GEO-AI Challenge for Crop Mapping by ITU - Antoine Saget's submission

This repository present the 2nd place solution for the Zindi GEO-AI Challenge for Cropland Mapping by ITU that achieve a 0.943 accuracy on the private leaderboard.

We provide two jupyter notebooks:

• full_study.ipynb : the full study. We recommend this notebook for people interested in the decisions that lead to the final submission. We strongly recommend reading section 6 where results are summarized and strengths, weaknesses and possible improvements are discussed. This notebook use the extra python files in src/ and the code is not very neat as it was used in the exploratory phase. This study consist of 6 parts:
  • 1 - Downloading the data from GEE
  • 2 - Data preprocessing to obtain aligned timeseries of same lengths
  • 3 - Study of different time spans (length) and periods (start date)
  • 4 - Study of different Sentinel-2 radiometric bands
  • 5 - Reproduction of the submitted solution
  • 6 - Discussion on strengths, weaknesses and possible improvements
• simple_reproduction.ipynb : a simplified version only reproducing the submitted solution. This notebook is self-contained and does not require any extra file. The code is more streamlined and probably easier to integrate in a production pipeline.

Solution

The solution is explained in details in full_study.ipynb. A summary is given below :

• Data:
  • Sentinel-2 time series data is used
  • The time series lengths and periods are optimized per country with CV on the training data
  • The choice of which Sentinel-2 radiometric bands to keep is optimized with CV on the training data
• Model:
  • A simple shallow RandomForest per country (using python scikit-learn implementation).

Installation

There are two ways to run the notebooks:

• With Google Colab
• Using a local python environment

Using Google Colab

• full_study.ipynb (<15min runtime on Google Colab)
• simple_reproduction.ipynb (<3min runtime on Google Colab)

Using a local python environment

The python environment can be installed using the requirements.txt file. We recommend using a virtual environment such as conda or virtualenv to avoid conflicts with your system python installation.

python3 -m venv env 
source env/bin/activate
pip install -r requirements.txt

Then you can run the notebooks using jupyter notebook or jupyter lab.

Folders descriptions

• data/: contains pre-downloaded GEE timeserie data for convenience.
• submission/: contains the original_challenge_submission.csv that has been submitted to the challenge. After running the notebooks and reproducing the results, other .csv submisssion files will be created in this folder.
• src/: contains extra python files used in the full_study.ipynb notebook. None of them are required if you only want to reproduce the submitted solution with the simple_reproduction.ipynb notebook.
  • downloader.py: Downloader class to download data from GEE
  • dataset.py:
    • Dataset and Dataset_country class that handle data loading and easy access to per country subsets
    • Dataset_training_ready and Dataset_training_ready_country class that handle data preprocessing and easy access to per country subsets
  • model.py: Model class that handle model training and prediction in various configurations such as per country models or global model
  • utils.py: various utility functions
  • constants.py

Redownloading the data

We provide pre-downloaded Sentinel-2 GEE timeseries data for convenience in the data/ folder.
The data can be redownloaded from GEE by removing the files in the data/ folder and running the notebooks.
Re-downloading the data from scratch will take up to 1h30min.
As the GEE requests are quite long, we use GEE tasks system. This requires a GEE account and the user must set PROJECT_NAME to it's own GEE project name in the notebooks. Otherwise, the GEE download tasks will fail.