Research — BuckAI Observatory

BuckAI's research program applies deep learning, computer vision, and physics-informed AI to large scientific datasets — from the full spectrum of satellite modalities to astronomical surveys and high-dimensional numerical simulations. While Earth observation is our primary focus, our six themes extend from applied remote sensing to the development and advancement of AI algorithms for science more broadly. Affiliated faculty and students span Earth Sciences, Geography, Mathematics, Astronomy, Ecology, and Environmental Sciences.

1

🌊

Hydrology & Water Resources

From Arctic rivers to the Amazon floodplain, AI tools are transforming our ability to monitor water bodies at global scale, with millimeter to meter precision, at temporal frequencies previously impossible with manual analysis.

River Classification at Sub-Meter Resolution

Fully convolutional neural networks trained on 30 cm WorldView/QuickBird commercial imagery achieve >90% precision and recall for river detection. Enables tracking of Arctic river width changes and snowmelt-driven discharge at seasonal to annual timescales.

FCN WorldView Remote Sensing Env. 2022

PI: Moortgat

Super-Resolution Water Mapping

Deep neural networks applied to free Sentinel-2 multispectral imagery produce 2 m effective water classification maps, extending high-resolution monitoring to the global, cloud-free Sentinel archive rather than costly commercial data.

Sentinel-2 Super-resolution J. Hydrology 2023

PI: Moortgat

River Discharge from Hypsometry + AI

Combining river channel geometry (hypsometry) with multi-temporal satellite imagery to improve remote estimates of river discharge — bridging the gap between satellite observation and hydrological modeling needs.

SWOT Remote Sensing Env. 2024

PIs: Moortgat, Durand, Alsdorf

Critical Zone Hydrology & Remote Sensing

AI analysis of multi-sensor satellite data (VIIRS, MODIS, GRACE) for large-scale monitoring of land-water-vegetation interactions, including changes in terrestrial water storage and land surface hydrology under climate change.

GRACE MODIS Critical Zone

PIs: Costa, Alsdorf, Yeo

2

🌊

Coastal & Marine Systems

The world's coastlines are among the most data-rich and most vulnerable environments on Earth. BuckAI applies physics-informed AI, gravity missions, and satellite altimetry to track bathymetry, sea level, and marine ecosystem health.

Coastal Bathymetry via Physics-Informed AI

Physics-informed neural networks (PINNs) fusing ICESat-2 lidar, Sentinel-2 multispectral imagery, and hydrodynamic constraints to map shallow-water depths with unprecedented accuracy. NSF CAIG proposal pending ($1.5M, 2026–2029).

PINNs ICESat-2 Sentinel-2 WACV 2025

PI: Moortgat

Sea-Level Change & Satellite Geodesy

Combining GRACE gravity data, satellite altimetry, and GPS measurements to separate ice-mass loss, thermal expansion, and hydrological contributions to sea-level rise — with AI used to improve signal separation and downscaling.

GRACE/GRACE-FO Altimetry Sea Level

PIs: Shum, Gómez

Coral Reef & Ocean Heat Stress Monitoring

AI analysis of infrared and microwave satellite data (NOAA CoralTemp, MODIS) to detect ocean heat stress events, map coral bleaching risk, and track long-term reef ecosystem change — supporting conservation and fisheries management.

MODIS Ocean Heat Stress Coral Bleaching

PI: Costa

Wetland Mapping & Carbon Dynamics

Multi-sensor satellite analysis to map wetland extent, inundation dynamics, and carbon fluxes — critical for both climate mitigation accounting and coastal ecosystem management.

SAR Optical Carbon

PI: Yeo

3

🌳

Forests & Ecosystems

Tropical deforestation, forest carbon stocks, and biodiversity loss demand global, continuous monitoring. BuckAI develops AI tools that turn satellite archives into actionable environmental intelligence — from Uganda to the Sahel.

Deforestation & Agroforestry in Sub-Saharan Africa

AI-driven satellite monitoring of forest loss and agroforestry expansion in Uganda and the Sahel, supporting CGIAR/SPIA food security and climate mitigation research ($1.25M, 2025–2028). Uses Sentinel-1/2, Landsat, and PlanetScope imagery with semantic segmentation models.

Sentinel Semantic Segmentation CGIAR/SPIA Active

PI: Moortgat

Forest Carbon Stocks & Biomass from LiDAR

Combining spaceborne LiDAR (GEDI, ICESat-2) with optical and SAR data and deep learning to estimate forest biomass, carbon stocks, and forest structure at national and global scales — inputs to carbon crediting and climate models.

GEDI LiDAR Forest Biomass

PI: Zhao

Biodiversity & Species Distribution Modeling

Machine learning models linking remote sensing variables (vegetation structure, land cover, phenology) to species occurrence data, enabling spatial predictions of biodiversity patterns and responses to land-use and climate change.

