Accurate identification of invasive Aedes mosquito species using low-cost imaging and geometric wing morphometrics

This study demonstrates that low-cost smartphone imaging combined with geometric wing morphometrics can accurately distinguish invasive *Aedes* mosquito species, offering a cost-effective alternative to professional microscopes provided that landmark digitization is performed by a single observer to minimize error.

The Gist

Imagine you are a detective trying to catch four very sneaky criminals: Aedes aegypti, Aedes albopictus, Aedes japonicus, and Aedes koreicus. These aren't just any mosquitoes; they are invasive species that can carry dangerous diseases like Dengue and Zika. The problem? They look almost identical to the naked eye. It's like trying to tell apart four identical twins who are wearing the same clothes.

Usually, to catch them, you need a highly trained expert with a fancy, expensive microscope and a lot of time. But what if you could use a cheap smartphone and a simple math trick instead?

That's exactly what this study set out to do. Here is the story of how they solved the case, explained simply:

The "Wing Fingerprint" Idea

Mosquitoes have wings, and just like human fingerprints, the pattern of veins on a mosquito's wing is unique to its species. The researchers realized that if they could measure the exact shape of these wings, they could tell the species apart without needing a PhD in entomology.

They used a technique called Geometric Morphometrics. Think of this as a digital "connect-the-dots" game. They placed 18 specific dots (landmarks) on the intersections of the wing veins. By measuring the distance and angles between these dots, they created a digital "shape signature" for each mosquito.

The Great Camera Showdown: Pro vs. Budget

The team wanted to know: Do we need a $10,000 professional microscope, or will a $50 smartphone with a clip-on macro lens work just as well?

• The Pro Team: They took photos using a high-end microscope.
• The Budget Team: They took photos using an iPhone SE with a cheap macro lens.

They then fed these photos into four different "digital brains" (computer algorithms) to see which one could best guess the mosquito's species based on the wing shape.

The Results: The Smartphone Wins (Mostly)

Here is the plot twist: The smartphone worked surprisingly well!

• The Accuracy: When a single, experienced person took the photos and marked the dots, the smartphone images were 92% accurate. The microscope images were slightly better at 95%.
• The Verdict: For a fraction of the cost, a smartphone is a powerful tool for catching these invasive mosquitoes. This is huge news for countries or regions that don't have expensive labs.

The "Human Error" Glitch

However, there was a catch. The study also tested what happened when six different people tried to mark the dots on the same photos.

• The Microscope: Even with different people looking at the high-quality microscope photos, they mostly agreed on where the dots went.
• The Smartphone: When different people looked at the smartphone photos, they disagreed more often. The lower image quality made it harder to see exactly where the veins crossed, leading to more "human error."

The Analogy: Imagine trying to draw a dot on a blurry photo of a street sign versus a crystal-clear photo. On the blurry one, one person might put the dot on the left edge of the letter, while another puts it on the right. On the clear one, everyone agrees it's in the middle.

The Solution: One Detective, One Tool

The study concluded that while the smartphone is a fantastic, low-cost tool, it works best if one single, experienced person does all the measuring. If you have a team of six people all marking dots on smartphone photos, the "noise" from their different interpretations might confuse the computer.

Why This Matters

This research is like giving local health workers a "superpower." Instead of waiting for a specimen to be shipped to a high-tech lab in a big city (which takes time and money), they can:

1. Catch a mosquito.
2. Snap a photo with a cheap phone.
3. Get an answer almost immediately.

This helps stop the spread of invasive mosquitoes and the diseases they carry, especially in areas where resources are tight. It turns a high-tech science problem into a low-tech, accessible solution.

In a nutshell: You don't need a Ferrari to win the race; sometimes a reliable bicycle (a smartphone) gets you there just as fast, as long as you have a skilled rider (a single experienced observer) steering the wheel.

Technical Summary

1. Problem Statement

Accurate identification of invasive Aedes mosquito species (specifically Ae. aegypti, Ae. albopictus, Ae. japonicus, and Ae. koreicus) is critical for assessing vector-borne disease risks in Europe. However, traditional morphological identification relies on taxonomic keys that require high levels of expertise and well-preserved specimens, which are often unavailable in field settings. Molecular methods are accurate but require expensive, well-equipped laboratories, limiting their use in resource-constrained environments.

While geometric wing morphometrics (analyzing the shape of wing veins via landmarks) offers a promising alternative, its broader application is hindered by two main factors:

1. Equipment Costs: High-resolution imaging traditionally requires professional stereomicroscopes.
2. Observer Bias: The digitization of landmarks is subjective; variations between different observers can introduce significant error, reducing the reproducibility and transferability of classification models.

The study aimed to determine if low-cost imaging (smartphones with macro lenses) combined with geometric morphometrics could provide accurate species identification comparable to professional microscopes, and to quantify the impact of observer variation on this process.

