Deep Learning Based Super Resolution Mapping (SRM) from Medium Resolution Satellite Imageries
National Technical Research Organisation (NTRO) · Smart Education · Software
The demo is striking and the data is free, but this problem is won on proving your output reconstructs real detail rather than hallucinating it — build the geographic-fidelity evaluation, because that is exactly where an NTRO judge will probe.
What it actually is
Free satellite imagery at ten to thirty metres covers wide areas frequently but is too coarse to make out small buildings, narrow roads or field boundaries. The ask is a deep learning super-resolution model that sharpens medium-resolution satellite images into finer detail — not just prettier, but reconstructing genuine fine-scale features while keeping the geographic and spectral data honest.
What to build
A super-resolution framework taking medium-resolution satellite imagery, such as Sentinel-2 at ten metres, and reconstructing finer spatial detail using a GAN, diffusion model or deep network trained on paired medium and high-resolution imagery, with the critical constraint that outputs preserve geographic registration and spectral consistency rather than hallucinating plausible but false detail, evaluated both on standard image-quality metrics and on whether downstream tasks like building or road extraction genuinely improve on the enhanced imagery.
Smallest thing that wins the room
Take a real Sentinel-2 tile, super-resolve it, and show alongside the true high-resolution image of the same location that your output recovers the actual road and field-boundary structure rather than inventing convincing but wrong detail, with the quantitative fidelity metrics reported.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 39% of the 226 · #138 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: defence, intelligence and space bodies drew small fields.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
3/5Feasible with free imagery and the NTRO association plus space theme means fewer teams, but generative super-resolution has a serious credibility trap — hallucinated detail that looks sharp but is geographically false — and an NTRO judge will test exactly whether your output invents structure, so honesty about fidelity is what determines the outcome.
Feasibility
3/5Copernicus provides free Sentinel-2 imagery and the model architectures are well-documented, but building genuine paired training data — the same place at both resolutions, co-registered — is the real difficulty, since high-resolution commercial imagery for the pairs is expensive and public high-resolution coverage is patchy.
Innovation scope
3/5Satellite super-resolution is an active research area with established GAN and diffusion approaches, so your room is in the paired-data construction and in enforcing spectral and geographic fidelity rather than in the architecture.
Clarity
4/5The description is explicit that the goal is reconstructing useful fine-scale detail while preserving geographic and spectral consistency, not merely visual sharpening, which is a precise and unusually well-stated requirement.
Effort
HeavyAssembling co-registered paired training data, training a generative super-resolution model and building a proper fidelity evaluation is focused but demanding work, with data assembly the slowest part.
Demo-ability
EasyA side-by-side of blurry input, your super-resolved output and the true high-resolution reference is immediately compelling, and the fidelity comparison gives it rigour.
In its favour
- Green flag: Copernicus Data Space provides free Sentinel-2 imagery, which is exactly the medium-resolution input the description targets
- Green flag: The before-super-resolved-truth triptych is one of the most immediately convincing demos possible
- Green flag: Evaluating whether downstream extraction actually improves is a rigorous framing that proves the enhancement is real rather than cosmetic
- Green flag: The NTRO association and the fidelity-not-prettiness framing signal a serious, well-scoped problem
Against it
- Red flag: Generative super-resolution hallucinates plausible detail, and inventing a road that is not there is worse than blur — the description warns against exactly this and an NTRO judge will test for it
- Red flag: Genuine paired training data at both resolutions for the same location is expensive and hard to assemble, and mismatched pairs teach the model the wrong mapping
- Red flag: Standard image-quality metrics reward sharpness even when detail is fabricated, so you need geographic-fidelity evaluation or your numbers mislead
- Red flag: For an intelligence organisation, a super-resolution that fabricates structure could produce a false interpretation, which is a serious failure mode to acknowledge
What you will be writing
- Real-ESRGAN / SwinIR / diffusion super-resolution
- Sentinel-2 imagery via Copernicus Data Space
- Paired data construction with co-registration
- Spectral consistency loss terms
- Downstream task evaluation (building/road extraction)
- SSIM / LPIPS + geographic fidelity metrics
- Satellite super-resolution
- Generative deep learning
- Remote sensing
Prior art to read before you start
satellite image super-resolution · generative detail reconstruction with fidelity constraints · medium-to-high resolution enhancement
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.