Repository containing notebooks to get started on the CV4A challenge at ICLR 2020 using eo-learn.
The cv4a-crop-challenge-to-eolearn notebook converts the data provided as .tiff files into smaller EOPatch format files. This allows to better handle and visualise data as rasters, and to easily apply processing pipelines.
The cv4a-process-and-train notebook shows how to set up a processing pipeline on EOPatch objects, such as cloud masking and feature interpolation. This way, different pre-processing methods can be quickly tested.
The pipeline shown in the notebook includes:
- conversion of cloud probabilities to cloud masks;
- dilation of cloud masks;
NDVIcomputation;- linear interpolation to fill missing values;
- utility tasks to get insights into data.
Features are then aggregated and used to train/evaluate a machine learning model.
In this starter's notebook, an untuned random forest classifier was trained on the temporal features, achieving a public score of 1.26628.
The SampleSubmission.csv template file is added for completion.
The notebook assume that the data has been downloaded according to the challenge instructions. Set the path to the data in the notebooks as ROO_DATA_DIR.
Installing eo-learn according to instructions should cover all dependencies used in the notebooks.
As already noted by organisers and participants, these additions should improve performance and generalisation of the methods:
- dealing with class imbalance (e.g. over-sampling, under-sampling, SMOTE);
- feature analysis and engineering adding domain specific indices. You can check this repo for inspiration on vegetation indices derived from Sentinel-2 data;
- use ML methods that better characterize temporal evolution of crops.
Good luck to all.