PlantNet Misclassification Patterns in Plant Families

ISEF Category: Plant Sciences

Subcategory: Systematics and Evolution  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A plant ID app can be right on one species and wrong on its close cousin. That mismatch is a clue, not a glitch. You can turn those mistakes into a real research project. Your job is to find out which plant families trip the model and which traits confuse it most.

What Is It?

This project asks how well PlantNet identifies plants from photos, then checks whether some plant families get mislabeled more often than others. Think of the app like a student who studies hundreds of flashcards, then still mixes up look-alike cousins. Your goal is to measure those mix-ups and look for a pattern.

You are not just asking, "Did it get this one right?" You are asking why it got it wrong. Maybe leaf shape, vein pattern, flower color, or growth habit makes two groups look too similar in photos. In systematics, those visible traits matter because they help people sort species into related groups. Your project tests whether the model fails in the same places a human might struggle, or in different places.

Why This Is a Good Topic

This makes a strong science fair topic because you can test it with public images, a clear scoring system, and real statistics. You do not need a wet lab. You need a plan for collecting images, organizing taxa, and comparing error rates across plant families. The project connects to biodiversity monitoring, invasive species detection, and citizen science, so the real-world value is easy to explain.

Research Questions

• How does PlantNet accuracy vary across plant families with similar leaf shapes??
• What is the effect of flower presence on PlantNet species-identification accuracy??
• Does image angle change the likelihood of misclassification within a plant family??
• To what extent do compound leaves increase misclassification compared with simple leaves??
• Which morphological traits, such as leaf margin, venation, or inflorescence type, best predict PlantNet errors??
• How does PlantNet perform on rare or underrepresented species compared with common species??

Basic Materials

• Smartphone or digital camera with a consistent photo setting.
• Computer with internet access.
• Spreadsheet software such as Google Sheets or Excel.
• PlantNet app or web interface.
• Public plant image sources, such as iNaturalist or USDA PLANTS, for comparison samples.
• Notebook for recording family, species, photo angle, and app output.
• Reference field guide or taxonomy database for checking names.

Advanced Materials

• High-resolution camera or DSLR for standardized images.
• Computer with spreadsheet software and basic statistics support.
• R or Python for accuracy analysis and plots.
• ImageJ for measuring visible traits such as leaf area, aspect ratio, and color contrast.
• Curated image set from herbarium databases or open biodiversity repositories.
• Taxonomy reference sources such as USDA PLANTS, Kew POWO, or GBIF for name verification.
• Optional GIS data if you compare geographic origin with misclassification patterns.

Software & Tools

• Google Sheets: Organizes predictions, true labels, and trait scores in a clean table.
• R: Runs accuracy tests, confusion matrices, and family-level comparisons.
• Python: Helps automate image sorting, data cleaning, and plotting.
• ImageJ: Measures visible shape features from plant photos.
• iNaturalist: Provides open observations and expert-vetted examples for comparison.

Experiment Steps

1. Define the plant groups you will compare, and decide whether you will test families, genera, or a smaller set of look-alike species.
2. Build a labeled image dataset with enough examples per group to make fair comparisons.
3. Decide which traits you will score by hand or with image analysis, such as leaf shape, margin, venation, or flower presence.
4. Set up a prediction table that records the top PlantNet match, the confidence score, and the true identity for every image.
5. Plan an error-analysis scheme that groups mistakes by family, trait, and image type.
6. Choose the statistics and graphs that will let you compare misclassification rates across groups.

Common Pitfalls

• Mixing photos from different sources with different backgrounds, which lets lighting and clutter affect the model more than plant traits.
• Using too few images in one family, which makes one odd sample look like a real trend.
• Comparing species names without checking the accepted taxonomy, which can turn naming changes into fake errors.
• Scoring traits too loosely, which makes leaf shape or venation categories too subjective to analyze.
• Ignoring photo angle, which can hide whether the model fails because the plant was partially visible rather than because the family is hard to identify.

What Makes This Competitive

A stronger project does more than report accuracy. It finds a pattern that explains the mistakes. You can compare multiple plant families, control for image quality, and test whether a specific trait set predicts failure better than family name alone. If you add a careful confusion matrix, a statistical test, and a trait-based error model, your project starts to look like real computational biology.

Project Variations

• Compare PlantNet accuracy on native versus invasive species from your region.
• Test whether flower photos or leaf photos produce fewer family-level errors.
• Analyze whether herbarium images and field photos lead to different misclassification patterns.

Learn More

• PlantNet: Search the PlantNet project pages and help section for how the model identifies plants and what data it uses.
• USDA PLANTS Database: Look up accepted plant names, family groups, and distribution data.
• GBIF: Search the Global Biodiversity Information Facility for specimen records and open occurrence data.
• iNaturalist: Use open observations to compare field photos and community identifications.
• Systematic Botany: Search the journal for studies on plant identification, morphology, and taxonomic traits.
• MIT OpenCourseWare: Search for ecology, evolution, or data analysis course materials that explain classification and biodiversity methods.