Hands-On Science: Train a Leaf or Rock Classifier Outdoors (Ages 10–16)

Why this outdoor ML project works (and what kids actually learn)

If your child likes collecting “cool rocks” or noticing different leaf shapes, you already have the beginning of a great ai science project outdoors. The twist is that your child becomes the scientist and the AI trainer.

In this activity, kids (ages 10–16) take photos outside, then use a simple tool to train an image classification project for students—a model that can tell “Maple leaf vs. Oak leaf” or “Smooth stone vs. Speckled stone.” It’s a hands-on way to learn how machine learning works without needing advanced math or coding.

This type of machine learning nature project for kids teaches three big ideas:

• Observation skills (real science): noticing patterns like edges, color, texture, veins, speckles.
• Data literacy: learning that AI depends on good data (and that messy data creates messy results).
• Responsible experimentation: testing, improving, and being honest about limitations.

Parents love it because it’s:

• Screen time with a purpose
• Outdoor time that feels like a quest
• A collaborative project you can do together in a weekend

What you need + smart setup (10 minutes of prep)

You can do this as a teachable machine image classifier activity using a phone, tablet, or laptop.

Supplies

• A phone or tablet camera (or a Chromebook/laptop with a camera)
• Internet access for training (you can collect photos offline first)
• A small bag or tray (optional, for rocks)
• Paper towels or a cloth (optional, to wipe rocks dry)
• A notebook (optional, for “field notes”)

Pick a classification challenge

Choose 2–4 classes (categories). Fewer categories is easier and usually gives better results.

Good starter options:

• Leaves: Maple vs. Oak vs. Birch
• Rocks: Smooth river rock vs. Rough rock vs. Speckled granite
• Nature textures: Bark vs. Grass vs. Soil

Safety and nature-care reminders

• Don’t pick protected plants; take photos instead.
• Avoid poisonous plants and unknown fungi.
• Wash hands after handling rocks/soil.
• Return any rocks/leaves you moved to where you found them.

Collecting great training photos: the “field scientist” method

Most “AI frustration” comes from the data, not the model. The goal is to give the classifier enough variety that it learns what makes a leaf (or rock) unique—without accidentally learning the background.

Aim for 50–120 images per class. More is better, but quality and variety matter most.

Here’s a practical guide you can follow together:

Step	What to do outdoors	Target	Why it helps the model
1. Choose classes	Pick 2–4 categories	2–4 classes	Fewer classes = clearer patterns
2. Gather variety	Photograph different specimens of each class	10–20 specimens/class	Prevents “memorizing one leaf”
3. Change angles	Top-down, side, slightly tilted	3–5 angles/specimen	Builds robustness
4. Change lighting	Sun, shade, cloudy	2–3 lighting types	Reduces lighting bias
5. Vary background (carefully)	Grass, pavement, hand, plain paper	2–3 backgrounds	Ensures it learns the object, not the scene
6. Keep it fair	Similar number of photos per class	Balanced counts	Prevents one class from dominating

Pro tips that make a big difference

• Fill the frame with the leaf/rock so the model focuses on the right thing.
• Avoid “signature backgrounds.” If all maple leaves are photographed on your red jacket, the model may learn “red jacket = maple.” Mix it up.
• Include “hard examples.” Torn leaves, wet rocks, partial views—these improve real-world performance.

Mini science lesson to share with your child

Explain it like this: “The model doesn’t ‘know’ what a leaf is. It only sees patterns of pixels. Our job is to show it enough examples that the patterns become reliable.”

Train your classifier (simple tools, real results)

You have two easy paths:

• Option A (fastest): Google’s Teachable Machine (great for beginners)
• Option B (slightly more advanced): a kid-friendly coding platform where you can connect the model to a project

For this post, we’ll focus on the classic family-friendly approach: Teachable Machine.

Training steps (15–30 minutes)

• Create a new Image Project.
• Make your classes (e.g., “Oak,” “Maple,” “Birch”).
• Add images for each class (upload photos you took, or use the webcam to capture them).
• Click Train Model.
• Test it live using the camera.

How to run a fair test (like a real experiment)

After training, don’t just test on the same photos you trained with. Try this:

• Pick 10 new leaves/rocks the model hasn’t seen.
• Test each item twice:
  • once in good lighting
  • once in tricky lighting (shade, partial view, background clutter)

Have your child record results like a scientist:

• Correct / incorrect
• What was different about the incorrect ones?
• Any patterns (e.g., “It struggles when the leaf is curled”)

Common failure modes (and what to do about them)

• It guesses “Maple” for everything: You may have way more maple photos than others. Balance the classes.
• It works at home but not outside: Your training photos had a different background/lighting. Add more variety.
• It confuses two categories: Your categories may be too similar (e.g., two very similar rocks). Either:
  • choose more distinct classes, or
  • add “feature-focused” photos (close-ups of veins, edges, speckles).

Upgrade challenge (for older kids 13–16)

Ask them to define a hypothesis:

• “If we add more shade photos, the model will improve in shade.”

Then run a simple A/B test:

• Model 1: trained mostly in sun
• Model 2: trained with equal sun and shade photos

Compare which model performs better in shade.

Make it a STEM weekend project for families (plus fun variations)

This can be a one-afternoon activity, but it’s also a fantastic stem weekend project for families when you add a little structure.

A simple weekend plan

• Day 1 (30–60 min outdoors): collect images + notes
• Day 1 (30 min indoors): train a first model, test it
• Day 2 (20–40 min): improve the dataset, retrain, compare results

Family roles (so everyone can participate)

• Data Collector: takes photos and labels them
• Quality Checker: makes sure photos are clear and balanced
• Tester: finds “new” leaves/rocks to challenge the model
• Science Reporter: writes the results and what changed

Fun variations to keep kids engaged

• “Mystery class” test: Add a “Not sure / Other” class with random backgrounds or non-leaf objects. This teaches the limits of classification.
• Seasonal redo: Train the same leaf model in summer vs. fall colors. Does it generalize?
• Neighborhood comparison: Try leaves from two parks—does the model still work?
• Ethical AI chat: Discuss what happens when AI is trained on biased data (easy example: only sunny-day photos).

This is also a great doorway into deeper AI learning: data collection, labeling, evaluating accuracy, and iterating—exactly what real ML teams do.

Next Steps: How to get started today (and keep going)

Use this checklist to go from “idea” to “working classifier” in one day:

• Pick your categories: 2–4 leaf types or rock types that are easy to tell apart.
• Collect balanced photos: 50–120 images per class with varied angles, lighting, and backgrounds.
• Train a first model: Use a teachable machine image classifier activity to get quick results.
• Test like a scientist: Try new samples and record where it fails.
• Improve the dataset: Add photos specifically targeting failures (shade, wet rocks, curled leaves, cluttered backgrounds).
• Retrain and compare: Keep notes on what changed and whether accuracy improved.

If your child enjoys this, the best next move is to turn the classifier into something interactive (a mini “nature scanner” game or a quiz). At Intellect Council, we love helping kids take projects like this beyond the first experiment—so they learn not just that AI works, but why it works and how to make it better.