Global EO-AI Research
Compendium

Existing fire spread models (FARSITE, FlamMap) rely on empirical fuel models poorly calibrated to non-North American ecosystems. Deep learning models ignore physical constraints, producing physically implausible spread predictions under novel climate conditions.

Develop a physics-informed neural network (PINN) encoding Rothermel fire spread equations as soft constraints during training, fused with Sentinel-2 fuel maps and ERA5 wind/humidity. Train on 5 years of global VIIRS active fire detections. Evaluate on Australian Black Summer (2019–20), Canadian 2023 megafires, and Amazonian 2020 fires. Key metric: 6-hour spread polygon overlap vs. NIFC perimeters.

NASA VIIRS active fire, Sentinel-2 fuel maps, ERA5, NIFC perimeters, FIRMS

Post-fire vegetation recovery rates are poorly quantified across biomes. Global assessments conflate regrowth with invasive species colonisation, and few studies exceed 10-year recovery windows.

Apply a spectral trajectory clustering approach (CCDC + k-means) to 40 years of Landsat to characterise vegetation recovery curves for 6 fire-prone biomes: Mediterranean shrubland, boreal forest, tropical savanna, temperate forest, chaparral, and peatland. Quantify how drought, soil type, and re-burn interval modulate recovery speed. Publish a global fire recovery database.

Landsat Collection 2, MODIS dNBR, SoilGrids, SPEI drought index

Peatland fires release disproportionate carbon per unit area, yet global carbon accounting poorly constrains smouldering combustion depth. Satellite-based estimates carry ±50% uncertainty.

Combine Sentinel-1 SAR coherence loss (sensitive to peat structure change) with GEDI LiDAR subsidence and TROPOMI CO column anomaly to estimate smouldering depth and total carbon flux during peatland fire events in Borneo (2015, 2019) and the Congo Basin (2022–23). Compare against GFED v4.1s.

Sentinel-1, GEDI, TROPOMI CO, GFED v4.1s, Indonesian peat depth maps

Brazil's DETER system flags deforestation at 25 ha minimum polygon size with 16-day Landsat latency. Sub-hectare selective logging and edge degradation are systematically missed, underestimating forest carbon loss by 30–50%.

Build a near-real-time sub-hectare deforestation and degradation detector using Planet NICFI basemaps (4.77 m) fused with Sentinel-1 SAR via a cross-modal attention transformer. Target: 1-ha minimum detection, 5-day latency. Validate against AmazonFACE plots and INPE field verification dataset. Deploy as open API.

Planet NICFI, Sentinel-1, Landsat, INPE PRODES, MapBiomas

Selective logging removes individual high-value trees leaving canopy largely intact, nearly invisible to 2D optical change detection. It accounts for ~50% of tropical forest carbon loss but is absent from most deforestation datasets.

Pair GEDI waveform LiDAR canopy height with Sentinel-2 canopy texture to detect selective logging gaps (<0.5 ha) across three concession types (FSC-certified, uncertified, illegal) in Malaysia, DRC, and Gabon. Quantify carbon losses attributable to logging intensity. Develop a Logging Intensity Index for forest governance.

GEDI L2A/L4A, Sentinel-2 canopy texture, FSC concession boundaries, Global Forest Watch

Mangroves store 3–5× more carbon per hectare than upland tropical forests, yet global mangrove carbon maps carry 40% uncertainty due to poor canopy height estimates and variable below-ground carbon ratios.

Integrate GEDI canopy height (for AGB estimation), Sentinel-1 SAR (for inundation and biomass proxy), and GMW v3.0 to produce a consistent 2010–2024 mangrove carbon stock change map at 25 m resolution. Partition loss drivers (aquaculture, erosion, sea level rise, storms). Validate against 200 field plots across 5 countries.

GEDI, Sentinel-1/2, GMW v3.0, SRTM, field biomass plots (CIFOR)

Ice sheet mass balance estimates from GRACE, ICESat-2, and InSAR diverge by up to 40 Gt/yr over East Antarctica due to firn densification uncertainty and different spatial footprints. Reconciliation is an open problem.

