43
B. Mendoza, E. Morales, C. Cruz, and L. Gomez,
“Morphological classification of hematophagous Diptera with Convolutional Neural Networks: A mapping of literature,”
Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.
Morphological
classication of
hematophagous Diptera
with Convolutional
Neural Networks: A
mapping of literature
ARTICLE HISTORY
Received 14 October 2025
Accepted 23 February 2026
Published 7 July 2026
Benjamín Paulino Mendoza Contreras
Veracruzana University
Faculty of Statistics and Informatics
Xalapa, Veracruz
benjaminpaulinom6@gmail.com
ORCID: 0009-0000-3491-6234
Emmanuel Morales García
Veracruzana University
Faculty of Statistics and Informatics
Xalapa, Veracruz
emmorales@uv.mx
ORCID: 0000-0002-6837-9227
Cecilia Cruz López
Veracruzana University
Faculty of Statistics and Informatics
Xalapa, Veracruz
ceccruz@uv.mx
ORCID: 0000-0002-9156-5669
Luis Enrique Gomez Medina
Veracruzana University
Institute for Research and Higher Studies in Administrative
Sciences
Xalapa, Veracruz
luisgomez04@uv.mx
ORCID: 0009-0009-1324-389X
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
This work is licensed under a Creative Commons
Attribution-NonCommercial-ShareAlike 4.0 International License.
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
44
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026
Morphological classification of hematophagous
Diptera with Convolutional Neural Networks: A
mapping of literature
Benjamín Paulino Mendoza Contreras
Veracruzana University
Faculty of Statistics and Informatics
Xalapa, Veracruz
benjaminpaulinom6@gmail.com
Cecilia Cruz López
Veracruzana University
Faculty of Statistics and Informatics
Xalapa, Veracruz
ceccruz@uv.mx
Emmanuel Morales García
Veracruzana University
Faculty of Statistics and Informatics
Xalapa, Veracruz
emmorales@uv.mx
Luis Enrique Gomez Medina
Veracruzana University
Institute for Research and Higher Studies in Administrative Sciences
Xalapa, Veracruz
luisgomez04@uv.mx
Abstract— This review analyzes studies that primarily address
the morphological classification of hematophagous Diptera, with
limited mention of other insects. These networks have become
increasingly important in morphological analysis through the
accurate and efficient automatic identification of species, surpassing
even traditional methods based on human observation. The main
architectures used, such as VGG-16, YOLOv5, Faster R-CNN,
Mask R-CNN, ResNet, and Swin Transformer-L are reviewed,
highlighting their applications in the detection and identification of
different anatomical parts. Common limitations are also mentioned,
such as the need for large volumes of classified data and variability
in image quality. Finally, current trends have been identified that
point to the development of more robust hybrid models capable of
recognizing new species and improving accuracy under real-world
conditions. This literature mapping provides greater certainty and
evidence regarding the most important identification methods in the
field of entomology. These findings highlight the gap in literature
related to the availability of public data, parameters used, data
volume, image quality, and model evaluation, providing a solid
foundation to guide future research in the field of entomology.
Keywords—Entomology, Organism Classification, Deep
Learning, Species, Identification, Morphology
I. INTRODUCTION
Image-based classification of hematophagous Diptera is
fundamental and has evolved thanks to the various
Convolutional Neural Networks (CNNs) used for image
recognition. These networks allow for the analysis of visual
characteristics (morphology) of species at a higher level and
more quickly, even under varying image conditions [1].
Classifying hematophagous Diptera from digital images
allows for faster, more accurate, and scalable identification
than traditional methods based on human observation. These
networks are capable of automatically extracting complex
visual features (shapes, textures, or patterns) that determine
the differences between species, without requiring the
researcher to manually define the important characteristics
[1].
One of the main advantages of CNNs is their ability to
process large volumes of data from different image capture
devices. This represents a significant change in field data
collection, as it automates repetitive tasks and reduces the
human workload, allowing experts and researchers to achieve
greater accuracy.
Some authors have developed hierarchical architectures that
incorporate taxonomic relationships between genera and
species within the model itself, improving performance and
reducing errors when classifying at more specific taxonomic
levels [2],[3]. This type of innovation is important for
classifying morphologically similar organisms, where visual
differences may be minimal.
Convolutional neural networks (CNNs) are one of the most
influential innovations in the field of deep learning, due to
their ability to automatically and efficiently process, analyze,
and classify visual data. Their main advantage lies in their
ability to extract hierarchical features directly from images,
reducing the need for human intervention in selecting
relevant features [4].
In the scientific and technological fields, CNNs have
demonstrated exceptional performance in tasks such as facial
recognition, medical imaging, organism classification, object
detection, and autonomous driving. Their structure, based on
convolutional, clustering, and fully connected layers, allows
for the identification of complex patterns [5],[6].
