LandCoverNet: Generating a Human-Verified Global Land Cover Classification Training Dataset

Abstract:

LandCoverNet is the first global training dataset for land cover classification using multispectral observations from Sentinel-2 satellites. Land cover classes are defined based on the annual time-series of Sentinel-2, and each pixel is labeled for one of the seven classes: water, natural bare ground, artificial bare ground, permanent snow/ice, woody vegetation, cultivated vegetation and (semi) natural vegetation. Generating labels for a large-scale training dataset such as LandCoverNet requires extensive coordination with labeling users, and deploying tools to facilitate pixel-level labeling. Moreover, human interpretation error is unavoidable at 10 m spatial resolution. Therefore, a consensus algorithm was implemented to generate the label for each pixel in the dataset using inputs from multiple users.

In this presentation, I will provide an overview of the version 1.0 of the dataset which covers the African continent and contains 135 million labeled pixels from Sentinel-2 time-series in 2018. I will then discuss how the accuracy of each labeling user was incorporated into the final product to generate a human consensus score for the labels. Finally, I will present the roadmap to generate the remaining training data for the other parts of the world.