FLAIR-HUB: Large-scale multimodal dataset for land cover and crop mapping
FLAIR-HUB: Large-scale multimodal dataset for land cover and crop mapping
Authors: Anatol Garioud, Sébastien Giordano, Nicolás David, Nicolas Gonthier Date: 2026-04-27 Paper ID: openalex:2506.07080
Summary
FLAIR-HUB is a new large-scale, multimodal Earth observation dataset designed to support advanced land cover and crop mapping research. It integrates six distinct sensor modalities, including aerial imagery, SAR, and multi-spectral satellite time series, paired with high-resolution 20cm ground-truth labels. The authors provide comprehensive benchmarks using various deep learning architectures, demonstrating the efficacy of multimodal fusion and multi-task learning for complex classification tasks. The dataset is intended to facilitate progress in both supervised and self-supervised learning for environmental monitoring.
Key Contributions
- Introduces FLAIR-HUB, a large-scale multimodal Earth observation dataset covering 2528 km² with 20cm resolution annotations across six aligned modalities.
- Establishes extensive baselines for multimodal land cover and crop mapping using state-of-the-art CNN and transformer architectures.
- Demonstrates that integrating six complementary data sources yields peak classification performance of 65.8% mean IoU, highlighting the benefits of multi-sensor fusion.
Open Questions & Future Work
Archivist Review
I have approved the FLAIR-HUB dataset as a significant new benchmark and retained the two proposed open questions as they effectively characterize systemic bottlenecks in remote sensing multimodal and multitask learning. The FLAIR-HUB concept was rejected as a concept entry because it is more appropriately tracked as a dataset.
Approved Open Questions
- Optimal Multimodal Fusion Architectures: Understanding optimal fusion mechanisms is critical for maximizing the utility of increasingly complex and multimodal Earth Observation datasets.
- Multitask Learning Interference Management: Efficient multitask learning is essential for deploying unified models for comprehensive environmental monitoring without the computational cost of separate models.
Rejected Candidates
- [concept] FLAIR-HUB (
flair-hub) - duplicate_existing: This candidate is a dataset and was moved to approved_datasets.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2506.07080
- paper_id
- 2506.07080
- paper_source
- openalex
- domain
- computer-vision
- tags
- computer-visionmultimodal-learningsemantic-segmentationremote-sensingtime-series-analysismultitask-learning
- architectures
-
- datasets
- flair-hub
- concept_slugs
-
- dataset_slugs
- flair-hub
- skill
- TimeSeriesSkill
- created_at
- 2026-04-30T06:05:45Z