Back in the spring, I had some fun doing experimentation around machine learning models for labelling satellite imagery with atmospheric conditions and various classes of land cover/land. You can check out my work in an open source Kaggle notebook. For this I trained on the data provided in the Planet Kaggle competition, Understanding the Amazon from Space.
This notebook is merge of https://github.com/planetlabs/planet-amazon-deforestation/blob/master/planet_chip_examples.ipynb and https://www.kaggle.com/code/hortonhearsafoo/fast-ai-v3-lesson-3-planet with port to fastai v2 and vast simplification of training.
If you choose to run on Kaggle the following settings might be useful:
- data If you need access to data from Kaggle: go to add data, then search for planet. It’s the top result; unfortunately only JPGs not TIFF
- settings In Kaggle, make sure to go to settings on the right, and choose GPU as accelerator and enable internet. I set the environment to the latest environment