SADER

Official checkpoints for SADER: Structure-Aware Diffusion Framework with Deterministic Resampling for Multi-Temporal Remote Sensing Cloud Removal.

SADER combines a multi-temporal conditional denoising network, a cloud-aware training objective, and deterministic resampling for cloud-free reconstruction from temporally aligned Sentinel-2 observations.

• Paper: arXiv:2602.00536
• Code and inference instructions: github.com/zyfzs0/SADER
• Issues: github.com/zyfzs0/SADER/issues

Released checkpoints

The two primary SADER checkpoints correspond to the benchmark settings reported in the paper.

Directory	Dataset / setting	Model output	Main checkpoint	MAE checkpoint used by deterministic resampling
newdesign/	Sen2_MTC_New; three temporal RGB+NIR observations	RGB	newdesign/final_ckpt/last.ckpt	pretrained/imputation_mask0.5_MAE_eurosat_rgb_dm192_bs16_lr0.001_Exp_0/checkpoint.pth
tsdesign/	SEN12MS-CR-TS(EA); three temporal Sentinel-2 observations with 13 bands	13 Sentinel-2 bands	tsdesign/final_ckpt/last.ckpt	pretrained/imputation_mask0.5_MAE_eurosat_ms_dm192_bs16_lr0.001_Exp_0/checkpoint.pth

The repository also contains:

• tsmask/: an additional mask-conditioned SEN12MS-CR-TS(EA) experimental checkpoint and its run snapshots.
• pretrained/: five EuroSAT MAE checkpoints used in the final configurations or auxiliary pretraining experiments.
• */configs/: the original project and PyTorch Lightning configuration snapshots.
• */testtube/: saved hyperparameters and training metric logs.

The configuration snapshots preserve the paths used for the original runs. Replace all dataset paths and checkpoint paths with locations on your system before running the code.

Download

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download the complete release (approximately 2.7 GB):

hf download zyfz0um/SADER --local-dir ./sader_weights

Download only the Sen2_MTC_New model and its MAE checkpoint:

hf download zyfz0um/SADER \
  newdesign/final_ckpt/last.ckpt \
  pretrained/imputation_mask0.5_MAE_eurosat_rgb_dm192_bs16_lr0.001_Exp_0/checkpoint.pth \
  --local-dir ./sader_weights

Download only the SEN12MS-CR-TS(EA) model and its MAE checkpoint:

hf download zyfz0um/SADER \
  tsdesign/final_ckpt/last.ckpt \
  pretrained/imputation_mask0.5_MAE_eurosat_ms_dm192_bs16_lr0.001_Exp_0/checkpoint.pth \
  --local-dir ./sader_weights

Use with the official code

Clone the code and initialize the MAE submodule:

git clone --recursive https://github.com/zyfzs0/SADER.git
cd SADER
pip install -r requirements.txt

For the closest match to the released checkpoints, use the environment specified by the official repository (environment.yaml, including PyTorch 2.3).

Download the required checkpoint pair into the code repository, then edit the corresponding YAML file:

Dataset	Official YAML file
Sen2_MTC_New	configs/example_training/sen2_mtc_new_TemporalAdm_repaint_mae.yaml
SEN12MS-CR-TS(EA)	configs/example_training/sentinelts_new_repaint_mae.yaml

Update the following entries in the YAML file:

1. model.ckpt_path: path to the SADER last.ckpt file.
2. sampler_config.params.mae_config.params.configs.model_path: path to the corresponding MAE checkpoint.pth file.
3. Dataset paths under data.params.

Run evaluation:

python main.py \
  --base configs/example_training/<CONFIG_FILE>.yaml \
  --enable_tf32 \
  -t false

Run prediction and save reconstructed images:

python main.py \
  --base configs/example_training/<CONFIG_FILE>.yaml \
  --enable_tf32 \
  -t false \
  --no-test true \
  --predict true

See the official GitHub README for dataset preparation, environment setup, training, testing, and prediction details.

Input and output settings

Dataset	Temporal input	Auxiliary information	Reconstruction target
Sen2_MTC_New	Three aligned observations	NIR is used together with RGB as conditional information	RGB image
SEN12MS-CR-TS(EA)	Three aligned Sentinel-2 observations	No additional auxiliary channel	All 13 Sentinel-2 bands

The released configurations use 256 × 256 image patches and T = 3 temporal observations.

Reported results

The following results use the evaluation protocol reported in the paper.

Dataset	PSNR ↑	SSIM ↑	LPIPS ↓	SAM (°) ↓
Sen2_MTC_New	20.941	0.729	0.228	—
SEN12MS-CR-TS(EA)	31.234	0.937	—	5.885

Intended use and limitations

These checkpoints are intended for research and reproducibility in multi-temporal remote sensing cloud removal. Performance may change for sensors, spectral configurations, geographic regions, temporal lengths, or preprocessing pipelines that differ from the two reported benchmarks. The full MTCDN is computationally demanding despite its compact parameter count.

The .ckpt and .pth files use PyTorch serialization, which may rely on pickle. Load checkpoint files only from a source you trust.

License

This repository currently declares the license as other; no separate license for the released weights has been specified. Please contact the authors with licensing questions.

Citation

@article{zhang2026sader,
  title={SADER: Structure-Aware Diffusion Framework with Deterministic Resampling for Multi-Temporal Remote Sensing Cloud Removal},
  author={Zhang, Yifan and Chen, Qian and Liu, Yi and Li, Wengen and Guan, Jihong},
  journal={arXiv preprint arXiv:2602.00536},
  year={2026}
}

Contact

For questions, please open an issue in the SADER GitHub repository or contact yifanzhg@umich.edu / 2250951@tongji.edu.cn.