Turning the growing multi-sensor satellite archive into robust, scalable measurements of ecosystem change.
Overview
Coastal waters, wetlands and croplands respond within days to heat waves, floods and nutrient pulses, and space-borne sensors offer global coverage for assessing natural and human-induced change. My work sits at the intersection of three areas: robust remote sensing, which makes observations consistent and trustworthy across sensors and time; multi-sensor integration, which combines multispectral, hyperspectral, thermal, radar and in situ data; and geospatial AI, which turns these records into maps and indicators at regional to global scales with few labels.
The work is application-focused: each method is validated against field data and delivered as a product others can use.
Research interests: remote sensing, micro- and macroalgae blooms, water quality, radiative transfer modeling, computer vision for Earth observation, and domain-specific foundation models.
Three areas of work
Satellite optical data processing on cloud and HPC
Flagship: AQUAVis
A physics-based processing chain that turns raw satellite radiance into water reflectance, correcting for atmosphere, sun glint, adjacency and clouds, and validated against in situ radiometry.
Read the details
The radiance a satellite measures is affected by atmospheric scattering and absorption, sun glint, adjacency effects from nearby land, and clouds and their shadows, so each application needs a carefully designed processing chain. My earlier work evaluated atmospheric, glint and adjacency corrections for multispectral and hyperspectral sensors, validated against in situ radiometry.
Building on this, I developed AQUAVis, a physics-based pipeline for Landsat-8/9 and Sentinel-2 that combines vicarious calibration, 6SV radiative transfer atmospheric correction, adjacency and sun glint corrections, and cloud, shadow and water masks (Science of Remote Sensing, 2025).
Flagship: a water-specific spectral bandpass adjustment
Rigorous ways to combine Landsat, Sentinel, hyperspectral and field data into one consistent record for water, so models learn the bio-optical signal instead of sensor noise.
Read the details
Differences in signal-to-noise ratio, spectral response and temporal aggregation can lead models to learn noise instead of the true bio-optical signal, so sensors have to be combined rigorously. I showed that NASA's Harmonized Landsat–Sentinel-2 product, derived for land, does not hold over water, and derived a water-specific spectral bandpass adjustment function (Science of Remote Sensing, 2025). AQUAVis then provides a harmonized Landsat–Sentinel-2 record for water, validated with in situ AERONET-OC.
I have also combined field radiometry and pigment analysis with Sentinel-3 OLCI and PRISMA hyperspectral data to map cyanobacteria (Harmful Algae, 2025; Remote Sensing, 2023), and integrated MODIS, climate and soil data to forecast soybean yield and avian influenza risk across the Mississippi River Basin.
Flagship: 30-m global floating algal bloom mapping
Vision Transformers and self-supervised foundation models that map floating algal blooms at 30 m globally, trained on harmonized data and run on HPC and the cloud.
Read the details
Deep learning and geospatial foundation models learn spatial and spectral patterns that indices and empirical models miss, and with cloud computing and HPC they can map Earth's surface at global scale. They depend on diverse, physically consistent training data and independent validation.
My doctoral research, funded by a NASA Early Career Research fellowship, explores these questions through Vision Transformers for 30-m global floating algal bloom mapping on harmonized Landsat–Sentinel-2 data, a self-supervised foundation model for global coastal waters, and compact coastal aquatic embeddings that encode the bloom signal and improve chlorophyll-a retrieval. The implementation runs on HPC resources with SLURM and PBS schedulers and on AWS.