LERSGAN_RS_PAPER

Paper-structured LersGAN dataset rebuild for low-light remote sensing image enhancement.

• Paper: LersGAN: A GAN-Based Model for Low-Light Remote Sensing Image Enhancement
• GitHub: https://github.com/TianqiLi11/LersGAN
• Dataset: https://huggingface.co/datasets/cod-tdq/lersgan

Dataset Description

This dataset follows the dataset protocol described in the paper's dataset section. The training and validation splits are unpaired low-light and normal-light image pools. The test split contains paired normal-light remote-sensing images and two generated low-light sets:

• RSDark1: standard low-light degradation with brightness scaling alpha in [0.2, 0.4] plus Gaussian sensor noise.
• RSDark2: extreme low-light degradation with brightness scaling alpha in [0.05, 0.2] plus Poisson-style shot noise and Gaussian read noise.

The paper reports aggregate selected counts, not exact per-source selected file lists. This rebuild records the available local reconstruction and is intended for LersGAN reproduction experiments.

Structure

.
├── train/
│   ├── low_light/
│   └── normal_light/
├── val/
│   ├── low_light/
│   └── normal_light/
└── test/
    ├── normal/
    ├── RSDark1/
    └── RSDark2/

Counts

• Train: 800 low-light images and 900 normal-light images.
• Validation: 100 low-light images and 100 normal-light images.
• Test: 500 normal-light images, 500 RSDark1 images, and 500 RSDark2 images.

Citation

@article{li2025lersgan,
  title = {LersGAN: A GAN-Based Model for Low-Light Remote Sensing Image Enhancement},
  author = {Li, TianQi},
  journal = {IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year = {2025},
  doi = {10.1109/JSTARS.2025.3608696},
  publisher = {IEEE}
}