The WorldCereal Processing Module
Generic overview
The WorldCereal Processing Module has been designed to generate custom cropland and crop type maps, anywhere on the globe. Users can either use our default models or train a custom classification model, finetuned for their region, time and crops of interest.
The figure below provides a full conceptual overview of the Processing Module:
In short, the Processing Module deals with the following steps:
- Fetching Earth Observation (EO) and ancillary input data needed for training/inference
- Pre-processing the fetched time series and ensuring all inputs are resampled to 10 m resolution
- Using the Presto deep learning framework to compute classification features, based on the extracted inputs
- Training a custom CatBoost classification model
- Applying a trained (or default) model on an area and season of interest
- Spatially cleaning the model predictions and associated class probabilities
NOTE: more details on the individual components of the processing module will be added here…