[14]
Earth2EMIT: Spatiotemporal Autoencoders for Orbital-Proximal Fusion
Georgia Tech · MIT MCSC · MIT Earth Intelligence Lab
Sujay Nair
My research sits at the intersection of machine learning and planetary-scale sensing — training foundation models that map Earth's mineral resources via satellite hyperspectral imagery, and building physics-guided systems that extract geochemical signal from high-dimensional observational data to advance carbon capture and critical mineral discovery.
Fourth-year double major in Computer Science and Mathematics at Georgia Tech. Research affiliate at the MIT Climate & Sustainability Consortium and MIT Earth Intelligence Lab, advised by Dr. Evan Coleman, Prof. Sherrie Wang, and Prof. Elsa Olivetti. Publications at AAAI, ICLR, AAS, PASP, and more.
01
Research Areas
01
MIT Earth Intelligence Lab — Earth Observation & Soil Intelligence
MIT Earth Intelligence Lab: Advised by Prof. Sherrie Wang & Dr. Evan Coleman
Applying machine learning to NASA EMIT satellite imagery to estimate global soil chemistry, combining satellite and field data and incorporating physical constraints to improve reliability and interpretability.
02
Geospatial Foundation Models & Mineral Prospecting
MIT MCSC: Advised by Dr. Evan Coleman, Prof. Elsa Olivetti & Prof. Sherrie Wang
Training models to predict mineral locations across continents by learning from geophysical data — fault lines, elevation, rock type, and soil surveys. Work has direct implications for critical mineral discovery and carbon capture.
03
Computer Vision & Embodied Perception
Fifth Set Analytics & GT Robot Learning Lab
Fine-tuning pose estimation models for automated sports officiating and human hand-tracking for robotic manipulation using depth cameras.
04 — Early Work
Astronomical Machine Learning
NASA JPL & NASA Exoplanet Watch
Early deep learning work on detecting exoplanets from photometric light curves using CNNs and LSTMs. Contributed to NASA's EXOTIC citizen science pipeline with 10 publications.
02
Education & Experience
Georgia Institute of Technology
B.S. Computer Science & B.S. Mathematics — Double Major
MIT Earth Intelligence Lab
Undergraduate Research Assistant
MIT Climate & Sustainability Consortium
Undergraduate Research Assistant
Fifth Set Analytics
Machine Learning Intern
GT Robot Learning & Reasoning Lab
Undergraduate Research Assistant
NASA Jet Propulsion Laboratory
Research Intern
NASA Exoplanet Watch
Student Researcher
- 2022–2026 Dean's Scholarship for College of Sciences, Georgia Tech
- 2022–Present Dean's List, Georgia Tech
- 2022, 2023, 2025 Faculty Honors, Georgia Tech
- 2021 Betty Neall Youth Award of Merit — East Bay Astronomical Society
- 2021 1st Place — Washington State Science & Engineering Fair
- 2021 NASA Earth Sciences Award — WSSEF
- 2021 Wolfram Research Award — WSSEF
- 2020 Select Interview — Research Notes of the AAS
- 2019 President's Award for Educational Excellence
03
Publications & Preprints
Highlighted Research
[14]
Earth2EMIT: Spatiotemporal Autoencoders for Orbital-Proximal Fusion
All-in-one soil property prediction pipeline fusing NASA EMIT hyperspectral satellite data with geospatial, atmospheric, and high-quality field measurements from YardStick PBC. Uses a physics-guided autoencoder that learns spectral-spatial embeddings to reconstruct soil spectra and predict soil properties.
[13]
Trained a U-Net ResNet to infill the presence of masked minerals based on surrounding mineral resources, geophysical data (fault lines, elevation, rock type), and agronomic data (RaCA). Model outperformed ViT and Kriging baselines with a test Dice coefficient of 0.31.
[12]
Used USDA RaCA data to reconstruct soil organic carbon spectra alongside SOC regression to improve out-of-distribution generalization. A physics-based pipeline provides an interpretable spectral signature of SOC and avoids generalization failures through fine-tuning.
Full Publication Record
Filter
[13]
[12]
[11]
[10]
[9]
[8]
[7]
Mid-Transit and Reference Star Analysis of HAT-P-37 b and Kepler-45 b
[6]
[5]
[4]
[3]
[2]
[1]
04
Methodology & Technical Stack
Deep Learning Frameworks
- PyTorch
- MMDetection / MMPose
- Timm (ViT, ResNet)
- Scikit-learn
Architectures & Methods
- U-Net / ResNet (Segmentation)
- Vision Transformers (ViT)
- Spatiotemporal Autoencoders
- CNN + LSTM Hybrids
- Masked Autoencoding (MAE-style)
- Physics-Constrained Networks
Data & Sensing
- NASA EMIT Hyperspectral Imagery
- Google Earth Engine
- USDA RaCA Soil Surveys
- USGS Geophysical Data
- Intel RealSense D435 (Depth)
Programming Languages
- Python
- R
- Java
- C / C++
- x86 & LC3 Assembly
Infrastructure
- AWS / Google Cloud
- Git / GitHub
- Jupyter / Colab
- JavaFX
Research Domains
- Geophysics & Soil Science
- Hyperspectral Remote Sensing
- Climate & Carbon Capture
- Astrophysics / Exoplanet Science
- Robotic Manipulation
- Sports Analytics