Develop a Bayesian data fusion framework combining GRACE-FO gravity anomalies, ICESat-2 elevation change, and Sentinel-1 ice flow velocities to jointly invert for mass balance, firn compaction, and bedrock uplift (GIA) over WAIS and EAIS. Compare reconciled estimates against IMBIE-3 intercomparison. Quantify 2020–2025 mass loss acceleration.

GRACE-FO, ICESat-2, Sentinel-1 InSAR, BedMachine Antarctic DEM, RACMO firn model

Sea ice thickness — the critical missing variable in Arctic shipping and climate models — is only sparsely measured by CryoSat-2 radar altimetry. Current ML models underperform in the marginal ice zone and during melt seasons.

Train a spatiotemporal transformer on 15 years of CryoSat-2 thickness retrievals, SMOS thin-ice products, AMSR2 sea ice concentration, and TOPAZ ocean reanalysis to predict pan-Arctic sea ice thickness at 25 km / weekly resolution. Quantify model uncertainty with conformal prediction. Validate against IceBridge airborne and submarine sonar data.

CryoSat-2, SMOS, AMSR2, TOPAZ4 ocean reanalysis, IceBridge airborne

High-Mountain Asia contains 100,000+ glaciers, most without in-situ mass balance records. The 'Third Pole' supplies water to 1.9 billion people, yet glacier-by-glacier mass change remains unknown for 80% of individual glaciers.

Apply a deep learning geodetic mass balance approach using differenced SRTM (2000), TanDEM-X (2011–2014), and ICESat-2 (2018–2024) DEMs to estimate individual glacier mass balance for all HKH glaciers. Train a regression model to predict mass balance from morphological attributes (slope, aspect, elevation, debris cover). Validate against WGMS benchmark glaciers.

SRTM, TanDEM-X, ICESat-2, RGI 7.0, WGMS benchmark, Landsat debris cover

In-situ coral bleaching surveys cover <1% of global reef extent. The 4th global bleaching event (2023–2025) is the most severe on record, yet satellite-based bleaching confirmation lags event onset by months.

Develop a coral bleaching probability model using Sentinel-2 bottom reflectance (10 m), SST anomaly (NOAA CoralTemp), and MODIS chlorophyll as predictors. Train on CoralNet / Global Coral Reef Monitoring Network field surveys. Deploy on GEE as a weekly bleaching alert system for 50 priority reef systems. Validate recall against GBRMPA aerial surveys.

Sentinel-2, NOAA CoralTemp, MODIS, CoralNet annotations, GBRMPA surveys

Ocean plastic accumulation zones are poorly mapped. Sentinel-2 can detect large surface slicks but cannot discriminate plastic from natural organic matter or foam. Dedicated hyperspectral missions (DESIS, EMIT) enable spectral unmixing.

Build a spectral library of floating plastic types (HDPE, PET, PP, fishing nets) and organic lookalikes (Sargassum, marine foam) from lab spectra. Train a spectral unmixing model on EMIT (380–2500 nm) and validate against aerial drone surveys and NCEAS plastic debris databases. Map global hotspot accumulation zones with mass-per-unit-area estimates.

EMIT hyperspectral, DESIS, Sentinel-2, drone validation surveys, NCEAS database

Illegal, unreported, and unregulated (IUU) fishing costs $10–23 billion annually. 'Dark vessels' that disable AIS transponders are undetectable by conventional means but visible in SAR imagery.

Train a YOLOv8-based ship detector on Sentinel-1 SAR imagery and cross-reference detections against AIS vessel registry to identify dark vessels. Apply to three high-IUU zones: South China Sea, Patagonian Shelf, and West African EEZ. Quantify AIS-dark vessel density, seasonal patterns, and flag state attribution. Partner with Global Fishing Watch for deployment.

Sentinel-1 SAR, AIS vessel registry, Global Fishing Watch, VIIRS boat detection

Existing 10 m global LULC maps (ESA WorldCover, DynamicWorld) have class disagreements of 20–40% in transitional zones and are produced at best annually. No operational annual update pipeline exists at this resolution.

