Space Technology Accelerator · IIT Kanpur

Planetary Intelligence
for Real-World
Decisions

We build the foundation models, data schemas, and agentic systems that transform Earth Observation into now-cast intelligence, scenario foresight, and real-time alert infrastructure — for India.

Explore Research Agenda Collaborate With Us
60
Research Topics Published
9
Thematic Domains
5
Capability Levels
3
Agenda Documents
Mission

Where Satellite Science
Meets Operational Impact

EO-AI Labs is the applied research and commercialisation arm of the Space Technology Accelerator (STAr) at IIT Kanpur. We occupy the space between frontier EO science and the decisions that governments, industries, and communities need to make — now, not in five years.

Our thesis: the satellite data exists, the AI methods exist, the compute exists. What's missing is organised, well-framed research that connects them to decisions that matter — at the speed those decisions need to be made. That gap is what we exist to close.

We operate as three interlocking units: a Research Engine generating frontier methods and IP; a Product Studio translating research into deployable APIs and data products; and a Market Access Unit connecting science to government agencies, industry, and global research networks.

🛰️
EO Sensing
Multi-sensor fusion across SAR, optical, hyperspectral, LiDAR, and passive microwave — from raw DN to analysis-ready data.
🧠
Foundation Models
Self-supervised geospatial FM pretraining, domain adaptation, and task fine-tuning with physics-informed constraints.
⚡
Agentic Systems
Perceive–Reason–Act–Learn pipelines with typed EO tool libraries and multi-agent orchestration patterns.
📡
Now-Casting
Ensemble Kalman Filter data assimilation fusing live EO streams with NWP downscaling for sub-6hr products.
🔮
Scenario Planning
Probabilistic ensemble modelling from 10-day agricultural outlooks to 50-year Net Zero pathway scenarios.
🚨
Alert & DSS
P1–P4 severity-graded real-time alert systems with stakeholder routing, HITL governance, and immutable audit trails.
🏛️
Policy Translation
Bridging EO-AI outputs to regulatory frameworks, government schemes (NDMA, ICAR, MoEFCC), and carbon markets.
💼
Commercialisation
IP strategy, tiered FM licensing, open-core release, grant acquisition, and phased path from PoC to ARR.
Research Agenda

60 Topics Across
Three Interlocking Agendas

Each topic defines a verifiable knowledge gap, a concrete thesis/research plan, key datasets, and AI/ML methods — ready for M.Tech/PhD adoption, collaborative grants, or PoC development.

AGENDA 01 · 30 TOPICS
🇮🇳
India EO-AI Research Frontier
Thirty applied research topics curated for India's agro-climatic diversity, governance systems, and development priorities — from kharif crop mapping to GLOF risk in the Himalaya.
Agriculture & Food Disaster Mgmt Water Resources Urban Growth Forest & Biodiversity Air Quality Coastal & Marine Infrastructure Novel India Topics
AGENDA 02 · 30 TOPICS
🌍
Global EO-AI Research Compendium
Thirty planetary-scale topics at the absolute frontier — each with a traceable India research link. Wildfires, cryosphere dynamics, ocean systems, tropical deforestation, and generative AI tools.
Wildfire & Biomass Cryosphere Ocean & Marine Tropical Forest Foundation Models Generative AI Conflict Assessment Critical Minerals
AGENDA 03 · ARCHITECTURE
🏗️
Applied Research Roadmap
The operational blueprint: foundation model architecture, canonical data schemas, agentic frameworks, now-casting engines, scenario planning platforms, and real-time DSS — for EO-AI Labs.
FM Architecture Data Schemas Agentic AI Now-Casting Scenario Planning Alert Systems Commercialisation
THEME · AGRICULTURE
🌾
Agriculture & Food Security Intelligence
Kharif/Rabi crop type mapping at district scale using TSViT on LISS-IV + Sentinel-2, paddy irrigation water use estimation, PM-FASAL yield prediction via spatiotemporal GNN, and locust early warning.
Sentinel-2 LISS-IV TSViT FASAL SMAP SHAP
THEME · DISASTER
🌊
Disaster Resilience & Emergency Response
Near-real-time Brahmaputra/Ganga flood mapping (SAR + attention U-Net), Himalayan landslide susceptibility (InSAR + XGBoost + SHAP), cyclone damage grading (coherence change + Cartosat-3).
Sentinel-1 SAR InSAR Attention U-Net NDMA GPM IMERG
THEME · CRYOSPHERE
🧊
Cryosphere, Water & Glacial Risk
GLOF risk monitoring for Sikkim–Uttarakhand (InSAR moraine stability + Sentinel-2 lake expansion), HKH glacier mass balance (TanDEM-X + ICESat-2 geodetic), GRACE groundwater downscaling.
ICESat-2 GRACE-FO TanDEM-X PS-InSAR GLIMS
THEME · ATMOSPHERE
💨
Air Quality & Atmospheric Intelligence
PM2.5 estimation over Indian cities from TROPOMI NO2 + MERRA-2 meteorology, stubble burning FRP + emission inventory (Sentinel-3 SLSTR), global methane super-emitter detection (EMIT matched filter).
TROPOMI EMIT Sentinel-3 SLSTR CPCB MERRA-2
THEME · FOREST
🌿
Forest, Biodiversity & Carbon
Northeast jhum cycle detection (BFAST + SAR-optical fusion), tiger corridor connectivity modelling (Circuitscape + GNN), global mangrove carbon MRV (GEDI + GMW v3.0), soil organic carbon for CCTS.
GEDI LiDAR BFAST GMW v3.0 Circuitscape REDD+
THEME · OCEAN & COAST
🐠
Coastal, Marine & Ocean Systems
Shoreline change detection across India's 7,500 km coast (DSAS + U-Net), HAB forecasting in Arabian Sea/Bay of Bengal (Sentinel-3 OLCI + LSTM), global coral bleaching probability (Sentinel-2 + CoralTemp).
Sentinel-3 OLCI DSAS CoralTemp LSTM CMLRE
View Full Research Agendas →
Technical Capabilities

