. Zardoya (Eds.), El árbol de la Vida: sistemáca y
evolución de los seres vivos. pp. 293–301.
Moore, N. W. (1997). Dragonies: Status Survey and Conservaon
Acon Plan. Gland, Switzerland, and Cambridge, UK: IUCN.
Narayan, Y. (2021). Hb vsEMG signal classicaon with me domain
and Frequency domain features using LDA and ANN classier.
Materials Today: Proceedings, 37, 3226–3230. doi:10.1016/j.
matpr.2020.09.091
Pal, K. K. & Sudeep, K. S. (2016). Preprocessing for image classica-
on by convoluonal neural networks. 2016 IEEE Internaonal
Conference on Recent Trends in Electronics, Informaon & Com-
municaon Technology (RTEICT), 1778–1781. doi:10.1109/RTEI-
CT.2016.7808140
Sáenz Oviedo, Kuhn, ... & Sanchez-Herrera Are wing contours good classiers for automac idencaon in Odonata?
106Internaonal Journal of Odonatology │ Volume 25 │ pp. 96–106
Paulson, D. R., Schorr, M. & Deliry, C. (2021). World Odonata List
· University of Puget Sound. hps://www2.pugetsound.edu/
academics/academic-resources/slater-museum/biodiversity-
resources/dragonies/world-odonata-list2/
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B.,
Grisel, O., .... Duchesnay, E. (2011). Scikit-learn: Machine Learn-
ing in Python. Journal of Machine Learning Research, 12(85),
2825–2830.
Raybaut, P. (2009). Spyder IDE (4.2.1). Pythonhosted. hps://www.
spyder-ide.org/
Samways, M. J. & Steytler, N. S. (1996). Dragony (Odonata) dis-
tribuon paerns in urban and forest landscapes, and recom-
mendaons for riparian management. Biological Conservaon.
doi:10.1016/S0006-3207(96)00032-8
Shahriar, M. T. & Li, H. (2020). A Study of Image Pre-processing for
Faster Object Recognion. ArXiv, October 2020. hps://arxiv.
org/abs/2011.06928
Sharma, P., Hans, P. & Gupta, S. C. (2020). Classicaon of plant leaf
diseases using machine learning and image preprocessing tech-
niques. Proceedings of the Conuence 2020 – 10th Internaonal
Conference on Cloud Compung, Data Science and Engineering,
480–484. doi:10.1109/Conuence47617.2020.9057889
Suárez-Tovar, C. M. & Sarmiento, C. E. (2016), Beyond the wing plan-
form: morphological dierenaon between migratory and non-
migratory dragony species. Journal of Evoluonary Biology, 29,
690-703. doi:10.1111/jeb.12830
Tharwat, A., Gaber, T., Ibrahim, A. & Hassanien, A. E. (2017). Linear
discriminant analysis: A detailed tutorial. AI Communicaons,
30(2), 169–190. doi:10.3233/AIC-170729
van Rossum, G., & Drake Jr, F. L. (2009). Python 3 Reference Manual.
CreateSpace. hps://dl.acm.org/doi/book/10.5555/1593511
Supplementary material
Supplementary Figure 1. Random Forest Classier Confusion Matrix.
Supplementary Figure 2. Support Vector Machines Confusion Matrix.
Supplementary Figure 3. K-Nearest Neighbors Confusion Matrix.
Supplementary Figure 4. Classicaon and Regression Trees Confu-
sion Matrix.
Supplementary Figure 5. Naïve Bayes Confusion Matrix.
Supplementary Figure 6. Logisc Regression Confusion Matrix.
Supplementary Table 1. Taxonomic informaon of specimens in-
cluded in the analysis.
Supplementary Table 2. Training accuracy scores.
Supplementary Table 3. Summary of classicaon report: Number
of classes found, accuracy, precision, recall, F1 score and support
values of the classiers tested.
Supplementary Table 4. ANOVA results for accuracy scores compari-
son.
Supplementary Table 5. Tukey mulple comparisons test for accu-
racy scores.
Supplementary Table 6. ANOVA results for F1 scores comparison.
Supplementary Table 7. Tukey mulple comparisons test for F1
scores.