The objective of this literature mapping is to show the trends,
methodological approaches, data types, and CNN
architecture used in the morphological classification of
Diptera reported in recent research. Although morphological
classification includes various biological groups, recent
studies show a significant concentration on hematophagous
Diptera, due to their relevance to public health issues.
Therefore, this research focuses on this group.
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
45
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
B. Mendoza, E. Morales, C. Cruz, and L. Gomez,
“Morphological classification of hematophagous Diptera with Convolutional Neural Networks: A mapping of literature,”
Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.
II. M
ETHODOLOGY
This study employs an exploratory and descriptive literature
review approach, aiming to identify trends, methodologies,
convolutional neural network (CNN) architectures, data types
used, and gaps in the literature regarding morphological
classification of dipterans using images. This type of review
seeks to provide a structured overview of the current state of
the art. Fig. 1 shows the general phases of the mapping review
process followed in this study.
Fig. 1. Article selection process.
A. Research Questions
RQ1. What pre-trained convolutional neural network
architectures have been used to detect morphology in
Diptera?
RQ2. What CNN-based approaches analyze Diptera
morphology?
RQ3. What types of results are reported in studies on wing
morphology?
RQ4. What computational approaches have been proposed
for detecting morphological patterns in Diptera?
B. Inclusion and exclusion criteria
TABLE I. SELECTION CRITERIA FOR PRIMARY STUDIES
Category
Inclusion
Exclusion
Type of
research
Practical research
on image
classification
methods using
convolutional
neural networks
Non-primary
studies: literature
review.
Publication year
interval
Articles
published from
2020 to 2025 to
ensure current
relevance
Studies published
before 2020.
Language
Articles in
English
Articles in
Spanish or
another language.
Search engines
Primary Search
Engines
(Publishers)
Search engines of
dubious scientific
quality
Applications
Emphasis on
medical imaging
and other areas of
study
Thematic
relevance
Research on the
classification of
diptera
Investigations
that divert
attention from the
main topic
C. Search Strategy
Fig. 2. Distribution of selected articles by digital library.
For this study, a search strategy was employed that was
designed to achieve broad and specific coverage of the
relevant literature, following a descriptive process (Fig. 2).
The following elements were considered:
1. Repositories used: ACM Digital Library, IEEE
Xplore, SpringerLink, and ScienceDirect.
2. Year range: Publications between 2020 and 2025,
with the aim of including recent work.
3. Language: Only publications in English.
4. Keywords: Terms related to CNNs, image
classification, Diptera, and public health.
To address the variability of terms present in the literature, a
set of keywords and synonyms were proposed, which were
combined using Boolean operators. The general search string
used was:
"convolutional neural network" OR CNN OR "deep learning"
OR "computer vision") AND ("image classification" OR
"image recognition" OR "object detection") AND
(morphology OR "morphological identification" OR
"morphometric analysis") AND (Diptera OR mosquito OR
mosquitoes OR Culicidae OR "hematophagous insects")
AND (wing OR winVuelgs OR "wing morphology" OR
"body morphology"
This string is adapted to the specific syntax of each
repository used.
The annual trend in publications in the selected subsample
was analyzed, as shown in Fig. 3. The frequency analysis by
year for the period 2020-2025 shows a growing trend in the
publication of articles that use Convolutional Neural
Networks (CNN) in the analyzed repositories.
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
46
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026
Fig. 3. Annual trend in publication frequency (2020-2025) of the subsample
of articles selected from the four repositories
In addition to the trend analysis, the authors with the
highest number of scientific publications were identified using
CNNs to classify organisms in the four academic repositories
shown in Fig. 4. The word cloud in Fig. 4 reveals a clear
concentration of publications by Asian authors, particularly
those with the surnames Zhang, Liu, Li, and Wang.
Fig. 4. Word cloud of the most frequent authors in the publication of articles
in the four repositories (2020-2025)
D. Study Selection Process
The search conducted across the selected academic
repositories yielded a total of 2,457 records (Table
II). To ensure the relevance of the studies included
in this mapping review, a multi-stage filtering
process was applied following the PRISMA 2020
guidelines (Fig. 5). First, duplicate records were
removed, resulting in the exclusion of 412 articles,
leaving 2,045 unique records for the screening
phase.
Second, a thematic filtering stage was performed
based on the analysis of titles, abstracts, and
keywords, which led to the exclusion of 1,820
records that were not aligned with the objectives of
this study. As a result, 225 articles were considered
potentially relevant and were retrieved for full-text
evaluation.
Third, during the eligibility assessment, the full
texts of the 225 remaining studies were examined
in detail. At this stage, 217 studies were excluded
because they did not meet the inclusion criteria.