Four-Tier Intelligence Stack

From raw satellite acquisition to autonomous decision support — a modular, upgradeable architecture built on open standards and physically grounded AI.

Tier 1 — Sensing Foundation Model
Self-supervised geospatial encoder pretrained on 100M+ global EO patches. Architecture: Vision Transformer (ViT-L) with spectral token embeddings and masked autoencoder (MAE) pretraining objective. Produces 768-dim patch embeddings compatible with all downstream fine-tuning tasks.
Architecture
ViT-L with custom spectral tokenizer supporting 2–200 band inputs across SAR, optical, hyperspectral, and LiDAR.
Pretraining
MAE objective on 100M+ Sentinel-1/2, Landsat, MODIS, and ISRO ResourceSat patches; multi-temporal context encoding.
India Adaptation
Domain-specific continued pretraining on monsoon-season mosaics, mixed smallholder landscapes, and urban-rural transitions.
Benchmarks
Evaluated on GEO-Bench, BigEarthNet, Sen12MS, and custom Indian EO benchmark suite with 5-zone stratification.
Tier 2 — Fusion Foundation Model
Cross-modal attention transformer that aligns representations across sensors, timesteps, and spatial resolutions in a unified latent space. Handles systematic missing data (cloud cover, orbit gaps) through learned imputation.
Cross-Modal Attention
Learnable positional encodings for spatial + temporal + spectral axes; SAR-optical alignment without paired labels.
Resolution Handling
Fuses 0.25m Cartosat-3, 10m Sentinel-2, 30m Landsat, 250m MODIS, and 1km atmospheric products in a single pipeline.
Cloud Removal
Conditional latent diffusion model for monsoon cloud removal, conditioned on co-registered SAR and adjacent clear-sky dates.
Uncertainty Propagation
Aleatoric + epistemic uncertainty decomposition propagated through all fusion outputs to downstream tasks.
Tier 3 — Reasoning Foundation Model
Domain-tuned vision-language model for geospatial QA, change event narration, scenario interpretation, and policy brief generation from model outputs. Bridges satellite analytics and human-readable intelligence products.
Architecture
Multimodal LLM with EO-specific visual adapter; fine-tuned on scientific literature, STAC item descriptions, and analyst annotation datasets.
Geospatial QA
Natural language query to spatial analysis: "What changed in the Yamuna floodplain between July 15 and August 3?"
Scenario Narration
Converts ensemble model output statistics into structured scenario briefs: exec summary, sector impacts, uncertainty statement.
Policy Translation
Auto-generates NDMA-compatible situation reports, CPCB compliance summaries, and FASAL-style pre-harvest advisories.
Tier 4 — Action Foundation Model
ReAct-style agentic model that translates inference outputs into structured action recommendations, alert payloads, and workflow triggers — with physical constraint checking and human authority gates at every P1 decision point.
Tool Library
8 typed tool categories: EO query, inference runners, geospatial analysis, climate/weather APIs, exposure/risk, alert routing, synthesis, and audit logging.
Multi-Agent Patterns
Supervisor–Worker, Pipeline Chain, Debate & Consensus, Human-in-the-Loop, and Continuous Background Monitor orchestration.
HITL Gate
All P1 (life-safety) alerts require human confirmation before external notification. Hardcoded at Act Layer; cannot be overridden.
RLHF Alignment
Monthly RLHF cycle from analyst confirmations, false positive reports, and outcome observations. Threshold calibration updated weekly.
Canonical Data Schema Stack
Five-layer schema stack from raw ingestion to decision payloads — STAC-compatible, OGC-compliant, and JSON-LD serialisable. The interoperability skeleton that makes intelligence portable across organisations and missions.
Layer 0–1: Ingestion → ARD
STAC 1.1 + EO/SAR Extension. Surface reflectance, sigma-naught, quality band, per-pixel uncertainty. Harmonised to common 10m UTM grid.
Layer 2: Feature Store
GeoParquet + pgvector. Spectral indices, GLCM texture, temporal statistics, 768-dim FM embeddings. H3 resolution-8 spatial index.