2. Methodology

Data Collection:

• Specimens: 670 female mosquitoes were analyzed, comprising four species: Ae. aegypti (184), Ae. albopictus (156), Ae. japonicus (166), and Ae. koreicus (164).
• Imaging Devices: Right wings were mounted and photographed using two devices:
  1. Professional: Olympus SZ61 stereomicroscope with an Olympus DP23 camera.
  2. Low-Cost: iPhone SE (3rd Gen) with an Apexel-24XMH macro lens.
• Landmark Digitization:
  • Single-Observer Dataset: An experienced entomologist digitized 18 anatomical landmarks on all 670 images.
  • Multi-Observer Dataset: Six different observers digitized landmarks on a subsample of 20 images per species per device (total 240 images) to assess observer error.
• Software: Landmarks were digitized using Fiji (ImageJ).

Statistical Analysis:

• Preprocessing: Generalized Procrustes Analysis (GPA) was performed using the R package geomorph to superimpose landmarks, removing size, position, and rotation to isolate shape variables.
• Classification: Four machine learning algorithms were compared using an 80:20 train-validation split (repeated 20 times):
  • Linear Discriminant Analysis (LDA)
  • Random Forest (RF)
  • Support Vector Machine (SVM)
  • XGBoost
• Error Analysis: A Procrustes ANOVA was conducted to partition shape variation among species, imaging device, observer, and image ID. Additionally, Euclidean pixel distances were calculated to quantify landmark-wise deviation among observers.

3. Key Contributions

1. Validation of Low-Cost Imaging: The study demonstrates that smartphone images equipped with macro lenses provide sufficient resolution for geometric morphometric analysis, achieving classification accuracies only slightly lower than professional microscopes when processed by a single observer.
2. Quantification of Observer Bias: The research explicitly quantifies how observer variation affects classification accuracy. It found that while single-observer data yields high accuracy, multi-observer data significantly degrades performance, particularly with smartphone images.
3. Algorithm Performance: The study evaluates the robustness of different classifiers against observer-induced noise, finding that machine learning models (RF, SVM, XGBoost) are more robust to observer variation than traditional parametric methods (LDA) in multi-observer scenarios.
4. Landmark-Specific Error Mapping: The authors identified specific landmarks (e.g., Landmark 2 and 13) that are prone to high inter-observer error due to anatomical ambiguity (e.g., merging veins or pale cross-veins), providing actionable insights for future protocol improvements.

4. Results

Classification Accuracy:

• Single Observer:
  • Microscope Images: LDA achieved the highest accuracy (95%).
  • Smartphone Images: LDA achieved 92% accuracy.
  • Note: Misclassification occurred primarily between the closely related Ae. japonicus and Ae. koreicus (80–92% accuracy), while Ae. aegypti and Ae. albopictus were distinguished with >95% accuracy.
• Multi-Observer:
  • Accuracy dropped significantly for all devices.
  • LDA performed the worst in the multi-observer setting, likely due to its sensitivity to non-normal data distributions caused by digitization errors.
  • RF, SVM, and XGBoost performed better than LDA in the multi-observer scenario, showing greater robustness to the "skewness" introduced by different observers.
  • The accuracy reduction was more pronounced for smartphone images than microscope images, confirming higher observer error with low-cost imaging.

Variation Analysis (Procrustes ANOVA):

• Species explained the largest proportion of shape variation (24.9%), confirming distinct morphological differences.
• Observer accounted for 7.7% of the variation, a statistically significant but smaller effect compared to species.
• Image ID (repeated measurements of the same image) accounted for 41.7% of the total variation, though the mean square was low, indicating small individual effects.
• Device had a small but significant effect (0.5%).

Landmark-Specific Findings:

• Landmark 13 (intersection of cubitus 1 and media) showed the highest variability among observers, attributed to the vein being pale and difficult to visualize.
• Landmark 2 (intersection of costa and subcosta) showed the highest variability associated with image ID and observer error, likely due to the gradual merging of veins and dense scale coverage.
• Smartphone images consistently yielded larger Euclidean pixel distances (higher error) than microscope images across all landmarks.

5. Significance and Conclusion

This study establishes that geometric wing morphometrics is a viable, cost-effective tool for monitoring invasive Aedes species in Europe, even when using low-cost smartphone technology.

• Practical Application: Surveillance programs in resource-limited settings can utilize smartphones with macro lenses to capture wing images. However, to maintain high classification accuracy (~92%), landmark digitization should be performed by a single, experienced observer to minimize inter-observer bias.
• Methodological Insight: If multiple observers are necessary, machine learning classifiers like Random Forest or XGBoost are preferable to LDA, as they are more robust to the noise introduced by varying digitization techniques.
• Future Directions: The authors suggest that further workflow optimization (e.g., using immersion oil for direct imaging without mounting) could accelerate the identification process, facilitating rapid integration into public health surveillance systems to prevent the establishment of invasive vector populations.

In summary, the paper bridges the gap between high-precision laboratory techniques and field-deployable solutions, offering a scalable strategy for early detection of disease vectors.