Semantic Retrieval and Multi-Temporal Change Analysis of Satellite lmagery.
Ministry of defence (MoD) · Space Technology · Software
One of the best-specified problems in the whole set — free public data, a rigorous held-out evaluation, a striking offline demo — but it is won on false-alarm suppression and honest incremental indexing, so build those rather than a semantic search box bolted to a naive change detector.
Data: Public / organiser-generated imagery only (GeoTIFF / COG); no classified data
What it actually is
Satellite archives are huge, and analysts can only search them by metadata — coordinates, date, sensor — which means they must already know where to look. The ask is a system that makes an imagery archive searchable by meaning ('newly built structures near a river') and by change over time, running entirely on-premises and offline, while carefully not reporting seasonal or atmospheric variation as real change.
What to build
A retrieval-and-change platform over a satellite archive with six required capabilities: free-text and image-to-image semantic search over tiles using a remote-sensing foundation model, ranked and filterable by area, date and sensor; multi-temporal change analysis that classifies construction, clearance, water-extent and road development within a time window and estimates the earliest observation supporting each change; false-alarm suppression that treats season, illumination, cloud, haze, shadow and mis-registration as confounders via quality masks and confidence estimates, favouring precision over recall; embedding-based discovery to cluster visually similar sites so one location of interest surfaces others; an analyst review queue with before-and-after evidence, provenance and feedback-driven reranking; and scalable incremental vector indexing that ingests new imagery without a full rebuild — all fully offline with network disabled during evaluation.
Smallest thing that wins the room
With the network physically disabled, type 'newly built structures near a river' and get ranked tiles from the archive, then pick an area and time window and watch the system surface a genuine construction change with before-and-after imagery and a confidence score, while correctly NOT flagging a seasonally greener field as change.
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 69% of the 240 · #75 of 240 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
4/5A strong pick — the spec is excellent, the data is public and free, the held-out evaluation rewards teams that build honestly, and the MoD framing plus offline constraint thins the field, though the false-alarm suppression and the incremental-indexing-plus-held-out-evaluation bar are demanding and a naive change detector will fail exactly where judging focuses.
Feasibility
3/5Remote-sensing foundation models (Clay, Prithvi, SatCLIP-class), vector indexing and change-detection methods all exist and public imagery is freely available, but the combination — semantic retrieval plus reliable change classification plus false-alarm suppression plus incremental offline indexing, all evaluated against held-out labels — is a genuinely large integration, and the false-alarm suppression is the hard, unglamorous core.
Innovation scope
4/5How you fuse semantic retrieval with change reasoning, and especially how you separate real change from seasonal, atmospheric and registration confounders, is genuinely open — most public change-detection work does not handle this suppression well, and it is exactly what the description prizes.
Clarity
5/5The description is near-professional: six numbered capabilities, explicit example queries, named confounders to suppress, offline and provenance constraints, supported formats (GeoTIFF/COG), and a held-out evaluation protocol with reported latency, storage and build-time metrics.
Effort
MassiveSemantic indexing, change classification, confounder suppression, clustering discovery, an analyst workflow with audit trail, and incremental offline indexing is six substantial subsystems that must work together and run air-gapped.
Demo-ability
EasyNatural-language search returning ranked satellite tiles, then a before-and-after change surfaced with confidence, is immediately compelling and reads instantly — and the offline constraint makes pulling the network cable a dramatic proof.
In its favour
- Green flag: Public satellite imagery is free and abundant and the organiser supplies the evaluation set, so there is no data-sourcing risk
- Green flag: Held-out semantic queries and labelled change/no-change cases mean your evaluation is rigorous, standardised and unarguable
- Green flag: Natural-language search over an image archive plus a network-cable offline proof is a striking, memorable demo
- Green flag: The false-alarm suppression is real intellectual work that separates a serious submission from a naive pixel-differencing change detector
Against it
- Red flag: Seasonal, atmospheric, illumination and registration variation are the dominant source of false change, and suppressing them well is the actual hard problem the held-out evaluation will target
- Red flag: Incremental indexing without a full rebuild plus a reported build-time and storage footprint is a real engineering bar most teams underestimate
- Red flag: Everything must run offline with staged models, so a solution quietly leaning on a cloud API is disqualified
- Red flag: Semantic retrieval over remote-sensing tiles needs a genuine RS foundation model — a general CLIP will underperform on the exact queries the description gives
What you will be writing
- Remote-sensing foundation model (Clay / Prithvi / SatCLIP)
- Vector index (FAISS / Qdrant) incremental ingestion
- Multi-temporal change detection + classification
- Confounder suppression (quality masks, radiometric normalisation)
- GeoTIFF / COG geospatial provenance
- Offline / air-gapped deployment
- Satellite image retrieval
- Change detection
- Geospatial intelligence
Prior art to read before you start
semantic satellite image search · multi-temporal change analysis · false-alarm suppression in change detection
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.