Layer 3: Inference Records
Pydantic v2 validated. Per-scene: model_id, confidence, uncertainty_band, spatial_mask, anomaly_score, change_flag, lineage_hash.
Layer 4: Decision Payloads
Structured alerts: event_type taxonomy, P1–P4 severity, stakeholder_routes, recommended_actions, expiry, HITL flag, audit_trail_url.
Agentic Framework
Full Perceive–Reason–Act–Learn loop on Apache Kafka + Flink stream infrastructure. Transforms passive model inference into active decision support with configurable trigger rules, multi-step reasoning chains, and feedback-driven learning.
Perceive Layer
Kafka/Flink stream processor ingesting EO inference outputs, IoT feeds, weather model outputs, and GDACS/USGS event feeds. Outputs prioritised event queue.
Reason Layer
ReAct-pattern LLM chain querying feature store, invoking EO tools, and formulating situation assessment with chain-of-thought + physics constraint checking.
Act Layer
Routes Decision Payloads to REST webhooks, SMS/WABA, GIS platform updates, SCADA OPC-UA alarm bus, NDEMS portal. All actions logged immutably.
Learn Layer
Collects analyst feedback, outcome observations, false positive reports → RLHF pipeline. Weekly threshold calibration. Monthly performance governance review.
Now-Casting Engine
5-module architecture producing continuously updated, gapless planetary state estimates with calibrated uncertainty — filling the gap between the last satellite overpass and the present moment.
Observation Fusion
24–48hr multi-sensor composite: SAR (cloud-penetrating) + optical (where available) + passive microwave + VIIRS NTL. Weighted by quality flags.
Temporal Dynamics
LSTM / Temporal Transformer trained on 5+ years of feature time series per land cover class. Provides prior estimate for data assimilation.
Data Assimilation
Ensemble Kalman Filter merging observation fusion with temporal prior. Full covariance matrix. Correct uncertainty propagation.
Products & Validation
8 now-cast products. CRPS, FSS, POD/FAR validation. SLA: <2hr SAR-based, <6hr optical. OGC API Coverages endpoint.
Scenario Planning Platform
6-module platform extending now-cast state into probabilistic futures under alternative climate, policy, and operational assumptions. From 10-day agricultural outlooks to 50-year Net Zero pathways.
Scenario Parameter Engine
NLP-to-JSON scenario specification: climate forcing (SSP1.9–SSP5.8), policy levers, extreme event forcings, operational decisions.
Ensemble Model Runner
N=50–500 ensemble members on cloud HPC. CMIP6-downscaled statistical emulators for long-term scenarios. CF-compliant NetCDF output.
Impact Assessment
Sector-specific surrogates: crop model (DSSAT/APSIM), flood damage curves (HAZUS-calibrated), carbon flux, infrastructure fragility.
Counterfactual Generator
DoWhy causal graph framework for physically consistent counterfactuals. OOD warning when scenarios exceed training distribution.
Real-Time Alert & Decision Support
End-to-end alert pipeline from stream detection to stakeholder action — with explainability-first design, graceful degradation, sovereign data compliance, and an immutable 7-year audit trail.
Alert Detection
Three detection modes: threshold crossing, rate-of-change (2σ anomaly), and compound event (multi-hazard co-occurrence). DBSCAN deduplication.
P1–P4 Severity
Multi-factor scoring: population exposure × asset value × confidence × time-to-impact. P1 requires HITL confirmation before external dispatch.
Stakeholder Routing
REST webhooks, SMS/WABA, ESRI ArcGIS push, Slack/Teams, SCADA OPC-UA alarm bus, NDEMS/IEMS portal integration.
DSS Principles
6 design principles: Explainability First, Graceful Degradation, Human Authority, Equity Auditing, Immutable Audit Trail, Sovereign Data Compliance.
Maturity Framework

Five-Level Capability Maturity Model

Organisations can self-assess against this framework and use the roadmap to advance systematically. Skipping levels creates technical debt that compounds at production scale.

