Smartphone Powdery Mildew Detection

ISEF Category: Plant Sciences

Subcategory: Pathology  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

Powdery mildew can spread before your eye catches it. That matters because early treatment works better than late treatment. Your phone camera can become a detector if you train it on the right leaf images. This project turns a garden problem into an image recognition challenge.

What Is It?

Powdery mildew is a fungal disease that leaves a white, dusty coating on leaves. Cucurbits are plants like cucumber, squash, zucchini, and melon. Early on, the spots can look mild or even easy to miss. Later, the leaves yellow, dry out, and lose their ability to feed the plant.

A convolutional neural network, or CNN, is a type of machine learning model that learns patterns in images. Think of it like a very fast pattern finder. Instead of looking for the whole leaf at once, it learns small clues, such as texture, shape, color shifts, and patch edges. You collect your own leaf photos, label them with help from a pathologist mentor, train the model, and then test whether it spots disease earlier or more consistently than a human scorer.

This topic sits at the border of plant disease diagnosis and computer vision. You are not just asking, "Can the model see mildew?" You are asking, "Can it beat or support human judgment on early symptoms, where decisions matter most?"

Why This Is a Good Topic

This is a strong science fair topic because you can measure real performance, not just make a pretty app. You can test accuracy, sensitivity, false alarms, and how well the model handles different leaf ages, lighting, and symptom stages. It connects to crop loss, home gardening, and disease monitoring, so the problem is real. You can also learn data collection, labeling, model training, and statistical comparison, which makes the project feel like true research.

Research Questions

• How does a CNN's accuracy change when it is trained on early-stage versus late-stage powdery mildew images?
• What is the effect of leaf age on the model's ability to detect powdery mildew?
• Does adding images from different lighting conditions improve the CNN's performance on new leaves?
• To what extent does the CNN outperform visual human scoring on early infection cases?
• Which image features, such as lesion texture or white patch coverage, most strongly influence the model's prediction?
• How does training set size affect the model's sensitivity to mild symptoms?
• What is the effect of using multiple cucumber-family species on the model's generalization to unseen leaves?

Basic Materials

• Smartphone camera with manual focus or exposure control.
• Home-grown cucurbit plants, such as cucumber, zucchini, or squash.
• Notebook or spreadsheet for symptom labels and metadata.
• Neutral background board for leaf photography.
• Tripod or phone stand to keep image framing consistent.
• Ruler or printed scale marker for image size reference.
• Access to a pathologist mentor for label review.
• Laptop or desktop computer for data organization and model training.
• Free cloud notebook account or local Python setup.
• External storage or cloud backup for image files.

Advanced Materials

• High-resolution smartphone or digital camera for standardized imaging.
• Uniform light box or controlled photo station.
• Stereo microscope or handheld digital microscope for close symptom checks.
• Leaf punch or sampling tools if mentor-approved verification is needed.
• Image annotation software for label review and quality control.
• Computer with GPU access for faster CNN training.
• Python, TensorFlow, or PyTorch environment for model development.
• Statistical analysis software for confusion matrices, ROC curves, and confidence intervals.
• Reference plant pathology images from a vetted collection for comparison.
• Optional multispectral or UV imaging device for advanced comparison.

Software & Tools

• Python: Organizes images, trains the CNN, and runs performance checks.
• TensorFlow or PyTorch: Builds and tests the image classification model.
• ImageJ: Helps measure lesion area, color change, and image quality metrics.
• Google Colab: Gives you a free place to train models if your laptop is slow.
• scikit-learn: Calculates confusion matrices, precision, recall, and other score metrics.

Experiment Steps

1. Define the disease stages you want the model to separate, especially early symptoms versus clear infection.
2. Plan a consistent image collection method so the model learns disease signals instead of background noise.
3. Build a label system with mentor review so your training data stays accurate.
4. Split your dataset into training, validation, and test groups so you can check generalization honestly.
5. Train a baseline CNN, then compare it with a human scoring method on the same images.
6. Test whether changes in lighting, leaf age, or species reduce model performance, and record where the model fails.

Common Pitfalls

• Labeling borderline leaves too quickly, which teaches the model the wrong symptom stage.
• Mixing plant species or cultivars without tracking them, which hides whether the model learned disease or host differences.
• Using the same plant in both training and test sets, which inflates accuracy by leaking near-duplicate images.
• Photographing leaves under changing light, which makes the model key in on shadows instead of mildew.
• Skipping false-negative analysis, which can make a weak early-detection model look better than it really is.

What Makes This Competitive

A competitive version of this project goes beyond simple accuracy. You compare the model with human scoring, separate early and late disease stages, and report where each method fails. You also test generalization across lighting, plant age, and cucumber-family species. Strong projects use clean labels, a held-out test set, and careful statistics, not just a high headline score.

Project Variations

• Use zucchini leaves instead of cucumber leaves to test whether the CNN generalizes across cucurbit species.
• Compare a CNN trained on close-up leaf crops with one trained on whole-leaf images to see which captures early disease better.
• Add lesion coverage estimates from ImageJ and test whether combining that feature with CNN output improves early detection.

Learn More

• USDA ARS Plant Disease resources: Search for extension-style guides and diagnostic photos on powdery mildew in cucurbits.
• NIH PubMed: Search for review articles on powdery mildew detection, plant disease imaging, and CNN classification.
• USDA National Plant Diagnostic Network: Find plant disease diagnostic references and sample submission guidance.
• MIT OpenCourseWare: Look for free computer vision and machine learning lecture materials to understand CNN basics.
• Plant Disease journal: Search recent peer-reviewed studies on image-based plant disease detection and powdery mildew diagnosis.
• NASA Earthdata: Explore remote sensing and image analysis concepts that help with pattern detection across image sets.