Deep Learning Based Super Resolution Mapping (SRM) from Medium Resolution Satellite Imageries
“Medium-resolution satellite images (like Sentinel-2 at 10m) cover large areas frequently but lack the fine details needed to see small buildings, narrow roads, or crop boundaries. We need to build an AI system that upscales these images to high-resolution (<4m) using generative models while ensuring the pixel colors and geographic locations remain scientifically accurate.”
Deconstructed Problem Blueprint
WHAT TO BUILD?
A robust deep learning-based Super-Resolution Mapping (SRM) web application and pipeline that ingests 10m medium-resolution satellite imagery, applies generative AI (like GANs or Diffusion models), and outputs sharp, high-resolution (<4m) analysis-ready raster images with quantified uncertainty maps.
WHO IS IT FOR?
National Technical Research Organisation (NTRO), urban planners, agricultural monitoring bodies, disaster response agencies, and defense analysts who require granular visual and spatial data from broad-scale satellite observations.
WHY DOES IT EXIST?
High-resolution commercial satellite imagery is extremely expensive and has low revisit frequencies. Medium-resolution data is free and abundant (e.g., Copernicus Open Access Hub), but its pixel blur prevents accurate local asset identification, small-scale change detection, and rapid disaster assessment.
INPUT → PROCESS → OUTPUT ARCHITECTURE PIPELINE
10m Sentinel-2 multi-spectral satellite imagery patches, paired high-resolution reference datasets (e.g., CartoSat or PlanetScope samples for training), geographic metadata (GeoTIFF / NetCDF formats), and user-selected region-of-interest coordinates.
Data pre-processing (normalization, cloud masking, band alignment), deep learning generative inference using CNN/GAN/Diffusion architectures trained on spatial textures and spectral consistency loss functions, and uncertainty estimation mapping.
Enhanced <4m Super-Resolved GeoTIFF images, side-by-side visual comparison dashboards, spectral integrity validation metrics, uncertainty confidence heatmaps, and downloadable analysis-ready layers for GIS tools.
CONSTRAINTS
- Must preserve spectral integrity and geographic coordinates (GeoTIFF spatial referencing) without hallucinating false infrastructure.
- Must explicitly quantify and display model inference uncertainty so analysts know which details are real vs. AI-generated.
- Must run efficiently on standard cloud GPU instances (like Google Colab / AWS T4/A10G) during 36-hour hackathon constraints.
JURY METRICS
- Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) against ground-truth high-res references.
- Spectral Angle Mapper (SAM) to ensure band-to-band spectral consistency.
- Inference processing latency per tile and UI/UX responsiveness of the web visualization dashboard.
- Accuracy improvement on downstream tasks like road extraction or crop boundary classification.
CORE BOTTLENECK
Preventing the generative model from hallucinating unrealistic structures (spectral/spatial hallucination) while maintaining true geospatial and radiometric fidelity during upscaling.
Why You Should Choose This PS (Pros)
- Direct alignment with national security, space tech, and strategic intelligence applications governed by NTRO.
- Clear, quantifiable mathematical evaluation metrics (PSNR, SSIM, SAM) that impress technical judges.
- High 'wow factor' through interactive side-by-side slider comparisons of blurry inputs vs. razor-sharp AI outputs.
Challenges & Risks to Watch (Cons)
- Training generative models from scratch within 36 hours is impossible; requires pre-trained weights or transfer learning.
- Handling large GeoTIFF multi-spectral raster files can crash standard web browsers if frontend rendering isn't optimized.
- Sourcing aligned pairs of 10m Sentinel and <4m high-res reference data requires careful dataset curation.
Recommended Tech Stack & Grand Finale Features
Killer Features to Impress the SIH Jury
View Official SIH 2026 Portal Raw Text▼
The expected solution is a robust AI-based super-resolution framework that can take medium-resolution satellite imagery as input, perform pre-processing, apply a trained generative model, and produce an enhanced spatial resolution image, suitable for analysis. The system should improve feature visibility, support better classification, change detection, crop monitoring, urban mapping, and disaster response. At the same time, it must clearly manage uncertainty because some reconstructed details are inferred by the model and not directly observed. Therefore, validation against high-resolution reference data is essential to ensure that the enhanced outputs are scientifically reliable and useful for real-world remote sensing applications.