Every season, farmers grow calories, not just crops. But to know how much food is being produced, or how climate, conflict, or drought are affecting it, we need to know what is growing where. Up-to-date and accurate crop maps are crucial in informing decisions about food security, water use, subsidies and more. Today, freely available satellite data (for example from the European Copernicus program), in combination with machine learning techniques, makes it possible to produce high-resolution crop type maps almost anywhere and at any time, covering many different crops and seasons. Yet, to train accurate crop identification algorithms, researchers rely on crop observations collected through field surveys, which are costly and typically performed by different organisations, which results in scattered, heterogeneous and typically closed data. This lack of solid reference data is one of the main barriers to building accurate, global crop type maps.
In this blog post, Lucile Jam explains how the WorldCereal project, funded by the European Space Agency (ESA), tries to bridge this gap. Lucile Jam is a master’s student in Applied Mathematics. She completed a six-month internship at ESA, where she worked on the WorldCereal project in collaboration with VITO and Wageningen Environmental Research, contributing to the use of existing crop type maps to address reference data scarcity.
To reduce this gap in trustworthy reference data, WorldCereal's Reference Data Module (RDM) brings together as much publicly available data as possible in one place. Figure 1 shows the current distribution of datasets. In parallel, many crop type maps have already been produced and made public. But these maps are not as reliable as ground-truth data: their quality, resolution and accuracy can vary widely. Therefore, the objective of this work has been to set up and apply a robust procedure that samples training points from high-quality crop/land-cover maps, and to explore how it could help strengthen crop mapping accuracies in regions where field data is scarce or completely missing.