The main reasons for exclusion included studies
not focused on morphological classification using
images, studies that did not employ deep learning
methods, articles lacking experimental evaluation
or performance metrics, studies based on non-
image data, and research outside the scope of insect
identification. After applying these criteria, a final
set of eight studies was selected for qualitative
analysis. These studies were analyzed to identify
the architecture, datasets, and methodological
trends used in automated morphological
classification of insects using artificial intelligence.
TABLE II. NUMBER OF ARTICLES RECOVERED FROM SELECTED
ACADEMIC REPOSITORIES
(2020–2025)
Repository
Total, number of items
ACM
165
IEEE Xplore
116
SpringerLink
397
ScienceDirect
1779
Fig. 5. Study filtering process
E. Descriptive analysis of the studies
Table III below shows information about the selected studies,
allowing you to view key information about each one. It
should be noted that no quantitative comparison is made.
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
47
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
B. Mendoza, E. Morales, C. Cruz, and L. Gomez,
“Morphological classification of hematophagous Diptera with Convolutional Neural Networks: A mapping of literature,”
Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.
TABLE III. ARTICLES DATA MATRIX
Author(s)
Organism
Analyzed part
Dataset
Task type
Metrics
Adhane, G., Dehshibi, M. M.,
& Masip, D. [1]
Mosquitoes (A.
albopictus)
Body (legs, abdomen)
Public
Classification
94.61% Accuracy
Minakshi, M. et al. [21]
Mosquitoes (9 species)
Thorax, wings,
abdomen
Private
Detection/Feature
extraction
95% Accuracy,
60% mAP
Sauer, F. G. et al. [22]
Mosquitoes (Aedes)
Wings
Private
Classification
99% F1-score
Nolte, K. et al. [23]
Mosquitoes (4 species)
Wings vs body
Private
Classification
87.6% (Wings) /
78.9% (Body)
Cannet, A. et al. [25]
Mosquitoes (Aedes
genus)
Interference patterns
(wings)
Private
Classification
95% Accuracy
Zhao, D. et al. [28]
Mosquitoes (17 species)
Whole body
Private
Classification
99.04% Accuracy
Lee, S., Kim, H., & Cho, B.-
K. [29]
Mosquitoes (11 species)
Whole body
Mixed
Detection
97.1% F1-score
Goodwin, A. et al. [30]
Wildlife species
Body
Private
Multilevel
97.04% Accuracy
(Known classes)
III. CONCEPTUAL FOUNDATIONS
A Convolutional Neural Network (CNN) is a type of deep
learning model designed primarily to process data with a grid-
like structure, such as images or spatial and temporal signals.
CNNs are inspired by the functioning of the visual cortex of
the human brain, which responds selectively to visual patterns
such as edges, textures, and shapes [4].
It is a machine learning system that mimics human visual
perception, capable of learning hierarchical patterns and
complex representations from large volumes of visual data.
Thanks to their generalization capacity and efficiency, CNNs
have become an essential tool in modern artificial intelligence
and in various areas of application [4].
The fundamental principle of a CNN is the convolution
operation, whereby the model applies filters (also called
kernels) to images to automatically extract relevant features at
different levels of abstraction. In the first layers, simple
features such as edges or colors are detected, while in the
deeper layers, more complex shapes such as complete objects
are recognized [4].
These networks have demonstrated robust performance in
computer vision tasks such as facial recognition, medical
image diagnosis, organism classification, object detection,
and autonomous driving. Unlike traditional methods, CNNs
do not require an expert to manually define visual features, as
they learn directly from the data, which increases accuracy
and reduces human bias. In addition, CNNs achieve very high
levels of accuracy in visual recognition tasks, outperforming
image classification or object detection. Thanks to the use of
graphics processing units (GPUs), these networks can handle
large volumes of visual information in a short time, making
them efficient and scalable [7].
However, CNNs also have limitations. They require large
amounts of labeled data to achieve adequate performance, as
well as high computational costs, which can limit their
application in contexts with limited data and resources.
Furthermore, when the dataset is small or lacks variety, the
model may overfit, i.e., learn specific patterns from the
training that do not generalize correctly to new data [3].
Another limitation relates to the lack of interpretability of
the results. CNNs are considered “black box” models because
it is not always possible to know exactly how they make
decisions, which creates uncertainty in areas where explaining
the process is as important as the prediction, such as in
medicine or biology. Similarly, their performance can be
affected by variability in external conditions, such as camera
angle, the image capture device used, or the background of the
images. Finally, if the training data is biased or imbalanced
between classes, the model may reproduce those same biases
in its results, affecting the accuracy of the predictions [8].
CNNs are a fundamental tool for automated image
processing and analysis, with applications in numerous fields
of knowledge. They have transformed the analysis of
biological images by offering an automated, efficient, and
accurate way to process large volumes of visual information.
In recent years, these architectures have become an important
tool in computational biology, ecology, and digital taxonomy,
enabling species identification, organism counting, and
morphological characterization from photographs or video
sequences. Likewise, they have shown outstanding results in
the identification of animal species captured by camera traps
or drones, achieving greater accuracy than experts [9].
