BELDE: Building a Large-scale Earth-observation Land-cover Dataset for Europe
ECCV 2026 | Ümit Mert Çağlar
, Alptekin Temizel | METU, Turkey
TL;DR: We introduce BELDE, a continental-scale RGB Earth-observation benchmark built from Sentinel-2 true-color imagery and ESA WorldCover maps, containing 1,088,385 curated 256×256 image-mask pairs across Europe at 10 m resolution. We evaluate 17 segmentation architectures and show that compact hybrid transformers (LALE-S2, 2.61M parameters) retain over 94% of the performance of a 100M-parameter model, while two out-of-domain extensions, BELDE-K (Korea) and BELDE-CA-NV (California-Nevada), reveal significant performance degradation under geographic domain shift.
Key Contributions
- Large-Scale Curated Dataset: BELDE, 1,088,385 geo-aligned RGB image-mask pairs across Europe at 10 m resolution, produced by an automated pipeline that prunes over 1.94 million non-informative, corrupted, or single-class dominated patches.
- Efficient Model Distillation: Demonstration that lightweight hybrid architectures (e.g., LALE-S2, 2.61M parameters) retain over 94% of the performance of 100M-parameter models, enabling deployment on bandwidth- and payload-constrained platforms.
- Cross-Region Generalization Benchmarks: Two out-of-domain extensions, BELDE-K (16,607 pairs, Republic of Korea) and BELDE-CA-NV (88,155 pairs, California-Nevada, USA), built with the same pipeline to quantify geographic domain shift.
- Architectural Evaluation: A comprehensive benchmark of 17 semantic segmentation architectures spanning CNN, transformer, and lightweight hybrid families, plus a data-scaling study showing that diffusion-based synthetic augmentation improves segmentation performance beyond full real-data capacity.
The BELDE Dataset
BELDE is constructed from Sentinel-2 true-color imagery (TCI) and ESA WorldCover 2021 land-cover maps (77.9% regional label accuracy over Europe), covering 13°W–50°E and 33°N–60°N. The original 11 WorldCover classes are harmonized into 7 (tree, shrub, grass, crop, built-up, barren, water), and patches with missing data or more than 90% water coverage are discarded, pruning 1,941,727 patches in total.
| Dataset | Patches | Classes | Resolution | Patch Size | Multi-spectral |
|---|---|---|---|---|---|
| BigEarthNet | 549,488 | 19 | 10m | 120×120 | No |
| ARAS (Real) | 100,240 | 7 | 10m | 256×256 | No |
| EuroSAT | 27,000 | 10 | 10m | 64×64 | Yes |
| FLAIR | 77,762 | 19 | 0.2m | 512×512 | Yes |
| LoveDA | 5,987 | 7 | 0.3m | 1024×1024 | No |
| DeepGlobe | 1,146 | 7 | 0.5m | 2448×2448 | Yes |
| LandCover.ai | 10,674 | 4 | 0.5m | 512×512 | No |
| YieldSAT | 113,555 | 12 | 20m | Varied | Yes |
| BELDE (ours) | 1,088,385 | 7 | 10m | 256×256 | No |
| BELDE-K (ours) | 16,607 | 7 | 10m | 256×256 | No |
| BELDE-CA-NV (ours) | 88,155 | 7 | 10m | 256×256 | No |
Semantic Segmentation Results
We evaluated 17 architectures spanning CNN (DeepLabV3, DeepLabV3+, FPN, LinkNet, PSPNet, UNet, UNet++), transformer (SegFormer, DeiT3, MaxViT) and lightweight hybrid (EfficientFormer, FastViT, LALE) families, ranging from 1.5M to 117M parameters, all trained under an identical AdamW + Dice-loss protocol.
| Architecture | F1 | IoU | Params |
|---|---|---|---|
| DeepLabV3 | 80.2 | 69.2 | 33.06M |
| DeepLabV3+ | 81.2 | 70.4 | 29.49M |
| FPN | 81.1 | 70.4 | 30.17M |
| LinkNet | 81.0 | 70.2 | 28.75M |
| PSPNet | 77.9 | 66.2 | 28.44M |
| SegFormer | 81.1 | 70.5 | 28.89M |
| UNet | 81.8 | 71.3 | 31.22M |
| UNet++ | 81.7 | 71.2 | 31.91M |
| DeiT3 | 81.5 | 71.0 | 117.08M |
| EfficientFormer-L1 | 81.6 | 71.1 | 29.83M |
| EfficientFormer-L3 | 82.5 | 72.3 | 49.13M |
| EfficientFormer-L7 | 83.0 | 72.8 | 100.32M |
| FastViT-mci0 | 81.6 | 71.1 | 29.07M |
| FastViT-sa12 | 82.2 | 71.8 | 29.24M |
| MaxViT | 82.9 | 72.8 | 60.80M |
| LALE-S3 | 78.4 | 66.8 | 3.66M |
| LALE-S2 | 78.2 | 66.5 | 2.61M |
Key Takeaways:
- Transformers Lead, Lightweight Models Compete: EfficientFormer-L7 (100.3M params) and MaxViT (60.8M params) achieve the highest F1-scores (83.0% and 82.9%), while LALE-S2 reaches 78.2% F1 with only 2.61M parameters, retaining 94.2% of peak performance at 38.4× fewer parameters.
- In-Domain Performance Nears a Label-Noise Ceiling: The narrow 1.5% IoU spread between CNN and transformer baselines suggests in-domain accuracy is bounded by the ~77.9% accuracy of the underlying WorldCover pseudo-labels.
- Shrub is the Hardest Class: Spectrally homogeneous classes (tree, water) segment well across all models, while the rare and visually ambiguous shrub class remains the most challenging under RGB-only observation.
Cross-Region Generalization
Models trained exclusively on BELDE (Europe) were evaluated zero-shot on BELDE-K and BELDE-CA-NV. All architectures show consistent performance degradation under domain shift, MaxViT and EfficientFormer-L7 lead on both extensions (58.2%/58.3% F1 on BELDE-K, 66.4%/66.2% F1 on BELDE-CA-NV), while LALE models trade some accuracy for stability and efficiency. This confirms that RGB-only segmentation models rely heavily on region-specific spatial and textural cues, motivating geographically diverse benchmarks like BELDE.
| Architecture | BELDE-K F1 | BELDE-CA-NV F1 |
|---|---|---|
| UNet | 56.6 | 65.1 |
| UNet++ | 56.5 | 65.1 |
| SegFormer | 55.4 | 65.0 |
| MaxViT | 58.2 | 66.4 |
| EfficientFormer-L7 | 58.3 | 66.2 |
| LALE-S3 | 52.1 | 60.9 |
| LALE-S2 | 51.3 | 60.3 |
For reference, in-domain (European) F1 scores range from 77.9-83.0%, so even the best zero-shot transfer models lose roughly 20-25 F1 points when moved to a different continent.
We further evaluated data scaling and diffusion-based synthetic augmentation on a BELDE subset: real-data scaling alone reaches 74.66% F1 at full capacity (80,000 pairs), and adding synthetic samples from a latent diffusion model conditioned on segmentation masks improves this further to 75.68% F1 (+1.02%), showing that generative augmentation provides utility beyond real-data availability.
Compute: 211 training runs totaling 1,077 GPU-hours on NVIDIA H100s — 544 hours for architecture search/conceptualization, 206 hours for the California-Nevada extension, 122 hours for the Korea extension, and the remainder split across the CNN, transformer, and lightweight architecture benchmarks. All results are reported as mean ± standard deviation over 3 repeated runs.
Dataset Availability and Reproducibility
The datasets and software framework created and used throughout this work are publicly available under the Creative Commons Attribution 4.0 International license (CC BY 4.0).
| Resource | Link |
|---|---|
| BELDE | datasets/caglarmert/BELDE |
| BELDE-K | datasets/caglarmert/BELDE-K |
| BELDE-CA-NV | datasets/caglarmert/BELDE-CA-NV |
| Source Code | github.com/caglarmert/BELDE |
Citation
If you find this dataset or codebase useful in your research, please consider citing:
@article{caglar2026belde,
title={BELDE: Building a Large-scale Earth-observation Land-cover Dataset for Europe},
author={Caglar, Umit Mert and Temizel, Alptekin},
journal={arXiv preprint arXiv:2606.20909},
year={2026}
}