A CNN model that spots sick plants before a human has to guess.
LeafSpark is a machine learning-based application that detects plant health conditions from leaf images, built as part of a 3rd-semester Artificial Intelligence course project to help lay users identify plant disease symptoms they'd otherwise struggle to recognize visually.
What needed to change
Novice farmers and ornamental plant enthusiasts often can't tell a diseased leaf from a healthy one until it's too late. The team set out to close that gap with a model accurate enough to trust, trained on a dataset that was smaller and messier than we'd have liked.
As full-stack developer and lead on the machine learning side, I sourced training data from Kaggle and got creative with image augmentation to work around the dataset's limitations. I led development of the CNN architecture in TensorFlow, running hyperparameter tuning across different convolutional layer configurations and dropout rates until the model reliably separated healthy from diseased leaves.
The final app pairs a Flutter mobile frontend with a TensorFlow CNN backend and Firestore for data handling, letting users photograph a leaf and get a health assessment back — no expertise required on their end.
Model accuracy achieved after tuning
From data collection to working model
Custom-tuned convolutional architecture