Design a self-updating LULC classification pipeline using Sentinel-2 quarterly composites and a knowledge-distilled SegFormer model with pseudo-label self-training. Target 11 IPCC classes at 10 m with annual cadence. Quantify class-specific accuracy improvement over WorldCover 2021 using 50,000 stratified validation points from LUCAS, GFSAD, and IIASA Geo-Wiki. Release as open GEE asset.

Sentinel-2, ESA WorldCover, DynamicWorld, LUCAS, GFSAD30, Geo-Wiki crowdsourced

SMAP L-band soil moisture (36 km) is too coarse for field-scale irrigation decisions. Downscaling to <1 km while preserving temporal accuracy is an open challenge, particularly over heterogeneous terrain.

Train a conditional diffusion model to downscale SMAP 36 km → 1 km soil moisture using Sentinel-1 backscatter, Sentinel-2 vegetation indices, and topographic wetness index as conditional inputs. Evaluate on 500+ ISMN in-situ stations across 6 continents. Compare against DISPATCH, SEN-ET, and classical regression downscaling. Assess drought classification accuracy for FAO early warning.

SMAP, Sentinel-1/2, ISMN station network, FAO drought declarations, TWI from SRTM

An estimated 400–950 Mha of cropland has been abandoned globally since 1990 due to rural depopulation, conflict, economic transitions, and climate stress. Re-cultivation of abandoned lands is poorly tracked, creating systematic errors in carbon accounting and food security models.

Apply CCDC-based land trajectory analysis to 30 years of Landsat to map global cropland abandonment and recultivation events at 30 m resolution. Link trajectories to drivers using socioeconomic covariates (GDP/capita, conflict data, climate anomalies). Quantify carbon uptake from fallow regrowth and release from recultivation.

Landsat archive, FAO GAEZ, World Bank GDP, ACLED conflict, SPEI drought

The oil/gas sector contributes 30% of anthropogenic methane emissions, but large 'super-emitter' point sources are responsible for a disproportionate share. Until EMIT (2022), their detection required expensive airborne campaigns.

Develop an automated methane plume detection pipeline using NASA EMIT hyperspectral data (matched filter approach) cross-validated against Sentinel-5P TROPOMI column anomalies. Build a global super-emitter database (>10,000 tonne CO₂eq/year per site) spanning oil fields, landfills, and coal mines across 20 countries. Quantify detection sensitivity and false positive rate.

NASA EMIT, Sentinel-5P TROPOMI, SRON methane dataset, Global Oil & Gas features

Multi-decadal aerosol trend attribution is contested — satellite AOD trends conflate emission changes with retrieval artefacts from surface reflectance changes (urbanisation, greening). Robust long-term homogenised records are lacking.

Construct a homogenised 2000–2024 AOD trend record by cross-calibrating MODIS Terra/Aqua, MISR, and VIIRS using AERONET as ground truth. Decompose trends into meteorological variability and emission drivers using causal inference (DoWhy framework). Analyse 30 global megacities including Delhi, Beijing, Cairo, Lagos, and São Paulo.

MODIS AOD, MISR, VIIRS, AERONET, ERA5, emission inventories (CAMS, EDGAR)

CMIP6 climate models run at 100 km resolution — too coarse for urban flood risk assessment in cities of 1–5 million inhabitants. Statistical downscaling approaches lack physical consistency during extreme precipitation events.

Train a physics-constrained GAN (PhyDNet architecture) to downscale CMIP6 precipitation fields to 1 km over 100 global cities using ERA5 as bridge observational dataset. Couple downscaled precipitation with a HAND-based inundation model to project 100-year flood extents under SSP2-4.5 and SSP5-8.5 for 2050 and 2100. Validate against historical flood insurance records.

CMIP6, ERA5, SRTM HAND model, satellite-derived flood extents, insurance records

Permafrost underlies 25% of the Northern Hemisphere land surface and stores twice the atmospheric carbon pool. Infrastructure built on permafrost faces accelerating failure risk. Continental-scale InSAR subsidence monitoring is computationally intractable with current workflows.

