ParamDev/SAT_ELITE_DATA

📦 SAT_ELITE_DATA

SAT_ELITE_DATA is a curated paired-image dataset intended for training and fine-tuning super-resolution models, particularly ESRGAN (Enhanced Super-Resolution Generative Adversarial Network). The dataset consists of aligned low-resolution and high-resolution image pairs, sourced from Sentinel-2 (10m/pixel) and NAIP (1m/pixel) satellite imagery, respectively.

---

🛰️ Dataset Overview

Domain	Source	Spatial Resolution	Description
Low-Res	Sentinel-2	10 meters/pixel	Multispectral satellite imagery (RGB subset)
High-Res	NAIP	1 meter/pixel	Aerial imagery from USDA's NAIP program

The dataset is organized to support paired image super-resolution tasks, where each Sentinel-2 patch (input) corresponds spatially and temporally to a high-resolution NAIP patch (target output).

---

📁 Directory Structure

The dataset is provided as two zip files:

SAT_ELITE_DATA/
├── train_set.zip   # Paired training images
├── val_set.zip     # Paired validation images

Each zip contains two folders:

train_set/
├── lr/             # Low-resolution Sentinel-2 images (input)
├── hr/             # High-resolution NAIP images (target)

The same structure applies to val_set/.

Each image pair shares the same filename (e.g., 12345.png in both lr/ and hr/), making it straightforward to load matching inputs and targets during model training.

---

📊 Image Specs

• Format: PNG (8-bit RGB)
• Patch Size (HR): 256×256
• Patch Size (LR): 32×32
• Normalization: No pre-applied normalization — users can apply their own preprocessing (e.g., mean/std normalization, scaling to [-1, 1]).

---

🎯 Use Case

This dataset is designed for:

• Super-resolution model training (e.g., ESRGAN, EDSR, SRGAN)
• Remote sensing image enhancement
• Satellite-to-aerial domain transfer learning
• Visual fidelity and detail recovery in earth observation pipelines

---

🔧 Suggested Training Pipeline (ESRGAN)

1. Preprocessing:

   • Resize Sentinel-2 images to match HR resolution using bicubic interpolation (if needed)
   • Normalize pixel values to [-1, 1]
   • Data augmentation (flip, rotate)

2. Training ESRGAN:

   • Generator: Residual-in-Residual Dense Blocks
   • Discriminator: PatchGAN-based
   • Losses: Content loss (VGG), adversarial loss, pixel-wise L1 loss

3. Evaluation:

   • PSNR, SSIM
   • Visual comparison with upsampled baseline (bicubic)

---

📜 License & Attribution

• Sentinel-2 data is provided by Copernicus Open Access Hub.
• NAIP imagery is provided by USDA Farm Service Agency.
• The dataset is intended for research and educational purposes only.

---

🤝 Citation

If you use this dataset, please cite the source repository or acknowledge:

@dataset{sat_elite_data,
  title        = {SAT_ELITE_DATA: Paired Sentinel-2 and NAIP Dataset for Super-Resolution},
  author       = {ParamDev},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/ParamDev/SAT_ELITE_DATA}},
  note         = {Paired low-res and high-res satellite imagery dataset}
}