Local Lizard Species Boundaries with Image AI Analysis

ISEF Category: Animal Sciences

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

The Hook

Two lizards can look like twins and still belong to different species. A laptop and a set of iNaturalist photos can help you spot those hidden splits. That matters because species counts drive conservation, biodiversity surveys, and how scientists describe a place.

What Is It?

Species boundaries are the lines scientists draw when one population stops looking like part of another species. In lizards, those lines can blur because color, body shape, and pattern can shift across age, sex, and location. A contrastive-learning model helps by turning each photo into a point in a map, then pulling similar images closer together and pushing different ones apart.

Think of it like sorting a box of mixed school photos. If the model separates the lizards into neat groups, you can ask whether those groups match known species labels, geography, or hidden traits. That gives you a way to test whether the local population contains more than one species, or just a lot of variation inside one species.

Why This Is a Good Topic

This project is a strong science fair choice because you can test a real taxonomic question with public data and clear metrics. It connects to biodiversity, conservation, and how scientists sort similar-looking animals into species. You can learn image cleaning, feature extraction, clustering, and basic validation without needing a wet lab.

Research Questions

• How does photo angle affect cluster separation in a contrastive-learning embedding model?
• What is the effect of filtering for verified iNaturalist observations on species-label agreement?
• Does a model trained on regional lizard photos separate local forms better than a model trained on broad global images?
• To what extent do embedding clusters match known species identifications in the local population?
• Which visual traits, such as dewlap color, body pattern, or tail length, explain the largest cluster differences?
• How does geographic distance between observation sites relate to embedding similarity among the lizards?

Basic Materials

• Laptop with at least 8 GB of RAM.
• Free iNaturalist account.
• Spreadsheet software, such as Google Sheets or LibreOffice Calc.
• Python installed through Anaconda or Miniconda.
• Jupyter Notebook or Google Colab access.
• Shared folder or external drive for image files and CSV exports.

Advanced Materials

• GPU workstation or cloud GPU access.
• High-resolution image archive from iNaturalist, museum collections, or field cameras.
• Python packages for deep learning, clustering, and plotting.
• Annotation tool such as Label Studio or CVAT.
• Standardized specimen imaging setup with fixed lighting and a scale bar.
• Digital calipers for optional morphometric measurements.

Software & Tools

• Python: Cleans observation data, trains models, and runs clustering analyses.
• Jupyter Notebook: Keeps code, notes, and plots in one place.
• PyTorch: Builds the contrastive-learning model and embedding pipeline.
• scikit-learn: Calculates cluster scores, principal component plots, and simple classifiers.
• Google Colab: Provides a free notebook environment with optional GPU access.

Experiment Steps

1. Define the exact lizard population, label set, and photo inclusion rules.
2. Filter out duplicate records, blurry images, and observations with weak identification support.
3. Build your embedding pipeline and pick a simple baseline, such as color or shape features.
4. Choose the metrics that will judge separation, label agreement, and geographic consistency.
5. Add controls for age, sex, background, and site bias before you run the full analysis.

Common Pitfalls

• Mixing juveniles, adult males, and adult females in one group, which can make body size and color look like species differences.
• Feeding the model photos with different camera angles and backgrounds, which can split clusters by pose instead of biology.
• Trusting unverified iNaturalist identifications, which can teach the model the wrong labels.
• Letting one island, park, or neighborhood dominate the sample, which can make place, not species, drive the result.
• Treating a clean cluster plot as proof of a species boundary, which skips the validation step that systematics needs.

What Makes This Competitive

A strong version compares the embedding model against simple baselines, not just a single cluster plot. It also checks whether the result holds after you control for age, sex, image angle, and geography. If you validate clusters against expert IDs, range overlap, or a second image set, you move from a nice demo to a real species-boundary test. That kind of careful validation is what makes systematics work convincing.

Project Variations

• Compare mainland and island populations to see whether geography predicts the embedding clusters.
• Test whether male, female, and juvenile photos separate differently, which can reveal sexual dimorphism effects.
• Swap in museum specimen images and ask whether preserved material produces clearer species groupings than field photos.

Learn More

• iNaturalist Help Center: Learn how observation quality, computer vision, and data exports work on the iNaturalist website.
• GBIF: Download occurrence records and map species ranges from the Global Biodiversity Information Facility website.
• PubMed: Search review articles on species delimitation, lizard evolution, and contrastive learning in biology.
• NCBI Bookshelf: Read free chapters on phylogenetics and species concepts in the NCBI Bookshelf library.
• OpenStax Biology 2e: Review species concepts, variation, and natural selection in a free online textbook.