Superpixel-Enhanced ESRGAN for Sentinel-2 Image Improvement Using Drone Imagery
Superpixel-based ESRGAN is a research and development project applying advanced image processing and AI to improve the spatial resolution of Sentinel-2 satellite data. By combining superpixel segmentation with Enhanced Super-Resolution GAN (ESRGAN) and training on high-resolution drone imagery, the project bridges the gap between medium-resolution (10–20 m) satellite data and fine-scale drone observations (~3–5 cm). The result is more accurate imagery for agriculture, land monitoring, and environmental research.
Purpose:
- Enhance the resolution of Sentinel-2 imagery while preserving multi-spectral integrity.
- Use superpixels to segment homogeneous regions and improve GAN training.
- Train ESRGAN with paired Sentinel-2 (low-res) and drone (high-res) datasets.
- Provide affordable, large-scale monitoring with near-drone-level accuracy.
- Support crop health monitoring, yield prediction, and climate research.
Who Uses It:
- Farmers & Agribusiness: Access near-drone-quality maps without frequent UAV flights.
- Remote Sensing Scientists: Develop and validate AI-enhanced datasets for precision research.
- Government & Agencies: Monitor crops, land use, and water resources with higher accuracy.
- AI/ML Researchers: Benchmark hybrid GAN + superpixel methods in geospatial imaging.
Key Capabilities:
- Superpixel Segmentation: Divides Sentinel-2 images into homogeneous regions (SLIC superpixels) to respect field and landscape boundaries.
- ESRGAN Super-Resolution: Learns mapping from Sentinel-2 to higher-resolution outputs using drone ground truth; restores sharper textures and boundaries.
- Cross-Resolution Training Dataset: Aligns Sentinel-2 tiles with drone orthomosaics; preserves NDVI and other vegetation indices.
- Evaluation Metrics: PSNR, SSIM, NDVI preservation, and crop classification accuracy.
Mobile / Web App:
- AI Map Viewer: Compare original Sentinel-2 images with GAN-enhanced versions.
- Analytics Layer: Generate vegetation indices from improved data.
- Exportable Data: Download GeoTIFFs for integration with GIS or farming platforms.
- Field Validation Mode: Overlay drone-collected ground truth for accuracy checks.
Typical Workflow:
- Collect: Acquire Sentinel-2 imagery and matching drone datasets.
- Segment: Apply superpixel segmentation to ensure boundary-aware training.
- Enhance: ESRGAN generates high-resolution imagery with preserved spectral detail.
- Validate: Compare with drone ground truth using PSNR, SSIM, and NDVI.
- Deploy: Provide enhanced maps to farmers, agencies, and researchers via web/mobile apps.
Governance & Data Integrity:
- Geo-referenced alignment between Sentinel-2 and drone imagery.
- Transparent evaluation metrics for scientific reproducibility.
- Role-based access for researchers, agencies, and agribusiness partners.
Outcomes & Value:
- More detailed satellite imagery for precision agriculture at scale.
- Cost-effective alternative to frequent UAV monitoring.
- Enhanced crop health, weed mapping, and disease detection.
- Stronger land-use and climate monitoring capabilities.