Hugging Face Releases GeoCR: A Generalist Cloud Removal Model for Heterogeneous Sensors and Spectral Domains
By Mr.Xu
Published:
Summary:Hugging Face has released GeoCR, a generalist cloud removal model capable of handling RGB-only, multispectral, and optionally SAR-guided cloudy observations within a single network. By extending a pretrained RGB autoencoder and leveraging a flow transformer, GeoCR achieves state-of-the-art performance across diverse datasets and configurations. This innovation marks a significant advancement in remote sensing image processing, offering a reusable and efficient solution for cloud removal tasks.
Release of GeoCR: A Generalist Cloud Removal Model
Hugging Face has introduced GeoCR, a groundbreaking generalist cloud removal model designed to overcome the limitations of traditional cloud removal methods that are specialized to individual datasets and input configurations. Key features of GeoCR include:
- Multi-modal Data Processing: GeoCR can handle RGB, multispectral, and optionally SAR-guided data, making it adaptable to various sensors and spectral domains.
- Extended Pretrained Architecture: By extending a pretrained RGB autoencoder while keeping its encoder and decoder trunks frozen, GeoCR ensures compatibility with different input and output configurations.
- Flow Transformer Architecture: GeoCR employs a flow transformer to jointly model clean RGB and non-RGB latents, conditioned on separate cloudy-observation streams and optional SAR tokens, enabling efficient cloud removal.
Technical Highlights
- Cross-Domain Generalization: GeoCR was trained on ten datasets comprising 883,331 cloud-free target images, learning a shared cloud removal prior across heterogeneous configurations. This allows GeoCR to perform direct inference without dataset-specific fine-tuning.
- Efficient Adaptability: GeoCR supports efficient adaptation through low-rank adaptation (LoRA) technology, enhancing its flexibility and practicality.
- Superior Performance: GeoCR achieved the best FID and DISTS scores on full-band SEN12MS-CR and Sen2_MTC_New and RGB-only CUHK-CR2, outperforming existing models in multiple benchmark tests.
Industry Impact and Developer Recommendations
GeoCR's release has significant implications for the remote sensing image processing field, particularly in the following areas:
- Multi-Sensor Data Fusion: GeoCR's ability to handle data from different sensors provides an efficient solution for multi-source data fusion.
- Cross-Domain Applications: The model's generalization capabilities make it suitable for various application scenarios, including weather monitoring, environmental protection, and agricultural management.
- Developer-Friendly: The open-source nature and efficient adaptability of GeoCR make it an ideal tool for researchers and developers.
Developers are encouraged to leverage GeoCR's pretrained model and utilize LoRA technology for rapid adaptation to meet specific application needs.
Conclusion
The release of GeoCR marks a significant milestone in cloud removal technology. Its cross-domain generalization capabilities and efficiency open new opportunities for the remote sensing image processing field.
— END —Source: Hugging Face Daily Papers (2026-09-26)
Tags: #Hugging Face #GeoCR #Cloud Removal #Remote Sensing #Multi-modal Data
Community Comments