BELDE: Building a Large-scale Earth-observation Land-cover Dataset for Europe

BELDE is a newly introduced large-scale dataset containing over 1 million RGB satellite image-segmentation pairs from Europe, designed to advance earth observation and land-cover segmentation models. The dataset achieves strong in-domain performance (83% F1 score) but reveals significant challenges in cross-geographic generalization, with accuracy dropping substantially on non-European regions.

BELDE addresses a critical infrastructure gap in AI model development by providing a continental-scale, publicly accessible benchmark for RGB-based earth observation. Earth observation imagery is fundamental to climate monitoring, urban planning, and disaster response, yet training datasets have historically suffered from limited geographic coverage or proprietary restrictions. The introduction of over 1 million curated image pairs from Sentinel-2 satellite data represents a substantial contribution to the open-source machine learning community.

The research reveals an important constraint in current deep learning approaches: models trained on European imagery experience significant performance degradation when applied to other regions, with F1 scores dropping from 83% to 58% on Korean data. This geographic domain shift problem mirrors challenges facing other AI domains and underscores the importance of diverse, multi-regional training data. The inclusion of test sets from Korea and North America enables systematic evaluation of cross-domain generalization, advancing the field beyond single-region benchmarks.

For developers and researchers, BELDE provides an immediately usable foundation for training production earth observation models while offering a testbed for developing domain adaptation techniques. Organizations working on climate tech, agricultural monitoring, or urban analytics can leverage this dataset to accelerate model development. The public release strategy democratizes access to high-quality satellite data, reducing barriers to entry for smaller research teams and startups. However, the documented performance gap on out-of-domain data signals that practitioners cannot rely on simple transfer learning approaches and must invest in region-specific model refinement or ensemble techniques for global deployment.

• →BELDE contains 1.088 million curated RGB satellite image pairs from Europe with 10m spatial resolution, establishing one of the largest publicly available land-cover segmentation datasets.
• →Models achieve 83% F1 score on European test data but drop to 58.3% on Korean data, demonstrating significant geographic domain shift challenges.
• →The dataset includes geographic variants (Korea, California-Nevada) enabling systematic evaluation of cross-region model generalization capabilities.
• →Public release of BELDE reduces barriers for climate tech and earth observation startups to access high-quality training data.
• →Results highlight the necessity for region-specific model adaptation rather than simple transfer learning in satellite imagery applications.