Intelligent classification of medically and veterinary important mosquito vectors using self-distillation-enhanced supervised learning

Mosquito-borne diseases remain a persistent global public health burden, with effective control strategies depending on the accurate and timely identification of mosquito species and estimation of their population density. The current gold standard, morphological identification via microscopic examination, is limited by inter-observer variability and a scarcity of local taxonomic expertise. While highly accurate, molecular techniques like polymerase chain reaction and DNA barcoding are often impractical for widespread field use due to their cost, procedural complexity, and reliance on specialized equipment and technicians. In response to these challenges, this study harnesses the potential of artificial intelligence (AI) to develop a robust, accessible diagnostic tool. We introduce an innovative application of the Distillation of knowledge with NO labels v2 (DINOv2-l14) algorithm, a self-supervised learning (SSL) model, for the classification of 30 distinct mosquito species from image data. A key focus was enabling the model to learn effectively from a highly imbalanced dataset, where the number of images per class ranged from 15 to 124. We also investigated the data efficiency of SSL by determining the minimum data partition required to match the performance of a conventional supervised model. Our image processing pipeline involved cropping, padding, and dividing images into 14 × 14 patches before converting them into embedded vectors for classification. The fully trained DINOv2-l14 model significantly outperformed a selected ResNet101 baseline, achieving 99.2% accuracy, 97.7% specificity, an F1 score of 86.1%, and a robust area under the curve (AUC) value of 0.996. Most notably, when fine-tuned on just 20% of the available data (averaging 11-13 images per class), the SSL model achieved performance comparable to the baseline model trained on the entire dataset, demonstrating exceptional data efficiency. Although limitations exist, including dataset imbalance and a lack of standardization for broader application, the model's high performance across five-fold cross-validation and strong Cohen's Kappa coefficients (up to 0.836) confirm its reliability. This work validates a novel, efficient, and highly accurate AI-driven approach to mosquito identification. By integrating such a model into a mobile application using an embedding AI framework, this technology holds significant promise for empowering field personnel with a practical, powerful tool to enhance vector surveillance and control programs worldwide.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-16
DOI
https://doi.org/10.1371/journal.pone.0362945
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Intelligent classification of medically and veterinary important mosquito vectors using self-distillation-enhanced supervised learning

Santhad Chuwongin, Sedthapong Laojun, Tawee Saiwichai, Veerayuth Kittichai et al.
4 links
PLoS ONE
Advanced Neural Network Applications
article

Intelligent classification of medically and veterinary important mosquito vectors using self-distillation-enhanced supervised learning

Santhad Chuwongin, Sedthapong Laojun, Tawee Saiwichai, Veerayuth Kittichai, Morakot Kaewthamasorn, Tanawat Chaiphongpachara, Naruchit Soiphet, Siridech Boonsang
article en

Abstract

Mosquito-borne diseases remain a persistent global public health burden, with effective control strategies depending on the accurate and timely identification of mosquito species and estimation of their population density. The current gold standard, morphological identification via microscopic examination, is limited by inter-observer variability and a scarcity of local taxonomic expertise. While highly accurate, molecular techniques like polymerase chain reaction and DNA barcoding are often impractical for widespread field use due to their cost, procedural complexity, and reliance on specialized equipment and technicians. In response to these challenges, this study harnesses the potential of artificial intelligence (AI) to develop a robust, accessible diagnostic tool. We introduce an innovative application of the Distillation of knowledge with NO labels v2 (DINOv2-l14) algorithm, a self-supervised learning (SSL) model, for the classification of 30 distinct mosquito species from image data. A key focus was enabling the model to learn effectively from a highly imbalanced dataset, where the number of images per class ranged from 15 to 124. We also investigated the data efficiency of SSL by determining the minimum data partition required to match the performance of a conventional supervised model. Our image processing pipeline involved cropping, padding, and dividing images into 14 × 14 patches before converting them into embedded vectors for classification. The fully trained DINOv2-l14 model significantly outperformed a selected ResNet101 baseline, achieving 99.2% accuracy, 97.7% specificity, an F1 score of 86.1%, and a robust area under the curve (AUC) value of 0.996. Most notably, when fine-tuned on just 20% of the available data (averaging 11-13 images per class), the SSL model achieved performance comparable to the baseline model trained on the entire dataset, demonstrating exceptional data efficiency. Although limitations exist, including dataset imbalance and a lack of standardization for broader application, the model's high performance across five-fold cross-validation and strong Cohen's Kappa coefficients (up to 0.836) confirm its reliability. This work validates a novel, efficient, and highly accurate AI-driven approach to mosquito identification. By integrating such a model into a mobile application using an embedding AI framework, this technology holds significant promise for empowering field personnel with a practical, powerful tool to enhance vector surveillance and control programs worldwide.

PLoS ONEVol. 21(11)
Chulalongkorn University (TH), Mahidol University (TH), Suan Sunandha Rajabhat University (TH), King Mongkut's Institute of Technology Ladkrabang (TH)
King Mongkut's Institute of Technology Ladkrabang
Good health and well-being
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon