IGNF/FLAIR-HUB_LC-I_swinbase-upernet

🌐 FLAIR-HUB Model Collection

• Trained on: FLAIR-HUB dataset 🔗
• Available modalities: Aerial images, SPOT images, Topographic info, Sentinel-2 yearly time-series, Sentinel-1 yearly time-series, Historical aerial images
• Encoders: ConvNeXTV2, Swin (Tiny, Small, Base, Large)
• Decoders: UNet, UPerNet
• Tasks: Land-cover mapping (LC), Crop-type mapping (LPIS)
• Class nomenclature: 15 classes for LC, 23 classes for LPIS

🆔 Model ID	🗺️ Land-cover	🌾 Crop-types	🛩️ Aerial	⛰️ Elevation	🛰️ SPOT	🛰️ S2 t.s.	🛰️ S1 t.s.	🛩️ Historical
LC-A	✓		✓
LC-D	✓		✓			✓
LC-F	✓		✓			✓	✓
LC-G	✓					✓
LC-I	✓				✓
LC-L	✓		✓	✓	✓	✓	✓
LPIS-A		✓	✓
LPIS-F		✓				✓
LPIS-I		✓			✓	✓	✓
LPIS-J		✓	✓		✓	✓	✓

🔍 Model: FLAIR-HUB_LC-I_swinbase-upernet

• Encoder: swin_base_patch4_window12_384
• Decoder: upernet
• Metrics:

mIoU	O.A.	F-score	Precision	Recall
43.49%	64.13%	58.09%	59.34%	58.11%

• Params.: 89.2

---

General Informations

• Contact: [email protected]
• Code repository: https://github.com/IGNF/FLAIR-HUB
• Paper: https://arxiv.org/abs/2506.07080
• Developed by: IGN
• Compute infrastructure:
  • software: python, pytorch-lightning
  • hardware: HPC/AI resources provided by GENCI-IDRIS
• License: Etalab 2.0

---

Training Config Hyperparameters

- Model architecture: swin_base_patch4_window12_384-upernet
- Optimizer: AdamW (betas=[0.9, 0.999], weight_decay=0.01)
- Learning rate: 5e-5
- Scheduler: one_cycle_lr (warmup_fraction=0.2)
- Epochs: 150
- Batch size: 5
- Seed: 2025
- Early stopping: patience 20, monitor val_miou (mode=max)
- Class weights:
    - default: 1.0
    - masked classes: [clear cut, ligneous, mixed, other] → weight = 0
- Input channels:
    - SPOT_RGBI : [4,1,2]
- Input normalization (custom):
    - SPOT_RGBI:
        mean: [1137.09, 433.26, 508.75]
        std:  [543.11, 312.76, 284.61]

---

Training Data

- Train patches: 152225
- Validation patches: 38175
- Test patches: 50700

---

Training Logging

---

Metrics

Metric	Value
mIoU	43.49%
Overall Accuracy	64.13%
F-score	58.09%
Precision	59.34%
Recall	58.11%

Class	IoU (%)	F-score (%)	Precision (%)	Recall (%)
building	57.23	72.79	72.23	73.37
greenhouse	49.83	66.52	64.05	69.18
swimming pool	13.76	24.20	39.65	17.41
impervious surface	53.18	69.43	68.13	70.78
pervious surface	39.98	57.12	56.88	57.37
bare soil	43.96	61.07	63.87	58.51
water	71.02	83.06	80.41	85.88
snow	61.98	76.53	69.44	85.23
herbaceous vegetation	36.86	53.87	56.88	51.16
agricultural land	48.21	65.05	57.30	75.23
plowed land	4.56	8.72	12.45	6.71
vineyard	42.06	59.21	69.71	51.47
deciduous	58.67	73.95	73.26	74.66
coniferous	52.95	69.23	71.47	67.14
brushwood	18.07	30.61	34.41	27.56

---

Inference

Aerial ROI

Inference ROI

---

Cite

BibTeX:

@article{ign2025flairhub,
  doi = {10.48550/arXiv.2506.07080},
  url = {https://arxiv.org/abs/2506.07080},
  author = {Garioud, Anatol and Giordano, Sébastien and David, Nicolas and Gonthier, Nicolas},
  title = {FLAIR-HUB: Large-scale Multimodal Dataset for Land Cover and Crop Mapping},
  publisher = {arXiv},
  year = {2025}
}

APA:

Anatol Garioud, Sébastien Giordano, Nicolas David, Nicolas Gonthier. 
FLAIR-HUB: Large-scale Multimodal Dataset for Land Cover and Crop Mapping. (2025). 
DOI: https://doi.org/10.48550/arXiv.2506.07080