EuroCropsML

EuroCropsML is a ready-to-use ML dataset combining EuroCrops reference data with Sentinel-2 reflectance data from 2021. It contains data from Latvia, Portugal, and Estonia and is intended for benchmarking few-shot crop type classification. We used Eurostat's GISCO dataset to map the EuroCrops parcels to their NUTS1-3 region.

The provided data comes in two stages:

1. raw_data.zip (stage 1): One dataframe per country containing a annual time series of observations for each parcel, as well as separate files for the parcels' geometries and classes (EC_hcat_c = 10-digit HCAT code indicating the hierarchy of the crop).
2. preprocess.zip (stage 2): Read-to-use .npz-files. Each data point is saved in an .npz-file along with its metadata (parcel's centroid in [lon,lan]; observation dates). In addition, we performed some cloud removal steps. Each .npz-file is saved with the following naming convention: <NUTS3region>_<parcelID>_<EC_hcat_c>.npz

Furthermore, split.zip contains .json-files that split the files from preprocess.zip into a pre-training/meta-learning (train and validation) and fine-tuning (train, validation, and test) dataset. In total, we provide two use cases:

• latvia_portugal_vs_estonia: pre-training on Latvia and Portugal (142 distinct classes), fine-tuning on Estonia (127 distinct classes, of which 34 have not been seen during pre-training)
• latvia_vs_estonia: pre-training on Latvia (103 distinct classes) and fine-tuning on Estonia (127 distinct classes, of which 46 have not been seen during pre-training)

For both use cases, the fine-tuning split is as follows:

• train: 1-, 5-, and 10-shot (for few-shot classification and benchmarking)
• validation: 1000 samples
• test: all samples

Changelog

• Version 4: The EuroCrops shapefilees sometimes contain a couple of parcels that lie outside the national borders. We now map them to the closest NUTS region within the country. Please rely on this version.
  
• Version 3: Some parcels have been clipped incorrectly. 
• Version 2: Remove datapoints that contain only cloudy observations (in preprocess.zip).
• Version 1: Initial publication