/iclr-radiant-crops

My solution and code for 2020 ICLR Workshop Challenge #2: Radiant Earth Computer Vision for Crop Detection from Satellite Imagery

Primary LanguagePython

ICLR Workshop Challenge #2: Radiant Earth Computer Vision for Crop Detection from Satellite Imagery

A Data Science competition on Zindi.

  1. Create a RadiantMLHub API Key, and store it in your project directory as api_key.txt. Add the file to your .gitignore.
  2. To download raw data (~16 GB on disk): Run download_data.py
  3. To extract pixel/field-level data: Run extract_pixel_data.py
  4. To calculate remote sensing indexes: Run get_features.py
  5. To calculate time series statistics per pixel: Run get_pixel_stats.py
  6. To train model and make predictions:

    • View script parameters: python3 main.py --h

    • Run script with parameters and save submission file under ./submissions/<current_time>-submission.csv:

      python main.py -cv -fd -md RandomForest -ne 100 -cw -rs 123 -sp

    • Write standard output to file, e.g. python3 main.py ...args... >> model_notes.txt

Docker Image

A docker image is available for a modeling development environment. The image was made to be light weight, and includes only the derived data and training scripts needed to run step 6 above to reproduce my results.

You can run the image by downloading docker and running the following commands (the -v flag binds your local directory with the container so you can access any new submissions files you create):

docker pull cambostein/iclr-radiant-crops:1.2
docker run -v "$(pwd)"/submissions:/app/submissions -it cambostein/iclr-radiant-crops:1.2

Once the container is running and you are connected via bash, follow the instructions above under step 6. You can read more about the project and modeling specifics in the Project Summary