Land-cover semantic segmentation relies on well-annotated satellite datasets, but the authors of a new arXiv paper argue that current datasets fail to capture the full range of geographic conditions. This gap can limit how well models generalize across different regions and climates.

To address this, the paper introduces a large-scale, climate-aware satellite image dataset. The dataset is explicitly built for domain adaptive semantic segmentation, meaning it is intended to help models transfer knowledge across diverse geographic settings rather than perform well only on the regions they were trained on.

The abstract does not provide details on the dataset's composition or experimental results, but the stated motivation is clear: without geographic diversity, land-cover models risk being biased toward the few areas that dominate existing benchmarks.