EO-AI Research Frontiers
for India

Global foundation models (Prithvi, SatMAE) are pretrained on Western-centric imagery archives. India's diverse agro-climatic zones, monsoon seasonality, and mixed urban-rural transitions create distribution shifts that degrade off-the-shelf model performance.

Develop and benchmark an India-centric self-supervised vision transformer pretrained on multi-year Sentinel-1/2 and ResourceSat-2 composites. Evaluate zero-shot and fine-tuned transfer across five ecological zones (Thar Desert, Indo-Gangetic Plain, Western Ghats, Deccan Plateau, Northeast Highlands). Key contribution: a publicly released checkpoint + curated Indian EO benchmark suite.

Bhuvan, ISRO ResourceSat-2, Sentinel Hub India tiles, LISS-IV archives

India operates a large fleet of domestic satellites (Cartosat, RISAT, EOS series) whose data is underutilized in AI pipelines due to domain gaps with publicly benchmarked sensors.

Design a contrastive domain adaptation framework to align feature representations between Sentinel-2 and Cartosat-3 / EOS-04 SAR imagery. Apply to crop mapping in Haryana and flood detection in Assam. Quantify label-efficiency gains using Indian Geodata Portal ground truth.

Cartosat-3, EOS-04, RISAT-1A, Sentinel-1/2

India's crop calendars are highly variable across states. Paddy in Punjab vs. West Bengal vs. Odisha differs by 6–8 weeks. National crop maps remain coarse (>30 m) and lag by months.

Build a multi-temporal transformer (TSViT or SITS-Former) trained on LISS-IV + Sentinel-2 time series for district-level crop type mapping across the Indo-Gangetic Plain. Validate against FASAL crop statistics and Kisan Suvidha ground truth. Target: 10 m resolution, <2-week latency.

LISS-IV, Sentinel-2, FASAL data, Crop Insurance (PMFBY) records

Paddy cultivation accounts for ~40% of India's agricultural water use. Remote sensing-based actual evapotranspiration (ETa) products lack sub-field accuracy for India's small-holder fragmented fields.

Integrate Sentinel-1 SAR backscatter with MODIS-based SEBAL/METRIC ETa models to estimate irrigation water consumption at field-parcel level across the Cauvery and Krishna basins. Compare against district-wise canal withdrawal records.

Sentinel-1, MODIS MOD16, Canal withdrawal data, Bhuvan field boundaries

India suffered its worst locust outbreak in decades in 2020. Current FAO early-warning systems rely on ground scouts; satellite-based vegetation anomaly detection is nascent.

Develop an anomaly detection pipeline combining NDVI time-series deviation (Sentinel-2), wind trajectory modeling (ERA5), and soil moisture (SMAP) to predict high-risk locust breeding zones in Rajasthan and Gujarat 2–4 weeks in advance. Validate against 2019–2022 outbreak records from the Locust Warning Organisation.

Sentinel-2 NDVI, ERA5 wind, SMAP, LWO India outbreak records

India's FASAL scheme provides pre-harvest crop forecasts using linear regression on coarse NDVI. Deep learning with weather co-variates can improve accuracy significantly.

Train a spatiotemporal graph neural network fusing Sentinel-2 phenology features, IMD gridded rainfall, and soil organic carbon maps to predict district-wise wheat and rice yields. Benchmark against FASAL historical forecasts (2010–2023). Assess explainability via SHAP for policy uptake.

Sentinel-2, IMD gridded rainfall, NBSS&LUP soil data, FASAL records

India has ~65 million slum dwellers across 2,613 towns. Official slum boundaries are outdated. Automated detection must handle highly heterogeneous roofing materials and densities.

Combine Cartosat-3 pan-sharpened imagery (0.25 m) with WorldView spectral features and OSM building footprints to train a semi-supervised instance segmentation model for slum delineation across Mumbai, Chennai, and Kolkata. Validate against 2011 Census slum boundaries and DCHB data.

Cartosat-3, WorldView-3, OpenStreetMap, Census 2011 slum data

India's rapidly expanding tier-2 cities face intensifying UHI effects. Landsat-based LST products miss intra-urban variability relevant for ward-level planning.

