Multimodal Garden Health Prediction

ISEF Category: Plant Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

One garden bed can hide three stories at once, yield, disease, and nutrient stress. A phone photo, a soil reading, and a weather log can each catch part of the story, but not all of it. When you combine them, you may predict plant health better than any single clue alone. That makes this project feel like farm tech, not just a class experiment.

What Is It?

This project asks a simple question with a modern twist. Can you predict how a school-garden plot will perform by combining three kinds of data, plant color from smartphone RGB images, soil electrical conductivity (EC), and weather conditions? RGB means red, green, and blue color values from a photo. Soil EC is a measure of how well soil carries electrical current, which often relates to dissolved salts and nutrient availability.

Think of it like reading a student report card. One grade tells you something, but the full report tells you much more. A photo may hint at chlorosis, which is leaf yellowing from stress. Soil EC may hint at nutrient problems. Weather can explain why plants slow down or get sick. A multimodal model tries to put all those clues together into one prediction.

You would compare that combined model against a baseline that uses only one data source, such as photos alone. Then you test whether the combined approach predicts outcomes like yield, visible disease, or nutrient stress more accurately. The point is not to make a perfect farm forecast. The point is to see whether mixed data gives you a smarter, more useful estimate than one type of measurement alone.

Why This Is a Good Topic

This is a strong science fair topic because you can measure real changes over time, compare a baseline against a combined model, and use data you can collect from a school garden. It connects to real problems in agriculture, like spotting disease early, estimating harvest size, and noticing nutrient stress before plants crash. You can learn image analysis, basic sensor data collection, and model evaluation without needing a wet lab full of expensive tools.

Research Questions

• How does combining smartphone RGB, soil EC, and weather data change yield prediction accuracy compared with photos alone?
• What is the effect of adding soil EC to a photo-based model for detecting visible nutrient stress?
• Does a multimodal model predict disease symptoms earlier than a single-image baseline?
• To what extent do weekly weather variables improve predictions of plant yield across garden plots?
• Which data source, RGB images, soil EC, or weather, contributes most to prediction accuracy for each plant health outcome?
• How does model performance change when you train on one crop species and test on another crop in the same garden?
• To what extent does image quality, such as shade or blur, affect disease and stress predictions?

Basic Materials

• Smartphone with a good camera.
• Digital kitchen or pocket scale for harvest mass.
• Soil EC meter or soil salinity meter.
• Notebook or spreadsheet for weekly field notes.
• Measuring tape or ruler for plant height and canopy size.
• Weather data from a local station or NOAA reports.
• Plant labels or garden map for tracking each plot.
• Consistent photo backdrop or simple white board for image standardization.

Advanced Materials

• Smartphone with manual camera controls or a tripod-mounted phone.
• Portable soil EC probe with repeatable calibration.
• Reference color card for image normalization.
• Laptop for model training and evaluation.
• Annotated plot map with repeated plant health scores.
• Optional handheld chlorophyll meter for comparison data.
• Access to a weather API or station archive.
• Standardized tray or scale setup for biomass or yield measurements.

Software & Tools

• Python: Organizes your data, trains baseline and multimodal models, and runs performance checks.
• Jupyter Notebook: Lets you clean data, test features, and document each modeling step.
• ImageJ: Measures color features from plant images and helps you compare plots over time.
• Pandas: Stores weekly measurements in tidy tables for analysis.
• scikit-learn: Builds baseline models and compares them with combined-data models.

Experiment Steps

1. Define one outcome first, such as yield, visible disease, or nutrient stress, so your model has a clear target.
2. Set up a weekly sampling plan that records each plot with the same photo angle, soil reading, and weather source.
3. Decide how you will label plant health, using simple categories or numeric scores that you can defend.
4. Build a single-modality baseline before you add extra data, so you can prove whether the extra inputs help.
5. Choose a way to combine image features, soil EC, and weather into one model, then compare that model against the baseline.
6. Plan validation carefully, so you test on held-out weeks or plots instead of grading your model on data it has already seen.

Common Pitfalls

• Taking photos in different lighting, which changes RGB values and confuses the model.
• Mixing plots with different plant varieties, which makes the model learn species differences instead of health differences.
• Using soil EC as a direct nutrient score without checking moisture, which can blur the meaning of the measurement.
• Labeling disease and nutrient stress with vague notes, which creates noisy training data.
• Testing the model on the same weeks used for training, which makes accuracy look better than it really is.

What Makes This Competitive

A class-level version of this project might stop at a simple comparison. A stronger version tests several model types, uses held-out validation by week or plot, and reports which data source adds the most predictive power. You can make it more competitive by separating yield, disease, and nutrient stress into different targets instead of treating them as one health score. A careful error analysis, especially for false alarms and missed stress events, can make the project feel much more like real research.

Project Variations

• Focus only on one crop, such as tomatoes, lettuce, or beans, to reduce variation and sharpen your model.
• Swap soil EC for leaf color indices extracted from images, then compare whether the model still improves over photos alone.
• Add disease severity scoring as the main target, then test whether weather history improves early warning more than soil data does.

Learn More

• USDA National Agricultural Library: Search for review articles on crop stress detection, remote sensing, and precision agriculture.
• NOAA Climate Data Online: Find local weather history and station data for your school garden site.
• NASA Earthdata: Explore free background resources on plant monitoring, vegetation indices, and environmental data.
• PubMed: Search for review articles on smartphone plant phenotyping, disease detection, and multimodal crop models.
• MIT OpenCourseWare: Look for free courses in machine learning, data analysis, and environmental systems that can help with model design.