LuojiaSET-OSFCR Dataset Overview

• LuojiaSET-OSFCR dataset is a globally representative remote sensing benchmark that provides spatially aligned quadruplets including cloudy optical, SAR, cloud-free optical, and land-cover labels.
• It is designed to support integrated research in cloud removal and land-cover segmentation with 20,000 high-resolution 256×256 patches representing diverse land-cover types and cloud conditions.
• The dataset underpins state-of-the-art joint models by enabling prompt-guided SAR-optical fusion and quantifiable benchmarking using metrics such as PSNR, SSIM, and mIoU.

The LuojiaSET-OSFCR dataset (Luojia Satellite Earth Observation Set for Optical-SAR Fusion, Cloud Removal & Segmentation) is a large-scale, globally representative remote sensing benchmark designed to advance research in multi-modal cloud removal (CR) and land-cover segmentation (LCS). Specifically constructed to provide spatially aligned quadruplets—cloudy optical, synthetic aperture radar (SAR), cloud-free optical, and land-cover label images—it supports both restoration of occluded optical imagery and downstream semantic scene analysis. The dataset underpins recent progress in joint CR-LCS models, such as task-driven @@@@3@@@@ frameworks, by delivering diverse, precisely co-registered, and analysis-ready data at high spatial resolution (Zhang et al., 17 Jan 2026).

1. Scope, Purpose, and Design

LuojiaSET-OSFCR is constructed to address critical bottlenecks in Earth observation, where persistent cloud occlusion in optical sensors impairs both the visual @@@@4@@@@ and semantic interpretability required for automated analysis. The dataset provides 20,000 globally sampled, 256 × 256 pixel patches at 10 m ground sampling distance (@@@@6@@@@), ensuring coverage across a broad spectrum of land-cover types and cloud conditions. Its explicit design objective is to support integrative research in joint cloud removal and land-cover segmentation by making available precisely aligned multi-modal inputs and reference outputs, facilitating experiments in analysis-ready data pipelines (Zhang et al., 17 Jan 2026).

2. Modalities, Channels, and Preprocessing

LuojiaSET-OSFCR comprises the following strictly co-registered modalities for each patch:

• Optical (Sentinel-2):
  • 13 spectral bands:
  • 4 bands at 10 m (B2, B3, B4, B8)
  • 6 bands at 20 m (B5, B6, B7, B8a, B11, B12)
  • 3 bands at 60 m (B1, B9, B10)
  • All bands are resampled to 10 m GSD and co-registered.
  • Preprocessing includes radiometric calibration, atmospheric correction (e.g., Sen2Cor), and per-band normalization to [0, 1].
• SAR (Sentinel-1):
  • Dual-polarization channels: VV and VH
  • Data processed to 10 m GSD and co-registered to the optical grid.
  • Preprocessing: radiometric calibration, speckle filtering (e.g., Lee filter), and terrain correction.

This consistent pre-alignment and normalization across modalities is critical for pixel-wise tasks in multimodal fusion frameworks.

3. Cloud and Segmentation Annotation Methodology

For each patch, the dataset contains:

• Cloudy Optical Image ($I_c$): Observed scene with cloud occlusion.
• Clear-Sky Optical Reference ($I_{gt}$): Matched scene under cloud-free conditions.
• Land-Cover Label ($Y$): Pixel-wise semantic annotation based on standard land-cover taxonomy (e.g., water, vegetation, built-up, bare soil).

Notably, no explicit cloud masks or “cloud thickness” labels are distributed. Instead, cloud presence is implicit in the difference between $I_c$ and $I_{gt}$. The underlying methodology for “degradation prompts” does not involve manual annotation; rather, a learned prompt-generator network $g_\theta$ produces a per-pixel, multi-channel degradation embedding $P \in \mathbb{R}^{C_p \times H \times W}$. This embedding encodes spatial uncertainty such as cloud boundaries and thickness and guides prompt-driven model fusion of SAR and optical features during training and inference.

