Aedes aegypti and Aedes albopictus mosquito classifier
Data files
Sep 21, 2026 version files 41.61 MB
-
mosquito-classifier-data.zip
41.60 MB
-
README.md
5.23 KB
Abstract
Mosquito-borne diseases transmitted by Aedes aegypti (Linnaeus, 1762) and Aedes albopictus (Skuse, 1894) — including dengue, Zika, chikungunya, and yellow fever — depend critically on rapid, accurate vector identification.
Deep learning achieves high accuracy on curated laboratory images, but performance degrades on community-submitted photographs that vary widely in quality, framing, and background.
We developed a multi-stage pipeline to distinguish these two morphologically similar vectors from real-world citizen-science images. Using 2112 mosquito images from the Global Mosquito Observation Database (GMOD), we combined: (i) a binary classifier screening for mosquito presence; (ii) a YOLO-based object detector; (iii) an image-quality assessment module evaluating brightness, sharpness, contrast, and bounding-box ratio; (iv) Segment Anything Model (SAM) segmentation to isolate specimens from background clutter; and (v) a YOLO classifier.
To target the diagnostic characters used in conventional morphological taxonomy, we refined detection to focus on the thoracic scutum — the region bearing the lyre-shaped pale-scale pattern of Ae. aegypti and the median white stripe of Ae. albopictus.
Baseline classification on raw images was poor (30.95% for Ae. aegypti; 78.4% for Ae. albopictus), and augmentation alone gave modest gains.
Whole-body SAM-mask classification improved overall accuracy to 68.21%, and scutum-focused classification achieved 87.5% and 83.3% accuracy for Ae. albopictus and Ae. aegypti respectively (overall 86.2%, 69/80).
Aligning machine attention with the morphological characters used by entomologists delivers meaningful gains, supporting scalable citizen-science vector surveillance.
Overview
This dataset supports Kruthiventi et al., which tests whether focusing a classifier on the mosquito thoracic scutum improves species identification from noisy, community-submitted photographs. The study analyzed 2,112 citizen-science images accessed through the Global Mosquito Observation Database (GMOD, https://www.mosquitodashboard.org). Images were split into training, validation, and test sets at 60:10:30; Ae. albopictus outnumbers Ae. aegypti roughly 3:1. Scutum-focused classification reached 86.2% overall accuracy on the held-out test set (87.5% for Ae. albopictus, 83.3% for Ae. aegypti), compared with 68.21% for whole-body classification and 30.95-78.4% for a baseline trained on raw images.
Associated publication: Kruthiventi et al. (in review). A scutum-focused deep learning pipeline for species-level identification of Aedes aegypti and Aedes albopictus from citizen-science images. Approaches in Entomology: Methods, Philosophies, and Ethics. Manuscript ID: 7743008.
Description of the data and file structure
Compressed folder: mosquito-classifier-data.zip
├── README.md
├── image_metadata.csv # dataset structure documentation (no record IDs)
├── 04_models/ # trained model weights
│ └── best_model.pth
└── 05_code/ # training and evaluation scripts
├── train_custom.py
├── test.py
└── image.png # sample input image for demonstration
image_metadata.csv
Documents the structure of the dataset used in the study, including column definitions for the image-quality metrics (brightness, sharpness, contrast, bounding-box ratio), species labels, and train/val/test splits. The original 2,112 source photographs are not included in this deposit. They were accessed from the Global Mosquito Observation Database (GMOD, https://www.mosquitodashboard.org) in January 2024.
Researchers who wish to apply the deposited pipeline may query GMOD for Aedes aegypti and Aedes albopictus photographs and use the provided code and model weights. For questions about data access or specific record retrieval, contact the corresponding author.
04_models
best_model.pth — trained ResNet18 weights (4-channel input) for the scutum-focused classifier reported in the paper.
05_code
train_custom.py— training script for the 4-channel ResNet18 classifier.test.py— evaluation script.image.png— a single sample image demonstrating the input format.
Code/software
Python, using PyTorch and torchvision for the classifier and Ultralytics v8.0.0 for YOLOv8 detection and Segment Anything Model segmentation. Bounding boxes were annotated in Roboflow. A GPU is used when available; the code also runs on CPU.
The pipeline runs in five steps: score images for quality (brightness, sharpness, contrast, bounding-box ratio) and discard those that fail; detect the mosquito with YOLOv8; segment it with SAM using the detected box as a prompt; composite the crop and mask into the 4-channel arrays; then train and evaluate the classifier with train_custom.py and test.py.
Code and models are also available at: https://github.com/KssNikhil2002/mosquito-classifier
Note on model architecture: The paper reports headline scutum-focused results from a YOLO classifier. The deposited train_custom.py and best_model.pth implement a 4-channel ResNet18. Users should pair the deposited scripts with the deposited weights to reproduce the ResNet18 results.
Access information
License: CC0 1.0 Universal Public Domain Dedication (Dryad default).
Derived from: Global Mosquito Observation Database (GMOD), https://www.mosquitodashboard.org, accessed January 2024.
Source images: The original citizen-science photographs are not redistributed here. They are available directly from GMOD. Users may query the database for Aedes aegypti and Aedes albopictus images and apply the deposited code and weights.
This study analyzed previously collected, de-identified citizen-science images; no human or animal subjects were involved. No ethical approval was required.
Funding
This work was supported by the University of Wisconsin-Madison School of Medicine and Public Health, the University of Wisconsin-Madison Office of the Vice Chancellor for Research, the University of Wisconsin-Madison Department of Population Health Sciences, the University of Wisconsin-Madison Data Science Institute, and the Alfred P. Sloan Foundation through the Open Source Program Office Internship Program.
Authors
Nikhil Kruthiventi, Allie Hannum, Abe Megahed, Sriram Chellappan, Ryan Carney, Finn Kuusisto, Johnny A. Uelmen (corresponding author: uelmen@wisc.edu).