Another limitation in the analysis of biological images is
the scarcity of labeled data, since obtaining high-quality
images for each species or taxonomic group is costly and
requires expert knowledge. This problem becomes more
relevant when there are unbalanced classes, causing the model
to favor the classes with more images and reduce the accuracy
of the minority classes [10].
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
48
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026
IV. LITERATURE
REVIEW AND DESCRIPTIVE
SYNTHESIS OF THE RESEARCH QUESTIONS
A. RQ1. What pretrained convolutional neural network
architectures have been used to detect morphology in
diptera?
It is designed for large-scale image classification and
recognition. It consists of 16 weighted layers (13
convolutional and 3 dense). Its key architectural principle is
the exclusive use of 3×3 convolutional filters stacked in
blocks, followed by Max Pooling layers (2×2). The process
begins with an input image (224×224 RGB), from which
features are extracted to be finally classified by the dense
layers with Softmax activation [11].
Faster R-CNN
It is a two-stage object detection model based on
Convolutional Neural Networks (CNN):
1. Region Proposal Network (RPN): Uses anchors with
various scales and aspect ratios to propose regions,
classifying them and adjusting their coordinates
(regression).
2. Detection (Fast R-CNN): It uses ROI Pooling/Align
to standardize the proposals, then performs the final
classification of the object and the final refinement of
the bounding box.
The parameters include the choice of backbone, the
configuration of the anchors, and the use of Soft L1 Loss for
box regression [12].
YOLOv5
It is a single-stage real-time object detector with an
architecture divided into three parts:
1. Backbone: Uses CSPDarknet53 for efficient feature
extraction.
2. Neck: Employs SPPF and PANet to fuse and enhance
features at different scales, which is key to its high
performance.
3. Head: This is the final layer that performs
simultaneous prediction of class, objectivity, and
bounding box coordinates.
The parameters are defined by model variants that control
the depth and width of the network, using the CIoU loss
function (for localization) and BCE (for classification) [13].
Darknet
It is the convolutional architecture that serves as the
backbone for the YOLO algorithm, optimized for real-time
object detection. The Darknet process (especially in versions
such as Darknet-53) is based on a fully convolutional
architecture that incorporates residual connections (like
ResNet) to increase depth. Its key parameters include the use
of 1×1 layers for dimensionality reduction and the Leaky
ReLU (or Mish) activation function to improve gradient flow,
allowing YOLO to predict bounding boxes and classes in a
single pass through the network [14].
ResNet101 DC-5
It is a 101-layer Deep Residual Neural Network that uses
residual connections to prevent accuracy degradation in deep
learning model training.
It uses ResNet base architecture with bottleneck blocks.
The DC-5 feature involves replacing the pooling layers in
Stage 5 with dilated convolutions. Its main application is
image classification, although it can be optimized for
detection and semantic segmentation tasks [15].
ResNeXt101
It is an architecture that extends ResNet to improve
accuracy by introducing cardinality as a new dimension of
scalability. It is based on Transformation Aggregation. Each
residual block divides it into multiple parallel and identical
branches that process different aspects of the input before
merging. Its main parameter is Cardinality (C), which is the
number of independent groups or branches, usually using a
value of C=32 [15].
ResNet18
It is the smallest and most efficient version of the Residual
Neural Network family, designed for image classification and
to serve as a fast backbone. Its most important process is the
use of residual connections so that the gradient flows directly
through its 18 layers with weights, solving the problem of
accuracy degradation in deep networks. Unlike larger
versions, it uses basic residual blocks (two 3×3 layers) and has
approximately 11.7 million parameters, which gives it speed
and efficiency in training [16].
RetinaNet
It is a single-stage image detector designed to achieve high
accuracy in dense object detection. Its architecture combines
a ResNet backbone with a Feature Pyramid Network (FPN) to
process information at multiple scales. Its key process and
characteristic parameter are Focal Loss, a modified loss
function that resolves the extreme imbalance between positive
and negative examples by heavily weighing difficult
(misclassified) samples and reducing the weight of easy
(background) samples, enabling effective training [17].
Mask R-CNN
It is an advanced model for instance detection and
segmentation that extends Faster R-CNN, which detects
objects and generates a precise pixel mask for each one. The
architecture uses a ResNet-101 backbone combined with a
Feature Pyramid Network (FPN) to create a multiscale feature
pyramid (neck).
Its process is ROI Align, a parameter that replaces ROI
Pooling to ensure precise alignment of the characteristics of
the Regions of Interest (ROI) in order to achieve accuracy in
the mask. It operates with three parallel heads (classification,
box regression, and an FCN for the mask) for application in
Instance Segmentation [18].
Swin Transformer-L
It is a hierarchical Vision Transformer architecture with
high capacity, designed to be a backbone for achieving high
performance in tasks such as object detection and
segmentation. Its hierarchical architecture mimics CNNs and
its process is Shifted Window Attention (SW-MSA). This
mechanism limits the self-attention calculation to local
windows, solving the quadratic complexity of traditional
Transformers and allowing communication between
neighboring windows, resulting in a network with
approximately 197 million parameters and superior efficiency
and accuracy in vision applications [19].
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
49
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
B. Mendoza, E. Morales, C. Cruz, and L. Gomez,
“Morphological classification of hematophagous Diptera with Convolutional Neural Networks: A mapping of literature,”
Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.
MobileNet
It is a lightweight CNN architecture for detection,
classification, and segmentation on mobile and edge devices
with limited resources. Its process and characteristic feature
are the use of Depth Separable Convolutions (DSC), which
divide the convolution operation into two steps (depth and
point), significantly reducing the number of parameters and
the computational load compared to standard convolutions. Its
scaling parameters (Width Factor and Resolution Factor)
allow the model to be adjusted for different latency and power
constraints [20].
B. RQ2. What CNN based approaches analyze Diptera
morphology?
Automatic identification of mosquito species has become
a central field of research, focusing on the use of convolutional
neural networks (CNN) and other deep learning techniques to
overcome the challenges of manual classification. Recent
studies demonstrate the feasibility and high accuracy of these
methods by evaluating the effectiveness of different model
architectures and specific anatomical parts used for
identification [1].
A key approach focuses on the use of citizen science
images, which introduce variability and field conditions. [1]
proposed a method for the automated classification of Aedes
albopictus mosquitoes using the VGG16 architecture. Unlike
other studies based on laboratory images, this model was
trained with a large dataset of field images collected by
volunteers through the citizen science initiative “Mosquito
Alert.” Despite the way these images were obtained, the
model's VGG16 architecture achieved a test accuracy of
94.61%, demonstrating the feasibility of using neural
networks to classify mosquitoes. They applied the Grad-CAM
algorithm, and this analysis revealed that the model focuses
on the white stripes located on the mosquito's legs, abdomen,
and thorax, the same characteristics that entomologists use for
identification. It was found that classification errors were
directly related to poor image quality, such as lack of clarity,
occlusion, or damage to key parts of the mosquito's body. As
a result, images of non-tiger mosquitoes with morphological
similarities could be misclassified, highlighting the
importance of image quality for model accuracy.
In the field of anatomical feature extraction, [21]
developed a Mask R-CNN-based framework to automatically
detect and extract anatomical components of mosquitoes
(thorax, wings, abdomen, and legs) from images obtained with
smartphones. They used 1,600 images of nine species for
training and validation, and evaluated performance with
metrics such as Precision, Recall, IoU, and mAP. The system
used ResNet-101 combined with Feature Pyramid Network
(FPN) and achieved 95% accuracy for thorax, abdomen, and
wings, with a mAP of 60% in validation and 52% in testing.
C. RQ3. What types of results are reported in studies on
wing morphology?
Wing morphology has proven to be a valuable feature for
automatic species classification. [22] developed a
convolutional neural network (CNN) to identify seven species
of Aedes from wing images. They used 1,155 images of Aedes
and 554 non-Aedes mosquitoes, and trained CNNs in
grayscale and RGB. This model achieved an F1-score of 99%
for differentiating Aedes from other mosquitoes and around
90–91% for classifying the seven species, with 100%
accuracy for Aedes albopictus. Classification errors occurred
mainly among similar native species.
Comparing effectiveness, [23] evaluated full-body and
wing images, finding that models trained with wing images
achieved higher accuracy (87.6%) than with body images
(78.9%) for the identification of four morphologically similar
Aedes species. Wing-based models required fewer images for
reliable performance. Likewise, model performance
decreased significantly when evaluated with images from
devices not included in training, although the study highlights
the viability of body- and wing-based classification methods.
[24] evaluated the geometric morphometry of wings to
identify six species of the genus Aedes in northeastern France.
They used 18 reference points on the wings, applied
Procrustes overlap analysis and Canonical Variant Analysis,
achieving 98% accuracy in reclassification.
[25] developed an automatic system to identify Aedes
species using wing interference patterns (WIPs) and deep
learning. With a set of 494 images, they trained several CNN
architectures, including MobileNet, ResNet18, and reduced
versions of DarkNet. The models achieved 95% accuracy at
the genus level, with perfect classification in half of the
species.
[26] proposed a two-stage method for the automatic
classification of midge species of the genus Culicoides based
on morphological analysis of their wings. They applied image
preprocessing techniques (filters and morphological
operations) and machine learning, achieving 95.31% accuracy
for wing segmentation and 94.79% for particle segmentation.
