Research

Change Detection in Satellite Imagery

I study how to find and measure changes over time in satellite data.

• Designing models that detect small changes over time in land use and infrastructure
• Combining data from different sensors (optical images, and radar called SAR) to get more accurate and reliable results
• Main uses: environmental monitoring, measuring the impact of disasters, and tracking how cities grow

Weakly Supervised Learning

I work on making deep learning succeed with very little supervision. Supervision means labels that tell a model the right answer.

• Building methods that learn from labels that are few, noisy, or incomplete
• Main uses: anomaly detection (finding unusual things) and semantic segmentation (labeling every pixel)
• Goal: cut the work of labeling data while keeping models accurate at large scale

Visual Anomaly Detection

• Invented a new method that finds anomalies in images using few labels (weak supervision)
• The core ideas led to a granted U.S. patent
• The method worked well across different industrial and quality control uses

Multi-modal Learning

• Studied ways to combine satellite images with other kinds of data
• Built pipelines that merge the different data types, to make predictions better and give them more context
• Applied the work to areas such as ecological monitoring and fast disaster response