saaga23/plant-disease-indabax2026

Reproducible MobileNetV2 training protocol for plant-disease classification on African smallholder crops. # Plant Disease Detection - IndabaX Nigeria 2026 A reproducible training protocol for edge-based plant-disease classification on African smallholder crops. ## What it does This repository contains a single Jupyter notebook that trains a lightweight image classifier to recognize crop diseases from leaf images. It is designed to run on CPU or edge hardware and targets reproducible research for IndabaX Nigeria 2026. ## How it works - **Data preparation:** images are organized by crop and disease class. Exact duplicate images are removed with MD5 hashing before splitting. - **Split:** stratified 70/20/10 train/validation/test split with a fixed seed so results can be reproduced. - **Model:** MobileNetV2 pretrained on ImageNet, with a custom dense head. - **Training:** - Phase 1 trains only the classification head for 10 epochs with the base frozen. - Phase 2 unfreezes the top 50% of the convolutional base and fine-tunes for up to 20 epochs at a lower learning rate. - An auto-fallback reverts to Phase 1 weights if Phase 2 does not improve validation accuracy by at least 0.5 percentage points. - **Augmentation:** random flips, rotation, zoom, and brightness changes are applied only to the training set. ## Tech stack - Python 3 - TensorFlow / Keras - NumPy, pandas, scikit-learn - Matplotlib, seaborn ## Results / Metrics Held-out test-set accuracy after fine-tuning: | Crop | Baseline (ImageNet) | Final accuracy | Improvement | |----------|--------------------:|---------------:|------------:| | Tomato | 34.69% | 97.96% | +63.27 pp | | Cassava | 52.42% | 98.39% | +45.97 pp | The notebook also reports per-class precision, recall, and F1, and saves confusion matrices and a results CSV. ## How to run 1. Open `plant-disease.ipynb` in Kaggle or Jupyter. 2. Provide the dataset. On Kaggle, mount a dataset containing `cassava/` and `tomatoes_combined/` directories; locally, place the data in the expected root path or u …