swissALPS: a Dataset for Imbalanced Land Cover Mapping in Alpine Regions

Our dataset consists of very high-resolution aerial images (50cm) and a digital elevation model (50cm) that covers approx. 2300 km2 of land above 2000m altitude in the southwestern part of Switzerland. Our land cover labels focus on alpine land cover that were generated through visual photointerpretation by experts from swisstopo based on the aerial images.The nine land cover types in our study area include bedrock, bedrock with grass, large blocks, large blocks with grass, scree, scree with grass, water area, forest and glacier. The distribution of pixels among classes presents a typical case of a long-tailed distribution with an imbalance factor, defined as the ratio of the most frequent to the rarest class, close to 1000. The classes are grouped by frequency into ‘head’, ‘body’, and ‘tail’ groups.