SatQuery AI - An Interactive Vision-Language Assistant for Multimodal Remote Sensing Image Analysis through Text Queries
Indian Space Research Organisation(ISRO) · Space Technology · Software
The datasets are handed to you and the optical-SAR angle is genuinely valuable, but VLM fine-tuning is compute-heavy and the multisensor reasoning is what tends to get dropped — commit to the paired and cross-sensor cases, since a single-image VQA model is the crowded easy version.
Data: BigEarthNet (Sentinel-1 SAR + Sentinel-2 optical); VRSBench, RSVQA for evaluation
What it actually is
Non-experts cannot easily get answers out of satellite imagery because it requires knowing GIS workflows, sensors and task-specific models. The ask is a vision-language assistant fine-tuned for remote sensing that answers natural-language questions about satellite images, including harder cases needing multiple images — optical plus SAR, or before-and-after pairs for change detection.
What to build
A remote-sensing vision-language assistant that takes a natural-language query and one or more satellite images and answers it, domain-adapted so it understands sensor characteristics and remote-sensing terminology rather than relying on a general VLM, using BigEarthNet's co-registered Sentinel-1 SAR and Sentinel-2 optical data for the multisensor adaptation, handling single-image questions, multitemporal change queries across image pairs, and fused optical-SAR reasoning where SAR adds structural and all-weather information, possibly orchestrating specialised task models behind the language interface, evaluated on VRSBench and RSVQA.
Smallest thing that wins the room
Ask the assistant a change-detection question over a before-and-after Sentinel pair — how has built-up area changed here — and get a grounded answer, then ask a question answerable only by fusing the optical and SAR views to show the multisensor adaptation working.
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 70% of the 226 · #69 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
3/5The datasets and benchmarks are named and the multisensor angle is genuinely valuable, but VLM fine-tuning is compute-heavy and the scope is large, so a team that ships a single-image RS-VQA model has answered the easy part while the multisensor and multitemporal reasoning that distinguishes this is exactly what tends to get dropped.
Feasibility
3/5BigEarthNet, VRSBench and RSVQA are named public datasets and remote-sensing VLM fine-tuning is an active area with base models to build on, but fine-tuning a VLM to reliably handle multisensor and multitemporal reasoning is compute-heavy and genuinely hard, and doing it well within a hackathon is ambitious.
Innovation scope
4/5Multisensor optical-SAR fusion and multitemporal reasoning through a single language interface is genuinely open — most remote-sensing VLM work handles single optical images, so the paired and cross-sensor reasoning is where the real contribution lies.
Clarity
4/5The description names the datasets, the required capabilities including multisensor and multitemporal reasoning, the adaptation requirement and the evaluation benchmarks, so the deliverable is well defined despite being broad.
Effort
MassiveDomain-adapting a VLM, handling multisensor and multitemporal inputs, possibly orchestrating specialist models, and evaluating on multiple benchmarks is a large, compute-intensive undertaking.
Demo-ability
EasyAsking a plain-language question and getting a grounded answer about a satellite image is immediately compelling, and the change-detection and optical-SAR cases make strong demo beats.
In its favour
- Green flag: BigEarthNet, VRSBench and RSVQA are named public datasets, so both training and evaluation data are handed to you
- Green flag: The optical-SAR fusion and multitemporal angles are genuinely differentiating in a field where most RS-VLM work is single-image optical
- Green flag: Natural-language querying of satellite imagery is an immediately compelling and legible demo
- Green flag: Parameter-efficient fine-tuning like LoRA makes adapting a base VLM tractable without full retraining
Against it
- Red flag: VLM fine-tuning is genuinely compute-heavy, and without adequate GPU access you will ship something under-trained
- Red flag: The multisensor and multitemporal reasoning is the hard, distinguishing part and is exactly what teams drop under time pressure, leaving a single-image VQA model
- Red flag: A domain-adapted VLM can still hallucinate confident wrong answers about imagery, which for an ISRO judge is worse than an honest refusal
- Red flag: The scope spans adaptation, fusion, change reasoning and multi-benchmark evaluation, so shallow breadth is a real risk
What you will be writing
- Remote-sensing VLM fine-tuning (LLaVA / Qwen-VL base)
- BigEarthNet Sentinel-1 + Sentinel-2 co-registered data
- Optical-SAR multimodal fusion
- Multitemporal change reasoning
- VRSBench / RSVQA evaluation
- LoRA / parameter-efficient fine-tuning
- Remote sensing
- Vision-language models
- Multimodal AI
Prior art to read before you start
remote-sensing VQA · optical-SAR multimodal reasoning · multitemporal change interpretation via language
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.