Additionally, [27] focused on the analysis of wingbeat
patterns, presenting a hybrid method for the classification of
mosquito species that combines different machine learning
and deep learning architectures (SVM, MLP, Random Forest,
Gradient Boosting, and kNN). They showed that hybrid
architecture outperforms individual algorithms, as they
achieved high accuracy and balanced performance in multi-
class classification.
D. RQ4. What computational approaches have been
proposed for morphological patterns in Diptera?
The integration of sophisticated architectures, such as
Transformers, has set new standards for accuracy. [27]
developed a deep learning model for automatic mosquito
species identification based on Swin Transformer. They
created a balanced dataset of 9,900 high-resolution images of
17 species and 3 subspecies. When comparing various
convolutional networks and transformer-based models, the
Swin Transformer-L variant was selected for its higher
accuracy (called Swin MSI), which achieved 99.04%
accuracy. In addition, this model demonstrated robustness by
achieving 96.26% accuracy when classifying species not
included in the training.
Along the same lines of advanced models, [29] developed
a deep learning image analysis method to identify eleven
mosquito species in Korea. They trained and compared five
object detection models: Faster R-CNN with Swin
Transformer, YOLOv5, ResNet101 DC-5, ResNeXt 101, and
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
50
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026
RetinaNet. The results showed that the combination of Swin
Transformer + Faster R-CNN achieved the highest accuracy
with an F1-score of 97.1%, and YOLOv5 with 96.4%. They
found that the combined use of RGB and fluorescent images,
together with the non-maximum suppression (NMS)
technique, improved the performance of all models. They
identified the small sample size in some species as a
limitation.
Finally, [29] developed a system to identify species using
convolutional neural networks with a multilevel model that
detects unknown species. They used a database of 12,977
images of 2,696 wild species, many with morphological
damage. The system combines CNNs for feature extraction
with classifiers (SVM, Random Forest) and a Gaussian
mixture model for low-confidence cases. It achieves 97.04%
accuracy in classifying 16 known species and 89.50%
accuracy in detecting novel species.
One of the trends for future research is to improve the
robustness and generalization of deep learning models. The
study by [1] already highlighted that, despite using a large
field dataset from the “Mosquito Alert” initiative,
classification errors were directly related to poor image
quality. This points to the need to develop models that are
more robust to variability. Future research should focus on:
preprocessing and data augmentation techniques; image
quality detection models that can filter or warn about
problematic images before classification and data collection,
as suggested by [29] when identifying the small sample size
in some species as a limitation, and the suggestion to improve
the capture process to increase the volume of training data.
There is a trend to further investigate and validate the
effectiveness of specific anatomical parts as descriptors,
particularly wings. Findings from [22], [23] demonstrated that
models trained with wing images achieved higher accuracy
and required less data than full-body images. This will aid
research into: the systematic comparison between the
geometric morphometrics of [25], wing interference patterns
(WIPs, Cannet et al., 2023), and wing-based CNNs to
determine the most efficient technique; the application of
anatomical extraction frameworks [20] to isolate and improve
the image quality of wings and legs before classification; and
the exploration of new motion-based descriptors, such as the
flapping pattern analysis proposed by [26].
The implementation of architectures such as Transformer-
based models to improve the identification of unclassified
species. The Swin MSI model [27] has already demonstrated
superior performance (99.04% accuracy) and generalization
ability (96.26% on unseen species). Future research will focus
on: exploring hybrid models that combine the high accuracy
of Transformers with the ability to classify unknown or low-
confidence species, as did the multilevel model of [30], which
achieved 89.50% accuracy in detecting novel species;
optimizing the combination of different image modalities,
following the example of [28] with the use of RGB and
fluorescent images to improve performance.
V. CONCLUSIONS
This research presents a mapping of the literature on the
application of convolutional neural networks (CNNs) for
classifying dipteran morphology. A descriptive analysis of
the included studies was used, based on which the
predominant architectures, the anatomical parts used, and the
most frequently employed computational approaches were
mapped, providing a current overview of the field.
According to the results found, the literature relies on pre-
trained CNN architectures, such as VGG-16 and ResNet, as
well as object detection models such as YoloV5, Faster R-
CNN, and Mask R-CNN, which have demonstrated good
performance in classification and morphological detection
tasks. These architectures are mainly used through transfer
learning, which allows for robust results even in scenarios
with limited datasets. In this regard, the analysis shows
evidence that wing morphology-based approaches
consistently report better performance metrics, with accuracy
and F1 scores exceeding 90% in most of the consulted
articles. This confirms that the wings constitute a highly
discriminating anatomical region, allowing for more precise
and efficient models compared to using the whole body. In
contrast, using whole-body images tends to employ more
complex pipelines, based on higher detection or other
strategies, to handle visual and structural variability.
