Loading Abhiram Reddy Palle

Abhi/ 26
Your time on this site
→All workResearch note 03 / 04

Research · Remote sensing

CERTUS-S2

When can you trust a pixel a model made up?

Type
Research · Remote sensing
Role
[Add role]
Year
[Add year]
Status
Research

(01)Motivation

Sentinel-2 gives free, global, multispectral imagery at 10, 20 and 60 m resolution — detailed enough for a lot, not enough for everything. Super-resolution fills the gap, but a sharper image isn't automatically a truer one. Without a way to validate outputs, a confident-looking image can quietly carry a wrong answer into a real decision.

(02)Abstract

Super-resolution models can make Sentinel-2 imagery look sharper than the sensor actually captured. That's useful right up until a decision depends on detail the model invented. CERTUS-S2 is research into trust and decision certification for super-resolved Sentinel-2 imagery: instead of asking whether an output looks good, it asks whether a decision made from it can be relied on.

Research question

When a super-resolved Sentinel-2 image feeds a real decision, can we certify when that decision is trustworthy — and flag it when it isn't?

  • Satellite imagery
  • Sentinel-2
  • Super-resolution
  • Trust
  • Validation
  • Decision certification
  • STAC
  • Microsoft Planetary Computer
  • Python

(03)Approach

Treat trust as something to validate, not assume: check super-resolved outputs and certify the decisions made from them, instead of scoring images on how sharp they look.

Written in Python. Imagery comes from Microsoft Planetary Computer, searched and loaded through its STAC API, so the pipeline starts from real, well-catalogued Sentinel-2 scenes — with testing and validation built into the workflow.

Pipeline
  1. 01→SearchSTAC query on Planetary Computer
  2. 02→ScenesSentinel-2 imagery
  3. 03→Super-resolveModel output
  4. 04→ValidateTrust checks
  5. 05■CertifyTrusted — or flagged
Language
Python
Data
Sentinel-2/Microsoft Planetary Computer/STAC
Method
Super-resolution/Validation/Decision certification

(04)Open problems

  1. 01

    Sharp isn't the same as right

    Standard image-quality scores reward outputs that look convincing. Trust needs a different yardstick.

  2. 02

    Certify the decision, not the picture

    The same image can be fine for one decision and dangerous for another, so trust has to be tied to what the output is used for.

  3. 03

    Knowing when to say no

    A useful system has to be able to flag an output as untrustworthy instead of always producing an answer.

(05)Status

[ADD STATUS & FINDINGS — paper, preprint, results]

Live[Add link]Code[Add link]Video[Add link]

(06)What's next

First versions are rarely the final versions. This one won't be either.

[Add what you'd improve next]

(07)Under the hood

Reconstruct → certify → abstainFrom the repo's claims ledger
Conformal risk control · site-disjoint splitsSentinel-2

The same picture the film runs — an illustration of how CERTUS-S2 thinks, not a screenshot.

(08)Figures

Fig. 1 — Input vs. super-resolvedFig 01
Fig. 2 — Trust mapFig 02
Fig. 3 — PipelineFig 03
Next — 04 / 18SIH1775 · AI / SecuritySocialGuard→Spotting the accounts that aren't who they say they are.