Projects
Machine Learning for Satellite Imagery
Vessel & Airplane Detection
I built deep learning systems that find ships and aircraft in medium-resolution satellite images.
- Adapted open-source object detection models for satellite data
- Worked with domain experts to design how we label training images, so the labels are high quality
- Delivered systems that detect and track objects at large scale in near real time
Semantic Segmentation for Roads & Buildings
I led a team that built segmentation models for roads and buildings. Segmentation means labeling every pixel in an image. The models work across the whole world.
- Built automated ETL pipelines (steps that collect and prepare data) and model training pipelines, using Kubeflow (a tool for running ML workflows)
- Deployed the models on Kubernetes (a system that runs software across many computers), with monitoring of their performance
- Supported analytics work for several teams
Change Detection
I built a complete pipeline that finds changes in infrastructure and land use over time.
- Analyzed satellite images of the same place from many dates, to see how it changed
- Released an open-source change detection dataset to support community research
- Connected the results to large language models (LLMs) that write a short summary of each event
Research & Development
Weakly Supervised Anomaly Detection
I designed and deployed a new method that finds unusual things in images (anomalies) using very few labels.
- Reached an F1 score of 0.85 (a score from 0 to 1 that balances finding real cases and avoiding false alarms), with low false-positive rates
- Led to a granted U.S. patent
- Published at the ICML 2019 Workshop on Climate Change and AI
Super Resolution for Satellite Imagery
I made satellite images sharper using deep generative networks. These are models that create new image detail.
- Modified ESRGAN (a well-known super resolution model) for remote sensing, which means studying Earth from satellites
- Designed custom loss functions (the rules a model learns from) that fit the textures and details of the ground
- Scaled the method to process imagery from the whole world