Plant Disease Detection Using CNN ??? Python & TensorFlow Project (Report, PPT, Viva, Source)
The #1 AgriTech final year project for 2026. PlantVillage dataset, ResNet50 vs MobileNetV2 vs EfficientNetB3, data augmentation pipeline, complete report outline, PPT, and 25 Viva Q&A.
??? Get the Full Working Project Instantly
Tired of broken GitHub code, missing datasets, and last-minute bugs? Get the SubmitKit All-in-One Bundle ??? full 1-click runnable source code, 60-page editable IEEE Black Book (.docx), Viva defense PPT, and top 25 examiner Q&A with full answers. Instant download.
???? Abstract & Problem Statement
India is the 2nd largest agricultural producer in the world, yet 20-40% of crop yield is lost annually due to undetected plant diseases. Farmers visit extension offices physically to get leaf diagnosis ??? 7+ days delay. The Plant Disease Detection using CNN project lets any farmer upload a single leaf photo via smartphone and receive an instant disease name, confidence %, severity, and pesticide recommendation in regional language.
???? PlantVillage Dataset ??? 54,305 Images, 38 Classes
Use the official PlantVillage dataset (Penn State University ??? CC-BY): 14 crop species (Apple, Blueberry, Cherry, Corn, Grape, Orange, Peach, Pepper, Potato, Raspberry, Soybean, Squash, Strawberry, Tomato) ?? 38 healthy / diseased classes. Train: 70% (~38,000), Validation: 15% (~8,100), Test: 15% (~8,100). Always report class-wise accuracy, not just overall.
???? Best Model Architecture 2026
- Input Preprocessing: Resize 224??224, Normalize (mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225])
- Data Augmentation: RandomHorizontalFlip + RandomRotation(??25??) + RandomZoom(??15%) + RandomBrightness + RandomContrast + CutMix (reduces overfitting by ~18%)
- Backbone: EfficientNetB3 (ImageNet pretrained) ??? better than ResNet50 / MobileNet in 2026 because it scales width + depth + resolution together
- Head: GlobalAveragePooling2D ??? BatchNorm ??? Dense 512 ReLU ??? Dropout 0.5 ??? Dense 38 Softmax
- Transfer Strategy: Phase 1 (10 epochs) freeze backbone, train head only. Phase 2 (25 epochs) unfreeze top 30 layers, fine-tune with LR 1e-5 + ReduceLROnPlateau.
- Optimizer: AdamW (weight decay 1e-4) + Label Smoothing 0.1
- Loss: Categorical Crossentropy with Class Weights for imbalanced classes
- Expected Test Accuracy: 98.2-99.1% on PlantVillage held-out (Easy 90+ marks in evaluation chapter)
???? 60-Page IEEE Black Book Chapter Plan
- Title, Certificate, Acknowledgement ??? 2 pages
- Abstract + Keywords ??? 1 page
- Table of Contents, Figures, Tables ??? 3 pages
- Chapter 1 ??? Introduction: Agriculture in India, Yield Loss, Farmer's Need, Problem Statement, 4 Objectives, Scope ??? 5 pages
- Chapter 2 ??? Literature Survey: 8 papers 2020-2026 (AlexNet, VGG, Inception, ResNet, MobileNet, EfficientNet, ViT on PlantVillage) ??? 8 pages
- Chapter 3 ??? Requirement Analysis: Functional (Upload ??? Predict ??? History ??? Admin), Non-Functional, Hardware/Software Tables ??? 5 pages
- Chapter 4 ??? System Design: Block diagram, Use Case, Activity, Sequence, ERD, UI Mockups ??? 8 pages
- Chapter 5 ??? Tools & Algorithms: CNN theory, Transfer Learning, AdamW, EfficientNet paper explanation, CutMix, Grad-CAM ??? 9 pages
- Chapter 6 ??? Implementation: Flask/FastAPI backend, React/Next UI, Model H5 save/load, Prediction function code ??? 10 pages
- Chapter 7 ??? Testing & Evaluation: Confusion matrix (38??38), Per-class F1, ROC-AUC macro, Comparison vs ResNet50 vs MobileNet ??? 7 pages
- Chapter 8 ??? Results & Discussion: Grad-CAM heatmap output screenshots, 3 sample correct + 2 failure mode analysis ??? 3 pages
- Chapter 9 ??? Conclusion + Future Scope: Hindi / Kannada / Telugu vernacular output, drone multi-plant disease, soil-NPK advice integration ??? 2 pages
- References (IEEE, 40+) + Appendix: ER Diagram + Sample Code + Snapshots ??? 3 pages
???? Top 15 Viva Questions (Plant Disease Detection)
Q1.Why EfficientNetB3 and not ResNet?
EfficientNet uses compound scaling (width ??, depth ??, resolution ??) with neural architecture search ??? it reaches higher accuracy at 1/10 the parameters of ResNet50. For 224??224 input: ResNet50 = ~96.5% accuracy vs EfficientNetB3 = ~99.0% on PlantVillage, and 6MB vs 98MB file size (critical for mobile deployment).
Q2.Why do we need data augmentation?
Without augmentation, CNNs memorize training data (overfit). Augmentations create visually-valid variations of each training image. CutMix pastes a disease patch from one image onto a healthy background ??? forces the model to look at the lesion region instead of background bias. Improves generalisation to unseen fields/lighting.
Q3.What is Grad-CAM and why did you show it?
Gradient-weighted Class Activation Mapping = explainable AI heatmap. It shows which pixel regions of the leaf contributed most to the disease prediction class. It proves the model looked at the actual lesion (not background dirt or leaf edge) ??? examiners LOVE XAI demonstrations, it pushes your project from B-grade to A+
Q4.What metrics besides accuracy do you report?
For imbalanced multi-class: per-class precision, recall, F1-score, macro F1, macro ROC-AUC, top-2 error rate, Matthews Correlation Coefficient. Accuracy alone is misleading. A model can get 95% accuracy by just predicting the majority class ??? the confusion matrix uncovers this.
Q5.How do you handle a disease class with only 100 training images?
3 techniques together: (1) Class weights in loss function, (2) SMOTE in embedding space after feature extraction, (3) Synthetic image generation using Stable Diffusion fine-tuned on that class. Also oversample that minority class 3:1 during each batch.
??? Get the Full Working Project Instantly
Tired of broken GitHub code, missing datasets, and last-minute bugs? Get the SubmitKit All-in-One Bundle ??? full 1-click runnable source code, 60-page editable IEEE Black Book (.docx), Viva defense PPT, and top 25 examiner Q&A with full answers. Instant download.