Hybrid Transfer Learning + Classical ML for Leaf Disease Classification

This repository contains the code used in the manuscript to classify lemon, mango, and pomegranate leaves as healthy vs diseased by:

1. extracting deep features from ImageNet-pretrained CNN backbones (EfficientNet-B0, DenseNet201, MobileNetV2, ResNet50),
2. selecting features with a mutual-information mRMR variant (no leakage; applied per training fold), and
3. training classical classifiers (SVM-RBF, Bagged Trees, kNN, MLP).

The pipeline reports mean ± SD across stratified 10-fold CV, includes imbalance-robust metrics (Balanced Accuracy, MCC, Cohen’s κ), per-class AUPRC, pooled confusion matrices, and generates qualitative activation heatmaps.

---

1. Environment

• Python ≥ 3.10
• PyTorch ≥ 2.0, torchvision ≥ 0.15
• scikit-learn, pandas, numpy, matplotlib, joblib, Pillow, scipy

Example (pip):

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121  # pick wheel for your CUDA/CPU
pip install scikit-learn pandas numpy matplotlib joblib pillow scipy


## 2. Data Layout

The code expects a folder with the structure below (class names are matched case-insensitively and remapped to ["Diseased","Healthy"]):
DATA_ROOT/
├── Healthy/
│   ├── img_001.jpg
│   └── ...
└── Diseased/
    ├── img_101.jpg
    └── ...
## 3. How to Run

Edit the CONFIG block at the top of the script:

DATA_ROOT: dataset root

OUT_DIR: output directory for tables and figures

optional: AUGMENT="light" to enable mild train-time augmentation

Run:python main.py
## 4. Outputs

Under OUT_DIR you will find:

tables_3_7.xlsx

Table3, Table4: best-by-F1 summaries (mean ± SD)

Table5: average Prediction Speed (obs/sec) and Training Time (sec)

Table6_CM: pooled confusion matrices for top-4 configs

Table7_mean and Table7_mean_sd_raw (to render “mean ± SD” per manuscript)

figure5_roc_best_pipeline.png/.svg

figure6_densenet201_svm_heatmaps.png/.svg

cache_full_[backbone]/feats_[backbone].npz (optional, if CACHE_FEATURES=True)