Plant Disease Detection using Transfer Learning

Plant diseases destroy a share of every season's harvest, and farmers usually notice infections only after they have spread beyond easy control. Expert visual diagnosis does not scale — an extension officer cannot visit every field, and misdiagnosis wastes both pesticide and yield. Transfer learning changes the economics of this problem: a network pre-trained on ImageNet already understands edges, textures and shapes, so fine-tuning it on leaf images reaches strong accuracy with tens of thousands of labeled photos instead of millions. This project builds that system as a complete student build: MobileNetV2 fine-tuned on the public PlantVillage dataset for 38 crop-disease classes, wrapped in a demo that classifies a leaf photo, explains its decision with a region heatmap, and shows a treatment note. Because the training is a logged two-phase experiment, the report and viva can discuss exactly what transfer learning bought — frozen-backbone accuracy versus full fine-tune.

1. PlantVillage images are resized to 224×224 and split 70/15/15 into train/validation/test; training images receive augmentation (rotations, flips, color jitter, random affine).
2. Phase 1: the MobileNetV2 backbone (ImageNet weights) stays frozen while only the new 38-way classification head trains for about 5 epochs.
3. Phase 2: the backbone is unfrozen and the whole network fine-tunes for about 15 epochs at a low learning rate with cosine decay and class-weighted loss.
4. Every epoch logs accuracy, loss and per-class F1; the best checkpoint on validation accuracy is kept and exported.
5. At inference, the demo app classifies the leaf photo and renders the top-3 classes with confidence bars.
6. A Grad-CAM-style heatmap overlays the influential leaf regions, and the treatment-note panel shows guidance for the top prediction.

Tech stack:

• Python 3.10, TensorFlow/Keras (MobileNetV2)
• NumPy, scikit-learn (metrics, confusion matrix)
• Matplotlib, Seaborn (training curves, per-class analysis)
• Flask-compatible inference module + single-file HTML/CSS/JS demo app
• PlantVillage dataset (Hughes & Salathé)
• Trained weights shipped as .h5 from the included training run

• [38-class leaf classifier] Predicts the crop and disease (or healthy) from a single 224×224 leaf photo using a fine-tuned MobileNetV2.
• [Top-3 predictions with confidence] Ranked class probabilities with animated confidence bars, so near-ties between similar diseases stay visible.
• [Region-influence overlay] A Grad-CAM-style heatmap showing which leaf areas drove the prediction — key viva material on model explainability.
• [Treatment note panel] A concise, crop-specific treatment note for the predicted condition, clearly labeled as general guidance.
• [Two-phase transfer-learning notebook] Phase 1 trains only the classification head with the backbone frozen; phase 2 fine-tunes end-to-end — both fully logged.
• [Imbalance handling] Class-weighted loss plus augmentation (rotations, flips, color jitter, random affine) to compensate for PlantVillage's skewed class counts.
• [Full evaluation logging] Accuracy, per-class F1 and the confusion matrix on a held-out test split — computed during the build, never pre-claimed.

• Complete source code (data pipeline, two-phase training, evaluation, inference, demo app)
• Jupyter training and evaluation notebook (buyer-run procedure: prepare data, train, evaluate)
• Project report PDF (background, dataset analysis, transfer-learning methodology, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, MobileNetV2, Grad-CAM, augmentation, class imbalance)
• Setup guide (environment, dataset download, training, running the demo)

• PlantVillage photos were shot in controlled conditions; cluttered field backgrounds, multiple leaves and poor lighting reduce accuracy — the report states this deployment gap explicitly.
• Similar-looking diseases are the main confusion source (early vs late blight); the confusion matrix documents it rather than hiding it.
• The system classifies only the 38 PlantVillage classes — crops and diseases outside the dataset cannot be recognized.
• Treatment notes are general guidance, not agronomist advice; the report frames the system as an educational prototype, not a diagnostic authority.
• Class imbalance is real (some classes outnumber others ~10×); the build mitigates it with class weights but rare classes still score lower.