Macroecology Species Distribution Trait-Based Ecology

PI: Jarzyna

Land Use, Land Cover & Urban Change

AI methods for large-scale land use and land cover classification, urban expansion mapping, and analysis of human-environment interactions — linking spatial data to socioeconomic and environmental outcomes.

LULC Urban Remote Sensing GIS

PI: D. Liu

4

🧊

Cryosphere & Climate

Ice sheets, mountain glaciers, and snow cover are Earth's largest freshwater stores and among its most sensitive climate indicators. AI applied to multi-decadal satellite records reveals the pace and pattern of cryospheric change.

Ice Sheet Dynamics & Glacial Retreat

Multi-sensor analysis of Greenland and Antarctic ice sheet change — combining optical, SAR, and altimetry data with deep learning to map ice velocity, margin retreat, surface mass balance, and subglacial hydrology at continental scale.

Sentinel-1/2 ICESat-2 Ice Dynamics

PI: Howat

Mountain Glaciers & Water Security

AI and remote sensing to quantify glacier mass balance, retreat rates, and downstream water availability in the Andes and Himalaya — regions where millions depend on glacial meltwater for agriculture and drinking water.

Andes Himalaya Water Security

PI: Mark

Snow Remote Sensing & Runoff Prediction

Machine learning models for snow water equivalent estimation from multi-frequency microwave and optical imagery — improving seasonal streamflow forecasts critical for water resource management in snowmelt-dominated basins.

AMSR Snow Water Equivalent Streamflow

PIs: Durand, Quiring

Ice Mass Balance from Satellite Geodesy

Combining GRACE gravity time-series, satellite altimetry, and GPS uplift measurements to separate ice-mass loss from hydrological changes and glacial isostatic adjustment — with AI improving signal decomposition.

GRACE Glacial Isostasy GPS

PIs: Shum, Gómez

5

⚡

Geodesy & Subsurface

Precise space-based geodesy, combined with AI-driven surface analysis, extends Earth observation into the subsurface — enabling prospecting for clean energy resources, earthquake early warning, and continental-scale deformation monitoring.

Natural Hydrogen Reservoir Prospecting

Semantic segmentation of "fairy circles" (semicircular surface depressions, or SCDs) using Segment Anything Model (SAM) + Sentinel-2 + Copernicus DEM for global mapping of potential natural hydrogen reservoirs. Covered by New Scientist, EurekAlert, and 10+ outlets; industry partner Koloma Inc. ARPA-E subcontract active.

SAM Sentinel-2 Natural Hydrogen Active

PI: Moortgat

GPS Seismology & Earthquake Monitoring

Real-time and near-real-time processing of high-rate GNSS data for seismic waveform extraction, earthquake source characterization, and tsunami early warning — including AI methods for rapid magnitude and slip estimation.

GNSS Earthquake Tsunami Warning

PI: Crowell

Satellite Gravity & Geoid Modeling

Advanced processing and AI-assisted interpretation of GRACE/GRACE-FO gravity time-series for hydrology, oceanography, solid Earth, and cryosphere applications — including improved downscaling and signal separation methods.

GRACE-FO Geoid Geodesy

PIs: Shum, Gómez

Critical Zone & Subsurface Geochemistry

AI integration of remote sensing with in-situ measurements to study weathering, hydrology, and biogeochemical cycling in Earth's critical zone — from surface vegetation to bedrock — and track changes driven by climate and land use.

Critical Zone VIIRS Geochemistry

PI: Costa

6

🧮

AI Methods & Scientific Computing

Advancing the AI algorithms themselves — developing scientific machine learning, uncertainty quantification, self-supervised learning, and ML-enhanced numerical methods that benefit the entire natural and mathematical sciences community.

Scientific Machine Learning & Uncertainty Quantification

Developing neural network architectures and training strategies tailored to scientific problems: learning from small labeled datasets, quantifying prediction uncertainty, enforcing physical constraints, and constructing data-driven reduced-order models for complex dynamical systems.

Scientific ML UQ Dynamical Systems

PI: Xiu

AI for Large Astronomical Surveys

Self-supervised learning, generative models, and foundation model approaches for spectroscopic and photometric survey data. Developing methods that transfer across observational modalities and enable efficient scientific inference from petabyte-scale astronomical archives.

Self-Supervised Learning Foundation Models Spectroscopy

PI: Ting

ML-Enhanced Numerical Methods

Combining classical numerical analysis with modern machine learning to build faster, more accurate solvers for differential equations and high-dimensional problems. Applications include hybrid PDE solvers, learned preconditioners, and neural-network-accelerated simulations.

Numerical Analysis PDEs Hybrid Solvers

PI: Han Veiga

Cross-Domain Foundation Models

A cross-cutting effort bringing together methods expertise (Mathematics, Astronomy) and domain data (Earth observation, ecology) to build and fine-tune foundation models that generalize across scientific disciplines — a key long-term goal of the Observatory.

Foundation Models Transfer Learning Cross-Domain

PIs: Xiu, Ting, Han Veiga, Moortgat