Accurate land cover classification of tundra is critical to modelling of future climate impacts in the Arctic. Existing classification systems (CAVM, NALCMS, GLC2000) are based on low- to medium-resolution remote sensing imagery which may result in low accuracy where land cover is variable. In this study we focus on southern Baffin Island to create a new classification based on 10-m resolution NDVI from 2018 Sentinel-2 satellite imagery, compare this new classification with existing systems, and characterize dependence of NDVI on spatial resolution based on three ~25-km2 0.5-m resolution WorldView-2 images. We recognized six different land cover types including Polar Semi-Desert (40%), Mesic Tundra (21%), and Wet Sedge Meadow (18%), based on previous studies of Baffin Island. Percent area of these six types within CAVM, NALCM, and GLC2000 cover classes was highly variable. As measured by the relative coefficient of variation, variations in NDVI were greatest at higher spatial resolutions, increasing by 800% with a shift from 0.5-10m2, versus 200% from 10-100 m2. Our results suggest that existing pan-Arctic classifications may not accurately capture land cover patterns on Baffin Island and likely other high-latitude regions, and highlight the need to use these classifications with caution when used to generalize Arctic ecosystem responses to climate change.