2. • Oil Palm is the tree that produces palm oil
• Mostly grown in Africa and South and Central
America
• Deforested land to grow by 53 million hectares
by 2050
• Important to understand where these oil palm
plantations are
• High resolution satellite images combined with
computer vision algorithms can help
5. Datathon Challenge
Planet and Figure Eight in collaboration with
West Big Data hub and WiDS Datathon committee
created a dataset of hi-res satellite imagery
Problem:
Develop a model to detect whether oil palm
plantations are present or not
6. Dataset on Kaggle
• Supervised Learning
problem
• 15K images in training
set
• 4K images in test set
8. Microsoft Custom Vision
• Azure hosted free platform service to build computer vision
model
• Great learning tool for beginners
• Uses AlexNet based Convolutional Neural Networks
9.
10. Custom vision
• Custom Vision Service (UI or API)
1. UI – upload, train and test your model using the UI
2. API – use the Azure Python SDK to train and test the model in Python
3. Hybrid – Train the model using Custom vision UI and export to your
favorite model such as Tensorflow to run Android apps with the
model
Use any of the above options to use the Custom Vision service. I
found the API – option #2 to be most effective
17. Steps
1. Install the Custom Vision SDK for Python
2. Get the training and prediction keys
3. Get the training images
4. Upload and tag images (hasoilpalm, nooilpalm)
5. Train the classifier
6. Get and use the default prediction endpoint
26. Good luck!
Vani Mandava
Microsoft Research
@vanimt
I’ll buy lunch for the first team who gets to .98+ AUC using Microsoft Custom
vision
(restaurant of your choice!)