Land cover maps are foundational geospatial data products for understanding our environment and its evolution through time. Typically, these maps (like NLCD made by USGS) take years to create, require manual/expert tweaking, and are limited in geographic scope. At Impact Observatory, we set out to create a global 2020 land cover map, with 10 discrete classes, from Sentinel-2 imagery at 10-meter resolution.
The map was developed with Esri and in partnership with Microsoft AI for Earth. This is an open dataset and freely available with a CC-by license, and you can download it in GeoTIFF format here.
I led the design and development of a deep learning-based segmentation model that takes an input 6-band (red, green, blue, nir, swir1, swir2) Sentinel-2 image at a single point in time, and outputs a 10-class land cover classification. A UNet model was trained from scratch on a massive global dataset consisting of over 5 billion hand-labeled pixels. A weighted categorical cross entropy loss function was used to account for pixel-wise class imbalance, weighting under-represented classes such as grass higher than over-represented classes such as forest.
While instantaneous land cover classifications on single scenes can be useful for some applications, we chose to incorporate many scenes across 2020 to create a land cover map that is representative of the whole year. This approach minimizes issues that inevitably arise due to cloud cover variability and data coverage gaps. More scenes are used in very cloudy parts of the world compared to cloud-free areas.
To take a stack of imagery over a region and convert it into a single map, the segmentation model is run over each scene in the stack. Then, a weighted mode is computed across the predictions, incorporating the probability of each classification along with a custom weight per class. The custom class weight can be thought of as a seasonal adjustment, which emphasizes ephemeral classes that may only occur a few times per year such as grass, and de-emphasizes classes that are transient such as snow/ice.
For more details on this work, check out the IEEE paper linked above. When this map was released, it was the highest resolution global land cover map ever made. Since then, other products have been released such as Worldcover and Dynamic World. I'm grateful for the opportunity to work on a project of such large scope, and learned a ton about geography and modeling in the process.