Predict Seed Longevity With Machine Learning

ISEF Category: Plant Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Some seeds stay alive for decades, while others fail fast. That difference matters for food security, conservation, and seed banks. You can test whether visible seed traits help predict that long shelf life better than family tree data alone.

What Is It?

This project asks a simple question with a smart twist: can you predict how long a seed stays viable by looking at its traits? Seed-bank longevity means how long a seed can still germinate after storage. Think of it like a battery life estimate for seeds. Some seeds have thick coats, small size, or other features that may help them last longer.

You start with data from KEW's seed information database and build two prediction models. One model uses only phylogeny, which means how closely related the plant is to other plants. The other model adds seed features such as mass and coat thickness. Then you compare their performance. If the feature-based model does better, that means seed shape and structure may add real predictive power beyond family relationships alone.

Why This Is a Good Topic

This is a strong science fair topic because it asks a clear yes-or-no question, but still leaves room for real analysis. You can work with existing data, so you do not need a greenhouse or a long germination study. The project connects to seed banking, crop preservation, and biodiversity conservation. You also learn how to clean data, compare baselines, and test whether a model really improves prediction.

Research Questions

• How does adding seed mass to a phylogeny-only model change prediction accuracy for seed-bank longevity? ?
• How does adding seed coat thickness to a phylogeny-only model change prediction accuracy for seed-bank longevity? ?
• What is the effect of using both seed mass and coat thickness together instead of phylogeny alone? ?
• Does model performance differ between small-seeded and large-seeded species? ?
• To what extent do closely related species share similar longevity once seed traits are included? ?
• Which feature set, phylogeny only, morphology only, or both, gives the best cross-validated longevity predictions? ?

Basic Materials

• Laptop or desktop computer with internet access.
• Spreadsheet software for cleaning and organizing data.
• Python with Jupyter Notebook.
• Free machine learning library such as scikit-learn.
• Access to KEW seed information database records.
• Spreadsheet or note app for tracking species names, traits, and missing values.

Advanced Materials

• University or public library access to botanical trait databases.
• Python with pandas, scikit-learn, and statsmodels.
• R with phylogenetic packages such as ape or phylolm.
• ImageJ for measuring seed coat thickness from microscope images.
• Digital microscope images or published trait images with metadata.
• External storage for large cleaned datasets and versioned analysis files.

Software & Tools

• Python: Cleans data, builds models, and compares prediction accuracy.
• Jupyter Notebook: Lets you document code, graphs, and results in one place.
• scikit-learn: Trains and tests regression or classification models for longevity prediction.
• RStudio: Supports phylogenetic analysis and baseline models if you use R.
• ImageJ: Measures seed dimensions or coat thickness from images when trait data need to be extracted.

Experiment Steps

1. Define the exact prediction target, such as survival years or a longevity class, and decide which species you will include.
2. Gather seed longevity records and trait data, then check how many entries have missing or inconsistent values.
3. Build a baseline model that uses phylogenetic information only, so you have a fair comparison point.
4. Add seed morphology features, then test whether the model improves under the same validation scheme.
5. Plan controls for data leakage, class imbalance, and species overlap between training and test sets.
6. Compare model results with simple statistics, then decide whether the added traits improve prediction in a meaningful way.

Common Pitfalls

• Mixing species names across databases, which creates duplicate records or mismatched traits.
• Letting closely related species appear in both training and test sets, which can make accuracy look better than it really is.
• Using longevity records with very different storage conditions, which adds noise that the model cannot explain.
• Treating missing trait values as zeros, which can distort seed mass or coat thickness patterns.
• Comparing models with different validation rules, which makes the phylogeny baseline and feature model unfairly matched.

What Makes This Competitive

A stronger project would go beyond a simple accuracy comparison. You could test whether morphology helps most in certain plant families, or whether the improvement holds after strict phylogenetic cross-validation. You could also compare multiple model types and report confidence intervals, not just one score. That kind of careful analysis shows you understand both the biology and the statistics.

Project Variations

• Use seed germination percentage instead of longevity as the prediction target.
• Compare morphology-based predictions with climate-origin variables, such as native habitat dryness or temperature.
• Test whether image-derived traits, such as seed shape and surface texture, improve predictions beyond mass and phylogeny.

Learn More

• KEW Seed Information Database: Search the database for seed longevity records and trait entries by species.
• USDA National Plant Germplasm System: Find seed and plant trait references that can help you compare storage behavior.
• PubMed: Search review articles on seed longevity, desiccation tolerance, and seed storage biology.
• Seed Science Research: Read peer-reviewed studies on seed aging, storage, and germination traits through journal access or abstracts.
• MIT OpenCourseWare: Look for free materials on machine learning, data analysis, and model evaluation.
• ImageJ documentation: Learn how to measure seed dimensions or thickness from images.