AI for Climate Change Hackathon

This repo contains the work my team and I have done in the context of the AI for Climate Change Hackathon that took place at MILA from September 13 to 15 2019.

Our team trained a convolutional neural network to identify plant deseases from a dataset of over 44 000 thousand images.

The aim was to deploy the application on mobile simulate real-time classification from a drone camera feed.

We have successfully trained our own model, but ultimately opted to retrain Tensorflow's mobilenetv2 to make a lightweight model and deploy quickly.

This repo contains all of our models as well as the template android app (source: Tensorflow examples).

The dataset can be downloaded at https://www.dropbox.com/s/7jvh6aeheja9qar/plant-original-dataset.zip?dl=1

Source: PlantVillage

Team: Ayoub El-Hanchi, Antoine Frau, and Artsiom Skliar

initial-model

This folder contains the CNN we built from scractch using the Keras library. We have reached an accuracy of 92% on the training set and 80% on the validation set.

transfer-learning-apples

This folder contains a retrained mobilenetv2 model trained to only recognize the apple specimens from the provided dataset.

transfer-learning-full-dataset

Once we successfully retrained the mobilenetv2, we attempted to train it on the full datasets.

andrdoid

This folder contains the app template to launch and deploy the tflite model of choice.