Level
Name
Characteristics
Gate Condition
L1
Data Ingestion
Automated, reproducible EO data pipelines; basic preprocessing and tile management across 3+ sensor streams at daily cadence.
Automated ingestion live; STAC catalog query <500ms; 10M+ tiles indexed.
L2
Feature Intelligence
Derived indices, change detection, anomaly scoring from single sensors. Task-specific fine-tuned models per domain.
Per-scene inference <30 min; F1 >0.80 on benchmark datasets; Feature Store <100ms query.
L3
Fusion & Now-Casting
Multi-sensor fusion, EnKF data assimilation, NWP downscaling, near-real-time products. First operational now-cast pipeline.
Fused product latency <6hr; CRPS skill >0.15 vs. climatology; 3+ stakeholder pilots onboarded.
L4
Scenario Planning
Ensemble modelling, counterfactual analysis, CMIP6-downscaled scenario generation with quantified uncertainty bounds.
10+ scenario variants in <4hr; ensemble runtime <4hr for 100 members; OOD flagging operational.
L5
Autonomous DSS
Full agentic reasoning, multi-step planning, NLP interfaces, stakeholder routing, HITL governance, immutable audit trail.
Alert-to-action cycle <15 min; POD >0.90 for P1; FAR <0.20; HITL audit trail complete.
Implementation Roadmap

From Research to Revenue:
A 5-Year Phased Plan

Each phase has quantified gate conditions — technical, commercial, and institutional milestones that must be met before advancing.

Phase 0
Q1–Q2 Yr 1
Foundation
  • Cloud HPC & MLOps infrastructure
  • STAC data lake + GeoParquet store
  • Sensing FM pretraining begins
  • Schema stack v1.0 published
  • 2 PoC partner agreements signed
  • IP disclosure framework established
10M+ tiles ingested · STAC <500ms · FM loss converging
Phase 1
Q3–Q4 Yr 1
Intelligence
  • Fusion FM trained
  • Reasoning FM fine-tuned
  • Feature store operational
  • First now-casting product live
  • Basic alert engine (P3–P4)
  • 3 peer-reviewed publications
  • Provisional patent filed
MRR >₹5L or grant >₹1Cr · POD >0.85
Phase 2
Q1–Q2 Yr 2
Operationalise
  • Action FM (agentic layer)
  • P1–P2 alerts with HITL
  • Scenario planning MVP
  • DSS dashboard v1.0
  • Industrial AI integration
  • ISO 27001 certification
ARR >₹2Cr · FAR <0.20 · 2+ enterprise clients
Phase 3
Q3 Yr 2 – Yr 3
Scale
  • 5+ domain verticals live
  • FM licensing programme
  • Multi-country deployment
  • Series A / Govt. contract
  • ESA BIC / ADB partnership
ARR >₹5Cr · FM licensed in 3+ products
Phase 4
Yr 3 – Yr 5
Lead
  • Global EO-AI Lab network
  • Sovereign FM deployments
  • Net Zero MRV platform
  • L5 autonomous monitoring
  • Patent portfolio >10 active
  • Series B / strategic exit
ARR >₹15Cr · 3+ countries · Top-3 EO-AI Lab ranking
Research Documents

Published Research Agendas

Three structured documents defining the EO-AI Labs research programme — available for academic collaboration, thesis adoption, and partnership development.

Research Agenda · India Focus
EO-AI Research Frontiers for India: 30 Applied Topics
Thirty research topics curated for India's specific geographic, ecological, and socioeconomic context. Each topic includes a defined research frontier, full thesis/research plan, key datasets (Bhuvan, ISRO ResourceSat, Sentinel, MODIS, IMD), and AI/ML methods. Spans agriculture, urban monitoring, disaster resilience, water resources, forest ecology, coastal environments, air quality, infrastructure, and novel India-specific challenges including sacred grove delineation, Kumbh Mela crowd estimation, and GLOF risk monitoring.
30 topics · 9 domains
View India Agenda
Research Agenda · Global Focus
Global EO-AI Research Compendium: 30 Planetary Topics
Thirty planetary-scale topics designed as a strategic complement to the India volume. Every topic includes an explicit "India Link" row mapping the global method to a specific Indian research opportunity. Covers wildfire spread physics-informed neural networks, cryosphere Bayesian fusion, ocean hyperspectral plastic detection, tropical deforestation sub-hectare detection, and foundational AI tools: EO-CLIP, cloud removal diffusion models, and self-supervised change detection without labels.
30 topics · 10 domains
View Global Agenda
Architecture Roadmap
Applied Research Roadmap: FM · Schemas · Agentic Systems
The operational blueprint for production-grade EO-AI intelligence infrastructure. Defines: four-tier foundation model architecture with pretraining protocols and Model Registry schema; five-layer canonical data schema stack (STAC, GeoParquet, OGC, CF, FAIR); agentic Perceive–Reason–Act–Learn framework with typed tool library; now-casting engine with EnKF data assimilation; scenario planning platform with counterfactual generator; and real-time alert DSS with P1–P4 severity taxonomy and six design principles.
9 parts · 5-year roadmap
View Roadmap
Leadership