Methodologically, the mapping shows a predominance of
direct classification approaches, complemented by detection
and segmentation, especially when it is necessary to locate
specific structures or work with damaged specimens. Articles
were identified that study hybrid models based on CNNs and
Transforms, which represent a promising line of research for
capturing complex spatial relationships, although their
adoption is limited. The findings of this mapping are useful
for researchers in computer vision and machine learning, as
well as for entomologists, biologists, and digital taxonomy
specialists, as it provides an overview of current trends in
CNNs applied to image classification of dipteran
morphology. This mapping contributes to a better
understanding of the current state of applications in dipteran
morphological classification, providing a solid foundation for
future research in medical entomology.
REFERENCES
[1] Adhane, G., Dehshibi, M. M., & Masip, D. (2021). A deep
convolutional neural network for classification of Aedes
albopictus mosquitoes. IEEE Access, 9, 72681–72690.
[2] Elhamod, M., et al. (2022). Hierarchy-guided neural
network for species classification. Methods in Ecology and
Evolution, 13(3), 642–652.
[3] Zhou, Z., et al. (2023). EchoAI: A deep-learning based
model for classification of echinoderms in global oceans.
Frontiers in Marine Science, 10, 1147690.
[4] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep
learning. Nature, 521(7553), 436–444.
[5] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017).
ImageNet classification with deep convolutional neural
networks. Communications of the ACM, 60(6), 84–90.
[6] Oliveira, M. B., et al. (2025). Classification of animal
species via deep neural networks and species distribution
modeling: A systematic review. Artificial Intelligence
Review, 58, 230.
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
51
DOI:
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
https://doi.org/10.33333/lajc.vol13n2.03
B. Mendoza, E. Morales, C. Cruz, and L. Gomez,
“Morphological classification of hematophagous Diptera with Convolutional Neural Networks: A mapping of literature,”
Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.
[7] Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep
learning. MIT Press.
[8] Litjens, G., et al. (2017). A survey on deep learning in
medical image analysis. Medical Image Analysis, 42, 60–88.
[9] Miao, Z., et al. (2019). Insights and approaches using deep
learning to classify wildlife. Scientific Reports, 9, 8137.
[10] Simonyan, K., & Zisserman, A. (2015). Very deep
convolutional networks for large-scale image recognition. In
Proceedings of the International Conference on Learning
Representations (ICLR).
[11] Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-
CNN: Towards real-time object detection with region
proposal networks. In Advances in Neural Information
Processing Systems (Vol. 28, pp. 91–99).
[12] Khanam, R., & Hussain, M. (2024). What is YOLOv5:
A deep look into the internal features of the popular object
detector. arXiv:2407.20892.
[13] Redmon, J., Divvala, S., Girshick, R., & Farhadi, A.
(2016). You only look once: Unified, real-time object
detection. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR) (pp. 779–
788).
[14] Xie, S., et al. (2017). Aggregated residual
transformations for deep neural networks. In Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) (pp. 5987–5995).
[15] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep
residual learning for image recognition. In Proceedings of the
IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) (pp. 770–778).
[16] Lin, T. Y., et al. (2017). Focal loss for dense object
detection. In Proceedings of the IEEE International
Conference on Computer Vision (ICCV) (pp. 2980–2988).
[17] He, K., et al. (2017). Mask R-CNN. In Proceedings of
the IEEE International Conference on Computer Vision
(ICCV) (pp. 2961–2969).
[18] Liu, Z., et al. (2021). Swin transformer: Hierarchical
vision transformer using shifted windows. In Proceedings of
the IEEE/CVF International Conference on Computer Vision
(ICCV) (pp. 10012–10022).
[19] Howard, A. G., et al. (2017). MobileNets: Efficient
convolutional neural networks for mobile vision applications.
arXiv:1704.04861.
[20] Minakshi, M., et al. (2020). A framework based on deep
neural networks to extract anatomy of mosquitoes from
images. arXiv:2007.14725.
[21] Sauer, F. G., et al. (2024). A convolutional neural
network to identify mosquito species (Diptera: Culicidae) of
the genus Aedes by wing images. Scientific Reports, 14(1),
3094.
[22] Nolte, K., et al. (2025). Potentials and limitations in the
application of convolutional neural networks for mosquito
species identification using wing images. PLoS
Computational Biology, 21(9), e1013435.
[23] Martinet, J.-P., et al. (2021). Wing morphometrics of
Aedes mosquitoes from north-eastern France. Insects, 12(4),
341.
[24] Cannet, A., et al. (2023). Wing interferential patterns
(WIPs) and machine learning for the classification of some
Aedes species of medical interest. Scientific Reports, 13(1),
11956.
[25] Venegas, P., et al. (2020). An approach to automatic
classification of Culicoides species by learning the wing
morphology. PLoS ONE, 15(11), e0241798.