Fuse ECOSTRESS (70 m, diurnal) with Sentinel-2 LULC classification and built-up density indices to model UHI intensity at ward level for 10 Indian cities. Analyze cooling contributions of urban forests, water bodies, and albedo variation.

ECOSTRESS, Landsat-8/9, Sentinel-2, Census ward boundaries

India's PMGSY tracks ~700,000 km of rural roads, but verification of construction quality and completeness using satellite data is not automated.

Develop a topology-aware road extraction network (D-LinkNet or RNGDet) using Cartosat-2S imagery to validate PMGSY road completeness in Odisha, Jharkhand, and Bihar. Quantify detection recall as a proxy for construction compliance; integrate with OMMAS portal.

Cartosat-2S, PMGSY-OMMAS data, OpenStreetMap

The Brahmaputra and Ganga basins experience annual catastrophic floods. Sentinel-1 SAR enables near-real-time mapping, but cloud contamination and speckle remain limiting for emergency response.

Design a SAR-based flood segmentation pipeline (U-Net with speckle-aware attention) integrated with NDWI from Sentinel-2 as a pseudo-label refinement signal. Deploy as a GEE-based near-real-time API for NDMA and State DMAs. Benchmark on 2017–2023 Assam and Bihar flood events.

Sentinel-1 GRD, Sentinel-2, MODIS NRT flood, IMD rainfall, NDMA records

India loses hundreds of lives annually to landslides. Current susceptibility maps use static geomorphological datasets; dynamic triggering via rainfall and seismic events is poorly captured.

Build a dynamic landslide susceptibility model combining InSAR-derived slope displacement (Sentinel-1), GPM near-real-time rainfall, and lithological maps. Train XGBoost with SHAP interpretability across Uttarakhand, Himachal Pradesh, and Kerala. Validate against BHUVAN Landslide Atlas (2020).

Sentinel-1 InSAR, GPM IMERG, NRSC Landslide Atlas, Geological Survey of India

India's eastern coastline is among the world's most cyclone-prone. Post-disaster building damage assessment currently takes weeks via field survey.

Develop a pre/post-event change detection framework using Sentinel-1 SAR coherence loss and Cartosat-3 optical change to classify building damage levels (Green/Yellow/Red) within 48 hours of cyclone landfall. Apply to Cyclones Fani (2019), Yaas (2021), and Michaung (2023).

Sentinel-1, Cartosat-3, NDMA event records, OSM building footprints

India has 5,200+ large reservoirs managed by CWC. Official storage reporting has a 2-week lag. Volumetric inference from satellite surface area remains challenging for irregular bathymetry.

Combine Sentinel-2 water surface area time series with ICESat-2 photon-counting LiDAR altimetry to derive area-elevation-volume (AEV) curves for 50 major Indian reservoirs. Validate against CWC daily storage bulletins (2018–2023). Develop GEE-based real-time storage anomaly dashboard.

Sentinel-2, ICESat-2, Landsat, CWC daily bulletins

GRACE/GRACE-FO has documented severe groundwater depletion in Punjab and Haryana. At 300 km resolution, GRACE cannot pinpoint district-level extraction hotspots needed for state-level regulation.

Downscale GRACE terrestrial water storage anomalies using Sentinel-1 InSAR subsidence, SMAP soil moisture, and ERA5 climate reanalysis via ML disaggregation. Validate against CGWB piezometer network. Identify villages where extraction rate exceeds recharge.

GRACE-FO, Sentinel-1 InSAR, SMAP, ERA5, CGWB piezometers

Vidarbha (Maharashtra) is India's most drought-vulnerable cotton belt, linked to high farmer distress. Multi-hazard drought monitoring integrating meteorological, agricultural, and hydrological signals is lacking at taluka level.

Develop a composite Drought Stress Index (DSI) fusing CHIRPS rainfall anomaly, MODIS VCI, and GRACE groundwater anomaly for Vidarbha's 62 talukas. Train a random forest classifier to identify 'severe drought' events aligned with government drought declarations (2012–2022). Provide a 4-week outlook using ML-downscaled IMD forecasts.