Research Team

EO-AI Labs brings together researchers spanning remote sensing, machine learning, atmospheric science, environmental economics, and policy — united by a commitment to research that reaches the decisions it's meant to inform.

Dr. Abhay Gupta
Dr. Abhay Gupta
Head, EO-AI Labs · Coordinator EO-AI & Commercialisation
STAr IITK · Adjunct Professor, SPASE · Director, Efficient World, Vancouver
Economist-engineer building evidence-driven AI grounded in observable reality. PhD Economics (UBC), B.Tech Chemical Engineering (IIT Kanpur). 20+ years spanning OECD Paris, European Investment Bank, House of Commons, and founding AI companies — now leading EO-AI research and commercialisation at IIT Kanpur.
Full Profile LinkedIn Email Website
RS
Research Scientists
EO-AI Core Research
STAr IITK · EO-AI Labs
Our research scientists work across the sensing-to-decision pipeline — from SAR-optical fusion and foundation model adaptation to scenario ensemble modelling and geospatial agentic systems. We are actively recruiting PhD researchers and postdoctoral fellows with expertise in remote sensing, geospatial ML, atmospheric science, and data assimilation.
Open Positions Apply
CE
Collaborating Faculty
Academic Partners
IIT Kanpur & Partner Institutions
EO-AI Labs operates in close collaboration with IIT Kanpur faculty across the Departments of Civil Engineering, Aerospace Engineering, Computer Science, and Earth Sciences — as well as external academic partners at ESA, NASA, JAXA, CSIRO, and leading Indian research institutions including NRSC, IMD, and ICAR.
Collaborate STAr IITK
Collaboration

We Are Seeking Deep
Collaborations — Not Citations

EO-AI Labs is at the beginning of a multi-year research and commercialisation programme. We are looking for partners who want to build something together — not just co-author a paper. If you see the same intelligence gap we see, let's talk.

🎓
Academic Research Partners
Joint PhD supervision, collaborative grant applications (DST-SERB, ISRO RESPOND, ESA BIC, USAID DIV), co-authored publications, and shared dataset creation across remote sensing, ML, hydrology, agriculture, and ecology.
🏛️
Government & Policy Partners
ISRO, NRSC, NDMA, IMD, ICAR, MoEFCC, CPCB, CWC, FSI, and State DMAs. We build research-to-operations pipelines aligned with national schemes including FASAL, PMGSY, PMFBY, and national environmental monitoring mandates.
🏭
Industry & Enterprise
Co-development of EO-AI intelligence products, PoC to production pipeline buildout, climate risk and supply chain intelligence, ESG disclosure support (BRSR, CSRD), and integration with enterprise digital twin platforms.
💹
Financial & Investment Partners
Commodity trading signal development, agricultural risk pricing, infrastructure climate risk assessment, carbon credit MRV, and alternative data intelligence products for asset managers and development finance institutions.
🛰️
International EO Missions
ESA, NASA, JAXA, CSIRO, DLR. We are actively pursuing access to Copernicus DIAS, Planet NICFI, MAXAR ARD, and EMIT hyperspectral data for research programme execution and joint instrument validation.
🎯
PhD & Postdoc Researchers
Seeking researchers with expertise in geospatial ML, data assimilation, SAR processing, atmospheric remote sensing, agentic AI systems, or research commercialisation. IIT Kanpur and co-supervised positions available.
Get In Touch
AG
Dr. Abhay Gupta
Head, EO-AI Labs · STAr IITK
✉️
abhaygupta@iitk.ac.in
🔗
linkedin.com/in/gabhayg
🏫
STAr IITK · IIT Kanpur

For research collaboration, document requests, grant partnership discussions, or PhD/postdoc enquiries — please reach out directly. We respond to all substantive enquiries within 3 working days.