My work sits at the intersection of machine learning, computer vision, and remote sensing, with deep collaborations across health, sustainability, and humanitarian partners. I co-lead Microsoft’s Geospatial Machine Learning Center, and also do fundamental research on generalization, transfer learning, conditional computation, and domain adaptation. Below is a snapshot of the projects I have led or co-led — see Publications for the full record.
A PyTorch domain library for geospatial data — datasets, samplers, transforms, and pretrained models that bring the ergonomics of torchvision to Earth observation. Adopted across academia and industry for remote-sensing deep learning.
A living atlas of the world’s utility-scale solar and onshore wind, mapped quarterly from satellite imagery with deep learning. Built with Planet and The Nature Conservancy to track the global energy transition and inform low-conflict siting.
Deep-learning system to triage retinopathy of prematurity (ROP) from neonatal fundus images, clinically validated to extend specialist screening into low-resource settings where pediatric ophthalmologists are scarce.
An AI dataset of every utility-scale solar installation in India, built from satellite imagery to inform renewable-energy policy and land-use planning. Joint work with The Nature Conservancy.
Open data and ML pipelines for monitoring the Hindu Kush Himalaya glaciers — supported by an $18,000 Microsoft AI Research Grant and featured by Microsoft On the Issues.
Detecting beached whales, cattle, and elk in very-high-resolution satellite imagery to support biodiversity monitoring and conservation partners.
Interactive systems where humans and ML models collaborate to produce robust land-cover maps that generalize across geographically diverse regions.
A drop-in normalization layer that revisits local normalization for dense-prediction networks, improving generalization and robustness in semantic segmentation.
Building maps of vulnerable communities in Uganda from satellite imagery to support flood and pandemic response operations.
Self-supervised pretraining and foundation models tailored to multispectral, multi-temporal Earth-observation data — see TorchGeo.
Rigorous evaluation of purported advances in remote-sensing change detection (arXiv 2024); temporal cluster matching for structures (COMPASS 2021).
PSMA PET/CT lesion segmentation; preventing shortcut learning in COVID-19 chest X-ray classifiers; federated and privacy-preserving training.
Defenses for deep models in remote sensing and imaging modalities beyond the visible spectrum (MILCOM 2018).
Architectural innovations that improve efficiency and generalization in dense-prediction tasks.
Fusion of hyperspectral data with image-based 3D models; spectral–elevation registration; SWIR terrain segmentation.
Code, datasets, and benchmarks I have led or co-led — building blocks the geospatial-ML community can reuse.