CHIRPS, MODIS VCI, GRACE-FO, IMD, Maharashtra drought declaration records

Northeast India hosts ~25% of India's biodiversity but faces intense jhum (shifting cultivation) pressure. High cloud cover >80% of the year limits optical monitoring.

Implement an all-weather forest disturbance detection algorithm combining Sentinel-1 SAR C-band backscatter with Landsat phenological composites. Map jhum cycles, secondary forest regrowth, and primary forest loss across 2010–2024 in the Northeast states. Compare with FSI State of Forest Reports.

Sentinel-1, Landsat, ALOS-2 PALSAR, FSI reports, Bhuvan forest maps

India holds 75% of the world's wild tiger population. Habitat fragmentation between tiger reserves threatens genetic connectivity. EO-derived land cover is rarely integrated into corridor modelling at 10 m resolution.

Produce a 10 m LULC map from Sentinel-2 for the Central Indian Landscape (connecting 10 tiger reserves). Feed into circuit theory habitat connectivity models (Circuitscape) and deep learning-predicted human-wildlife conflict zones (using NCRB records). Propose three priority wildlife corridors for legal notification.

Sentinel-2, NCRB wildlife conflict data, NTCA tiger census, OSM roads

The Indian Sundarbans are retreating due to sea level rise, cyclone impacts, and aquaculture encroachment. Automated multi-decadal change detection at species level is nascent.

Use Sentinel-1/2 and Landsat time series (2000–2024) to map mangrove extent, species composition proxies, and drivers of loss (aquaculture vs. erosion) across the Indian Sundarbans delta. Build a pixel-based attribution model for change drivers. Validate with WWF and Forest Department ground surveys.

Sentinel-1/2, Landsat archive, DESIS hyperspectral, WWF ground survey

India's coastline experiences both erosion (Kerala, West Bengal) and accretion (Gujarat mudflats). NCSCM assessments are decadal and cannot track annual monsoon-driven dynamism.

Apply DSAS coupled with a deep learning shoreline extraction model (trained on Sentinel-2 water-land boundaries) to produce annual shoreline change rates for all 11 coastal states (2015–2024). Identify hotspots of >2 m/yr erosion and link to upstream dam construction using river discharge data.

Sentinel-2, Landsat, NCSCM shoreline data, CWC discharge records

Noctiluca and Trichodesmium blooms are increasing in Indian coastal waters, threatening fisheries (3.5 million fisherfolk). MODIS chlorophyll products have 1 km resolution — too coarse for coastal HAB delineation.

Fine-tune a Sentinel-3 OLCI neural network for Indian coastal waters. Combine with SST (MODIS Aqua) and wind (ERA5) drivers in an LSTM early warning model for HAB onset in the Arabian Sea. Validate against CMLRE ship cruise data and fisheries distress alerts.

Sentinel-3 OLCI, MODIS Aqua, ERA5, CMLRE cruise data

India has 14 of the world's 20 most polluted cities. CPCB ground monitors cover <200 locations nationally. Satellite-based PM2.5 inference offers nationwide spatial coverage.

Train a geographically weighted random forest using TROPOMI NO2/aerosol index, MERRA-2 meteorology, and CPCB ground PM2.5 data to estimate daily PM2.5 at 1 km resolution across 100 Indian cities. Quantify exposure disparities between low-income and high-income wards using census socioeconomic data.

Sentinel-5P TROPOMI, MERRA-2, CPCB PM2.5, IIT monitoring network

Annual crop residue burning in Punjab-Haryana causes severe winter air quality crises in Delhi. NASA FIRMS provides fire hotspots but cannot quantify emission intensity at field level.

Combine Sentinel-3 SLSTR fire radiative power with Sentinel-2 burned area mapping to construct a field-level paddy stubble burning emission inventory (PM2.5, BC, CO). Link to PMFBY crop insurance data to identify smallholder incentive structures. Compare estimates against IITM-Pune atmospheric transport model forecasts.

Sentinel-3 SLSTR, Sentinel-2, NASA FIRMS, PMFBY records, IITM model output

India's solar capacity crossed 70 GW in 2024. Satellite-based mapping of utility-scale and distributed rooftop solar lags behind actual installations, hindering grid planning.

