🌿 PlantVillage Disease Classification

A deep learning pipeline for plant leaf disease recognition using CNNs and transfer learning.

---

📌 Project Overview

This project presents an end-to-end image classification pipeline for plant leaf disease identification using the PlantVillage dataset. The pipeline covers the full machine learning workflow, from dataset preprocessing to quantitative evaluation and visual interpretability analysis.

We investigate the performance of two convolutional neural network (CNN) architectures under a transfer learning setting:

• ResNet-18, leveraging residual connections for stable optimization
• EfficientNet-B0, employing compound scaling for parameter-efficient learning

Both models are fine-tuned from ImageNet-pretrained weights and evaluated under a consistent experimental setup to enable fair comparison.

The proposed pipeline consists of the following components:

• Dataset preprocessing and stratified splitting into training, validation, and test subsets
• Model training and fine-tuning, including a two-phase optimization strategy
• Quantitative evaluation, reporting accuracy, macro-averaged F1 score, and confusion matrices
• Qualitative model interpretability using Grad-CAM visualizations to highlight discriminative image regions
• Automated result visualization and experiment archiving, ensuring reproducibility across model variants

This repository is designed to be modular and extensible, allowing additional architectures or datasets to be incorporated with minimal changes to the codebase.

---

🗂 Dataset

Dataset: PlantVillage
Source: https://www.kaggle.com/datasets/emmarex/plantdisease/data

After downloading, place the dataset at:

data/raw/PlantVillage/

Summary

• Number of classes: 15
• Total images: ~20,600
• Image format: RGB, mostly 256×256
• Split ratio:
  • Train: 70%
  • Validation: 15%
  • Test: 15%

Classes (15)

Pepper (bell):

• Bacterial spot
• Healthy

Potato:

• Early blight
• Late blight
• Healthy

Tomato:

• Bacterial spot
• Early blight
• Late blight
• Leaf mold
• Septoria leaf spot
• Spider mites (Two-spotted spider mite)
• Target spot
• Tomato yellow leaf curl virus
• Tomato mosaic virus
• Healthy

---

🧠 Models

ResNet-18 (Baseline Transfer Learning)

• Pretrained on ImageNet
• Two-phase training:
  • Phase 1: Train classifier head only
  • Phase 2: Fine-tune full network with reduced learning rate

EfficientNet-B0 (Baseline Transfer Learning)

• Pretrained on ImageNet
• Same training strategy and hyperparameters as ResNet-18
  (ensures fair comparison)

---

⚙️ Training Setup

• Framework: PyTorch
• Device: Apple M1 Max (32GB RAM)
• Accelerator: MPS
• Optimizer: AdamW
• Loss: Cross-Entropy
• Early stopping: Enabled
• Image size: 224×224
• Normalization: ImageNet mean & std

⏱ Approximate training time per model:
~30 minutes (early stopped)

---

📊 Results

Final evaluation is performed on the held-out test set.

Quantitative Metrics

Model	Test Accuracy	Macro-F1	Top-3 Accuracy
ResNet-18	~0.883	~0.870	~0.977
EfficientNet-B0	~0.816	~0.774	~0.948

Exact metrics are stored in each model’s results.json.

---

ResNet-18 — Training Curves, Confusion Matrix, Errors

---

EfficientNet-B0 — Training Curves, Confusion Matrix, Errors

---

🧠 Model Explainability — Grad-CAM

Grad-CAM highlights image regions that most influenced the model’s prediction.

Example (ResNet-18)

Example (EfficientNet-B0)

Full Grad-CAM galleries:

• reports_archive/reports_20260126_143813_resnet18/figures/gradcam/
• reports_archive/reports_20260126_151938_efficientnet/figures/gradcam/

---

🧪 Reproducibility

Install Dependencies

pip install -r requirements.txt

Run Full Pipeline

chmod +x scripts/run_all.sh
./scripts/run_all.sh

Switch Models

To switch models, change one line only in scripts/run_all.sh:

CONFIG=configs/resnet18.yaml
# CONFIG=configs/efficientnet_b0.yaml

---

📁 Project Structure

configs/                # Model & training configs
src/
├── data/               # Dataset, transforms, splits
├── models/             # CNN & transfer models
├── explain/            # Grad-CAM
├── train.py
├── eval.py
├── visualize.py
reports_archive/        # Timestamped experiment outputs
scripts/
requirements.txt
README.md

---

⚠️ Limitations

• Clean background bias: Images are captured under controlled conditions.
• No real-world field images: Limited lighting and background diversity.
• Generalization risk: Performance may degrade in real agricultural environments.

---

🚀 Future Work

• Add real-world field datasets
• More realistic data augmentation
• Additional crop species
• Model ensembling
• Lightweight deployment (ONNX / CoreML)
• Prediction uncertainty estimation