Apply PS-InSAR time series analysis (Sentinel-1) at continental scale using cloud HPC (AWS/GEE) to produce pan-Arctic annual surface deformation maps at 20 m resolution for 2015–2024. Correlate subsidence hotspots with thermokarst lake expansion (Landsat), permafrost temperature records (GTN-P), and infrastructure at risk (OpenStreetMap). Estimate methane release potential from thermokarst formation.

Sentinel-1 InSAR archive, Landsat, GTN-P borehole temperatures, OpenStreetMap infrastructure

Megadeltas (Mekong, Nile, Mississippi, Ganges-Brahmaputra) combine land subsidence, sediment deficit, and sea level rise into compound hazards. Most exposure assessments use static DEMs ignoring dynamic sediment processes and human-induced subsidence.

Integrate InSAR land subsidence (Sentinel-1), GRACE groundwater depletion, ICESat-2 delta elevation change, and sea level projections (IPCC AR6 regional) to produce compound sea level exposure maps for 10 megadeltas. Quantify cropland, population, and critical infrastructure at risk under 2050 and 2100 scenarios.

Sentinel-1 InSAR, GRACE-FO, ICESat-2, IPCC AR6 sea level, tide gauges, WorldPop

The CBD Kunming-Montreal framework requires countries to report on 30×30 biodiversity targets. Most Essential Biodiversity Variables (EBVs) lack satellite-based operationalisation, creating a reporting accountability gap for 196 signatories.

Operationalise four EBVs from satellite data: (1) canopy chlorophyll content from Sentinel-2 Red Edge, (2) species habitat suitability from EO-derived LULC + climate, (3) ecosystem disturbance frequency from Landsat, and (4) phenological diversity from MODIS NDVI. Produce global maps and validate against GBIF occurrence records and NEON field sites.

Sentinel-2, Landsat, MODIS, GBIF, NEON field sites, CBD national reports

Sub-Saharan African savannas face interacting pressures: overgrazing, bushmeat hunting, charcoal extraction, and agricultural expansion. Distinguishing degradation from natural seasonal and interannual variability is a key unresolved challenge.

Develop a vegetation functional type change detection model combining MODIS long-term NDVI phenology (anomaly-based), Sentinel-2 fractional cover (grass/shrub/bare), and GEDI canopy height. Classify degradation trajectories for 5 African savanna countries over 2001–2024. Validate against GlobeLand30 and national land degradation assessments.

MODIS NDVI, Sentinel-2, GEDI, GlobeLand30, UNCCD national reports

Ramsar-listed wetlands (2,400+ globally) lack systematic satellite-based condition monitoring. Ecosystem services (carbon sequestration, water purification, flood buffering) are being lost faster than they are documented.

Build a global wetland mapping and condition index using Sentinel-1 SAR inundation frequency, Sentinel-2 NDWI seasonal composites, and GEDI surface water elevation to classify wetland type, inundation regime, and vegetation condition. Generate annual condition reports for 100 priority Ramsar sites. Validate against IWMI WaterPix and Copernicus GSW products.

Sentinel-1/2, GEDI, Copernicus GSW, Ramsar site boundaries, IWMI WaterPix

Vision-language models (CLIP, GPT-4V) underperform on satellite imagery due to domain gaps in spatial resolution, spectral range, and overhead perspective. A dedicated EO-language foundation model does not yet exist at sufficient scale.

Pretrain an EO-language model (EO-CLIP) using 50M satellite image-text pairs scraped from Copernicus Open Access Hub, Bhuvan, USGS Earth Explorer, and scientific paper figure captions. Evaluate zero-shot performance on 12 downstream tasks (scene classification, change captioning, disaster QA) across GEO-Bench and EarthVQA. Quantify gains from domain-specific pretraining vs. CLIP fine-tuning.

Copernicus Open Access Hub, USGS Earth Explorer, GEO-Bench, EarthVQA, scientific figure captions

Cloud cover renders 60–80% of optical satellite imagery unusable in the tropics. Existing cloud removal methods (gap-filling, temporal compositing) produce blurry outputs during persistent cloud events (monsoon seasons).

