Sen4AgriNet: A Harmonized Multi-Country, Multi-Temporal Benchmark Dataset for Agricultural Earth Observation Machine Learning Applications

This work introduces a one of its kind labeled Earth Observation based benchmark dataset, Sen4AgriNet, for agricultural applications in Europe. The dataset is labeled using farmer declarations for the period 2016–2020, available as open data only very recently. The Sen4AgriNet contains 42.5 million parcels hence is significantly larger than the existing archives and is tailored to be used as a training source in the context of deep learning. It consists of two sub-datasets: Object Aggregated Dataset (OAD) and Patches Assembled Dataset (PAD). OAD dataset capitalizes zonal statistics of each parcel, thus creating a powerful label-to-features instance for classification algorithms. On the other hand, PAD structure generalizes the classification problem to parcel extraction and semantic segmentation and labeling. Key advantages from other similar datasets are the inclusion of all bands, multicountry and multi-year time span, and standardization of the crop type taxonomy across Europe. Finally, we showcase the potential of Sen4AgriNet through machine learning experiments. All data and code are accessible here: https://sen4agrinet.space.noa.gr/