Develop a semantic segmentation model (SegFormer) trained on Cartosat-3 + Sentinel-2 to automatically detect, classify, and size solar PV installations across Rajasthan, Gujarat, and Tamil Nadu. Track annual capacity additions and compare against MNRE reported capacity.

Cartosat-3, Sentinel-2, MNRE capacity reports, LISS-IV

India is executing the world's largest railway expansion (dedicated freight corridors) and National Highway development programs. Physical progress monitoring relies on manual site inspections.

Design a construction progress monitoring pipeline using Sentinel-1 SAR coherence change (for earthwork) + Cartosat-2S optical change (for surface completeness) along DFC-East and -West corridors. Define a Construction Progress Index (CPI) and validate against DFCCIL quarterly reports.

Sentinel-1, Cartosat-2S, DFCCIL reports, NHAI construction MIS

India has ~100,000 sacred groves (Devvans, Nandavana) — biodiversity refugia managed by indigenous communities. These are unmapped at national scale and receive no formal protection.

Use high-resolution Cartosat-3 imagery and spectral texture analysis to identify and delineate sacred grove patches across Kerala, Meghalaya, and Rajasthan. Train a few-shot classifier using culturally-informed field annotations from tribal community partners. Estimate total carbon stock and biodiversity contribution.

Cartosat-3, Sentinel-2, community survey data, Forest Survey of India

India hosts the world's largest human gatherings. Real-time crowd density estimation is critical for stampede risk management. Satellite imagery offers the only overhead view at full event extent.

Develop a density map estimation model (CSRNet adapted for overhead satellite view) using PlanetScope 3 m imagery at Prayagraj Kumbh 2025. Train using synthetic density maps from lidar-validated pedestrian simulations. Validate against police headcount checkpoints. Propose a real-time crowd risk score API for disaster management authorities.

PlanetScope, Pleiades, SDMA event records, police crowd estimates

The Deccan Plateau harbors thousands of undocumented megalithic and early historic sites. Remote sensing-based archaeological prospection is nascent in India.

Apply anomaly detection on SRTM DEM derivatives combined with Sentinel-2 vegetation anomalies and soil signatures to identify candidate archaeological features in Karnataka and Telangana. Combine with GPT-4 Vision analysis of drone orthophotos for feature classification. Validate 20% of candidates with ASI ground surveys.

SRTM/ALOS DEM, Sentinel-2, ASI records, drone surveys

Illegal sand mining in Indian rivers (Yamuna, Mahanadi, Godavari) causes severe ecological damage and infrastructure risk. Enforcement is hindered by poor monitoring. SAR is cloud-independent and sensitive to riverbed morphology changes.

Develop a change detection pipeline using Sentinel-1 SAR backscatter time series to identify riverbed morphological anomalies indicative of sand extraction in 10 major Indian rivers. Combine with nighttime light (VIIRS) anomalies to detect mining operations outside permitted hours. Validate against NGT court-documented mining locations.

Sentinel-1, VIIRS nighttime light, NGT enforcement records, CWC surveys

India is developing a national voluntary carbon market (CCTS). Soil organic carbon credits for regenerative agriculture require transparent, scalable remote sensing-based MRV.

Build a predictive SOC mapping model using Sentinel-2 bare soil composites, hyperspectral proxies, and soil survey data from ICAR-NBSS&LUP for three agro-ecological zones. Use explainable machine learning to validate driving factors. Assess suitability for Indian carbon credit certification under BIS standards.

Sentinel-2, DESIS hyperspectral, ICAR soil survey, climate gridded data

The 2023 Sikkim GLOF disaster highlighted India's vulnerability. Over 200 potentially dangerous glacial lakes exist in Indian Himalayan states. Near-real-time monitoring and risk classification is urgently needed.

Develop an automated glacial lake inventory and expansion monitoring pipeline using Sentinel-2 multitemporal imagery and SAR-coherence for ice dam detection. Train a GLOF risk classifier combining lake area growth rate, moraine stability (InSAR), and downstream exposure indices. Pilot for Sikkim and Uttarakhand with NDMA integration.

Sentinel-1/2, ALOS DEM, GLIMS glacial lake database, NDMA records