[26] Gireesh, A., & Noortaj, S. (2025). Hybrid machine
learning approach for mosquito species classification using
wing beat analysis. International Journal of Scientific
Research in Science, Engineering and Technology, 12(3),
1062–1070.
https://ijsrset.com/index.php/home/article/view/IJSRSET25
12124
[27] Zhao, D., et al. (2022). A Swin transformer-based model
for mosquito species identification. Scientific Reports, 12(1),
18664.
[28] Lee, S., Kim, H., & Cho, B.-K. (2023). Deep learning-
based image classification for major mosquito species
inhabiting Korea. Insects, 14(6), 526.
[29] Goodwin, A., et al. (2021). Mosquito species
identification using convolutional neural networks with a
multitiered ensemble model for novel species detection.
Scientific Reports, 11(1), 24040.
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
52
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
AUTHORS
Benjamín Paulino Mendoza Contreras es estudiante de octavo semestre
de la Licenciatura en Estadística en la Universidad Veracruzana, con
sede en Xalapa, Veracruz, México. Sus intereses de investigación
se centran en el aprendizaje automático (machine learning) y el
aprendizaje profundo (deep learning), particularmente en el desarrollo
y aplicación de modelos estadísticos y computacionales para el
análisis de datos. Actualmente participa en actividades académicas
relacionadas con la ciencia de datos y la modelación estadística.
Licenciado en Ciencias y Técnicas Estadísticas y Especialista en
Métodos Estadísticos por la Universidad Veracruzana, con Maestría
en Ciencias de la Información Geoespacial por el Centro Geo, CDMX
(Centro CONAHCYT), actualmente cursando el Doctorado en Ciencias
de la Computación en la Universidad Veracruzana. Profesor en la
Licenciatura en Estadística, en la Especialidad en Métodos Estadísticos
y la Maestría en Economía y Sociedad de China y América Latina en
la misma universidad, donde he dirigido 15 tesis de licenciatura y 4 de
especialidad. También tengo experiencia como Analista Estadístico
en la Oficina del Programa de Gobierno del Estado de Veracruz.
Mis líneas de investigación incluyen metodologías de cómputo,
programación estadística, estadística multivariada, análisis espacial,
ciencia de datos y modelos estadísticos, con aplicaciones en biología,
medicina, ciencias administrativas y sociales. He participado en
diversos congresos nacionales e internacionales.
Benjamín Mendoza
Emmanuel Morales
B. Mendoza, E. Morales, C. Cruz, and L. Gomez,
“Morphological classification of hematophagous Diptera with Convolutional Neural Networks: A mapping of literature,”
Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.
ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026
53
LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026
AUTHORS
Profesora de tiempo completo en la Facultad de Estadística e
Informática de la Universidad Veracruzana (UV). Es Doctora en
Investigación Educativa por la UV, Maestra en Ciencias con especialidad
en Estadística Aplicada por el ITESM Campus Monterrey, así como
Especialista en Métodos Estadísticos y Licenciada en Estadística
por la UV. Actualmente coordina la Especialización en Métodos
Estadísticos y pertenece al Sistema Nacional de Investigadores (SNII)
como Candidata.
Sus líneas de investigación abarcan la Educación Estadística, la
Metodología Estadística Aplicada y la integración de la estadística con
tecnologías emergentes como Machine Learning, Ciencia de Datos y
Análisis Espacial. Ha colaborado en proyectos sobre sustentabilidad,
alfabetización digital y actitudes hacia la estadística en estudiantes
latinoamericanos.
Entre sus publicaciones recientes destacan trabajos sobre aprendizaje
supervisado, formación en consultoría estadística y aplicaciones
multivariantes, además de capítulos sobre alfabetización digital y
redes sociales. Ha dirigido más de 70 tesis, combinando docencia,
investigación e impulso al uso estratégico de la estadística para el
desarrollo sostenible.
Cecilia Cruz
B. Mendoza, E. Morales, C. Cruz, and L. Gomez,
“Morphological classification of hematophagous Diptera with Convolutional Neural Networks: A mapping of literature,”
Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.
Es doctor en Administración y doctor en Finanzas Públicas. Estudió
la Maestría en Impuestos, la Maestría en Contabilidad, así como la
Especialidad en Administración. Es licenciado en Contaduría Pública
Certificado por el IMCP y Licenciado en Psicología.
En su experiencia docente se desempeña como coordinador
del Posgrado en Administración modalidad Virtual, asimismo es
integrante del Cuerpo Académico Consolidado “Las organizaciones
y su entorno”. Investigador de Tiempo Completo del Instituto de
Investigaciones y Estudios Superiores de las Ciencias Administrativas.
El Dr. Luis Enrique es miembro del sistema nacional de investigadores
e investigadoras, miembro certificado por Conocer, perfil deseable
PRODEP y nivel 6 de productividad UV.
Además de ser autor y coautor de varios libros y revistas.
Luis Gomez