Segmentation labels are distributed as single-channel, 256 × 256 integer masks, generated by expert-reviewed automatic classification on cloud-free Sentinel-2 composites and manually refined. The class taxonomy itself is inherited from the original LuojiaSET release; the specific number of classes is not detailed in (Zhang et al., 17 Jan 2026), but global land-cover datasets typically include 6–10 classes.

4. Dataset Partitioning and Statistical Properties

LuojiaSET-OSFCR adopts a random split for supervised learning tasks:

Split	Number of Patches	Percentage
Training	16,000	80%
Validation	2,000	10%
Test	2,000	10%

Cloud coverage within the dataset traverses the spectrum from 0% (clear) to over 80% (strong overcast), providing a challenging basis for both restoration and semantic segmentation tasks. No explicit stratification or per-split cloud-coverage histograms are reported.

5. Evaluation Metrics and Baseline Performance

LuojiaSET-OSFCR supports quantitative benchmarking through standardized restoration and segmentation metrics:

• Cloud Removal:
  • Peak Signal-to-Noise Ratio (PSNR)

  $$\mathrm{PSNR} = 10\,\log_{10}\left(\frac{\mathrm{MAX}^2}{\mathrm{MSE}}\right), \quad \mathrm{MSE} = \frac{1}{HW}\sum_{i,j}(\hat{I}_{ij}-I_{gt,ij})^2.$$ - Structural Similarity Index (SSIM)

• Segmentation:
  • Pixel Accuracy (PA)

  $$\mathrm{PA} = \frac{\sum_{c} \mathrm{TP}_c}{\sum_{c}(\mathrm{TP}_c+\mathrm{FP}_c+\mathrm{FN}_c)}.$$ - Mean Intersection over Union (mIoU)

  $$\mathrm{mIoU} = \frac{1}{C}\sum_{c=1}^C \frac{\mathrm{TP}_c}{\mathrm{TP}_c + \mathrm{FP}_c + \mathrm{FN}_c}.$$

Documented baseline results facilitate direct comparison of new methods:

Method	PSNR (dB)	SSIM	mIoU (%)	PA (%)
EMRDM (CR only)	32.92	0.925	—	—
TDP-CR (CR only)	32.41	0.922	—	—
TDP-CR (Full)	33.10	0.925	61.8	81.7
Multi-stage EMRDM→SegFormer	—	—	59.2	79.8
CloudSeg (joint)	—	—	60.4	80.3

TDP-CR’s joint framework establishes new state-of-the-art (SOTA) results on both restoration and segmentation tasks in the presence of clouds (Zhang et al., 17 Jan 2026).

6. Access, Organization, and Usage Considerations

The dataset is officially referenced as “LuojiaSET.” Download instructions are typically provided via the original project page, institutional repositories, or the Open Science Framework (OSF). A standard directory architecture simplifies access for training and evaluation:

LuojiaSET-OSFCR/ ├── train/ │ ├── cloudy_optical/ (.tif) │ ├── sar/ (.tif) │ ├── clear_optical/ (.tif) │ └── labels/ (.tif or .png) ├── val/ (same structure) └── test/ (same structure)

Recommended preprocessing steps include per-band normalization to [0, 1] or z-score scaling, data augmentation strategies (e.g., random cropping to 128 × 128, flips), and stringent assurance of pixel-wise alignment across all modalities and annotations for model training.

7. Context and Applications

LuojiaSET-OSFCR underpins development and fair benchmarking of algorithms for multi-modal fusion, restoration, and semantic segmentation under cloud occlusion. Its careful construction enables methodologies such as prompt-guided SAR-optical fusion (e.g., the TDP-CR framework) to learn data-driven strategies for cloud-adaptive feature integration and supports rigorous quantification of downstream semantic utility, not just visual fidelity. A plausible implication is that the dataset’s diverse global coverage and high spatial precision facilitate robust generalization analyses in Earth observation, remote scene understanding, and disaster monitoring contexts (Zhang et al., 17 Jan 2026).