Back to Projects
C
Data Science / Analysis

Cassava Disease Detection

ML Engineer (Research)

TensorFlowOpenCVPython

About the Project

This research project developed a computer vision model to identify diseases in cassava crops — a staple food for over 500 million people in Africa. Using image classification techniques, the model can distinguish between healthy plants and four common cassava diseases from leaf photographs, achieving 92% accuracy in field tests. The goal was to give smallholder farmers a practical tool for early disease detection.

What It's Made Of

Image dataset & augmentation

Field-captured leaves · heavy augmentation

Thousands of leaf images captured in real field conditions — varying lighting, angles, and low-end phone camera quality — expanded through aggressive augmentation. The augmentation strategy is what carried the model from lab accuracy to field accuracy, because the training distribution had to look like what a farmer's phone actually produces.

Classification model

TensorFlow · transfer learning · 5 classes

A transfer-learned image classifier distinguishing healthy plants from four common cassava diseases, reaching 92% accuracy in field tests. Transfer learning from pre-trained weights was necessary rather than preferred: the labelled dataset was far too small to train a competitive vision model from scratch.

Preprocessing & inference

OpenCV · Python

The OpenCV stage that normalises an incoming photograph before it reaches the model — cropping, colour and lighting correction, resizing — and the inference path around it. Designed to be cheap enough to run on a mid-range phone, since the deployment target is a smallholder farmer in a rural area, not a server.

Key Highlights

  • Achieved 92% classification accuracy across 5 categories (healthy + 4 disease types)
  • Trained on thousands of field-captured leaf images with data augmentation techniques
  • Built with TensorFlow and OpenCV for image preprocessing and model inference
  • Implemented transfer learning from pre-trained models to work with limited training data
  • Designed for deployment on mobile devices for use in rural farming communities

Technical Challenges

Real-world agricultural images are messy — varying lighting, angles, multiple diseases on one leaf, and image quality from low-end phone cameras. Heavy data augmentation and careful preprocessing were essential to build a model that generalized beyond lab conditions to actual field use.