Train a conditional latent diffusion model for cloud removal conditioned on (1) SAR co-registered imagery, (2) adjacent cloud-free dates, and (3) land cover priors. Train on a global paired cloud/cloud-free dataset with synthetic cloud injection. Evaluate on RICE, Sen2_MTC, and custom monsoon-season Indian datasets. Outperform SEN12MS-CR and GLF-CR baselines on SSIM and downstream task accuracy.

Sentinel-1/2, RICE cloud dataset, Sen2_MTC, custom Indian monsoon dataset

Labelled change detection datasets are expensive to produce and highly domain-specific. Models trained on urban change fail on agricultural change; fire scar models fail on flood inundation. A universal, label-free approach is an open frontier.

Develop a contrastive self-supervised change detection framework (ChangeSSL) that learns change representations purely from co-registered bitemporal pairs without change labels, using a masked autoencoder pre-text task. Fine-tune with <1% labelled data per domain. Benchmark on LEVIR-CD (urban), xBD (disaster), SECOND (multi-class), and S2Looking. Demonstrate generalisation across 5 change types without retraining.

LEVIR-CD, xBD disaster dataset, SECOND, S2Looking, Sen1Floods11

VIIRS NPP night-time light conflates economic activity, electricity access, gas flaring, fishing boats, and conflict-related outages. Decomposing these signals for SDG7 (energy access) monitoring is unresolved at sub-district scale.

Build a spectral + temporal decomposition model for VIIRS DNB separating stable economic light, gas flaring (VIIRS SWIR), ephemeral events (disasters, conflict), and seasonal agricultural activity. Apply to track electricity access progress in Sub-Saharan Africa and South Asia, and conflict-related blackouts in Ukraine (2022–24) and Gaza (2023–24). Validate against DHS electrification surveys.

VIIRS DNB, VIIRS SWIR, ACLED conflict, DHS electrification surveys, World Bank

The NASA Global Landslide Catalog captures <5% of global events. Remote mountain ranges have near-zero detection coverage. An incomplete inventory systematically underestimates global landslide risk and derails infrastructure planning.

Combine Sentinel-1 SAR coherence loss, Sentinel-2 surface disturbance, and Planet high-cadence imagery to build an automated global landslide detection pipeline. Represent the landscape as a graph (nodes = slope units, edges = hydrological connectivity) and apply a GNN susceptibility model trained on known inventory data. Publish a completed global landslide inventory for 2017–2024.

Sentinel-1/2, Planet, NASA Global Landslide Catalog, OpenTopography DEM

Conflict damage assessment in active war zones is impossible via ground survey. Satellite-based building damage mapping (UNOSAT, Copernicus EMS) relies on manual analyst time, creating severe bottlenecks during mass-casualty events.

Develop an automated conflict damage assessment pipeline using Sentinel-1 SAR coherence change + Maxar/Planet VHR optical change, fine-tuned on the xBD building damage dataset. Apply to Ukraine (2022–24), Sudan, and Gaza conflict zones. Benchmark against UNOSAT verified damage polygons. Develop an active learning loop for analyst efficiency improvement.

Sentinel-1, Maxar/Planet VHR, xBD dataset, UNOSAT damage assessments

The global energy transition requires 3–6× more lithium, cobalt, nickel, and copper by 2040. Mine footprint expansion for these critical minerals is poorly tracked globally, creating blind spots in ESG disclosure and community impact assessment.

Build a global open-cast mine detection and expansion monitoring system using Sentinel-2 spectral indices (bare soil, ferric oxide, gossan) and Sentinel-1 SAR (surface roughness change). Develop a mine activity index from temporal change rates. Map 500+ critical mineral mines across DRC, Chile, Australia, Philippines, and Indonesia for 2015–2024. Link expansion rates to commodity price signals and deforestation incidents.

Sentinel-1/2, SNL Metals & Mining database, Global Forest Watch, commodity price data