Metadata

This is a very interesting one, where a standard (CORINE) classification is mimicked with OSM where enough good data is available, and then sparser areas are filled in with satellite data and standard ResNet’s. A useful pattern that’d be interesting to see if it works in less standard setups as well.

Highlights

large-area fusion of OpenStreetMap and Copernicus data at a spatial resolution of 10 m or finer and can be applied globally.

land use labels from OpenStreetMap and remote sensing data to create a contiguous land use map of the European Union as of March 2020.

Country-specific deep learning convolutional neural networks and Sentinel-2 feature space composites of 2020 at 10 m resolution were employed. The overall map accuracy is 89%, with class-specific accuracies ranging from 77% to 99%. The data set is available for download from https://doi.org/10.11588/data/IUTCDN and visualization at https://osmlanduse.org.

LULC products benefitted most notably by the use of remote sensing, its proliferation through open data policies4 and artificial intelligence5. Currently the further accelerated use of such technology is primarily limited by the availability of sufficient thematically labelled data

Given the fragmented availability of OSM data and opportunistic availability of thematic content and depth, resulting LULC products are consequently incomplete in terms of spatial coverage. Such gaps in unlabelled LULC data can be addressed by performing basic classification of remote sensing data, using known areas as training data

Copernicus Sentinel 2 (S2) 10 m multi spectral feature space processed through Food and Agriculture Organization (FAO) sepal.io system. Note: This is an interesting project, note to explore further. Feature space was a best pixels medoid composite of Sentinel 2 bands red, green, blue (RGB) and near infrared (NIR) at 10 m of the past three years as of April 2020.

Our decision to train separate classification models per country stems from the varying completeness and likely slightly different tagging cultures of OpenStreetMap data across Europe, may differ at national borders.