Abstract
Widely used European land cover maps such as CORINE are produced at mediumspatial resolutions (100 m) and rely on diverse data with complex workflowsrequiring significant institutional capacity. We present a high resolution (10m) land cover map (ELC10) of Europe based on a satellite-driven machinelearning workflow that is annually updatable. A Random Forest classificationmodel was trained on 70K ground-truth points from the LUCAS (Land Use/CoverArea frame Survey) dataset. Within the Google Earth Engine cloud computingenvironment, the ELC10 map can be generated from approx. 700 TB of Sentinelimagery within approx. 4 days from a single research user account. The mapachieved an overall accuracy of 90% across 8 land cover classes and couldaccount for statistical unit land cover proportions within 3.9% (R2 = 0.83) ofthe actual value. These accuracies are higher than that of CORINE (100 m) andother 10-m land cover maps including S2GLC and FROM-GLC10. We found thatatmospheric correction of Sentinel-2 and speckle filtering of Sentinel-1imagery had minimal effect on enhancing classification accuracy (< 1%).However, combining optical and radar imagery increased accuracy by 3% comparedto Sentinel-2 alone and by 10% compared to Sentinel-1 alone. The conversion ofLUCAS points into homogenous polygons under the Copernicus module increasedaccuracy by <1%, revealing that Random Forests are robust against contaminatedtraining data. Furthermore, the model requires very little training data toachieve moderate accuracies - the difference between 5K and 50K LUCAS points isonly 3% (86 vs 89%). At 10-m resolution, the ELC10 map can distinguish detailedlandscape features like hedgerows and gardens, and therefore holds potentialfor aerial statistics at the city borough level and monitoring property-levelenvironmental interventions (e.g. tree planting).