This page was last updated in August 2022.

Machine Learning

My master’s thesis project applies rotation-equivariant convolutional neural networks (CNNs) to detecting clouds in satellite imagery. I expect to submit my thesis and receive my master’s degree in January 2023. I plan to open-source all of my thesis code when the project is complete – currently, I’ve open-sourced my implementation of focal loss for PyTorch.

In 2021, I worked on developing cloud-detection algorithms for BeaverCube-2, a 3U CubeSat scheduled for launch in 2023 and designed and built by MIT Space, Telecommunications, Astronomy, and Radiation Lab (also known as STAR Lab). I contributed to a conference paper comparing the performance of rule-based methods (luminosity thresholding and random forest detection) to deep learning (specifically, a U-Net) for detecting clouds in visible-spectrum and long-wave infrared satellite imagery.

In 2022, I joined a collaboration that tested algorithms designed for BeaverCube-2 onboard OPS-SAT, a 3U CubeSat operated by the European Space Agency (ESA), and ultimately tested my code on-orbit and on the OPS-SAT engineering model, and contributed to a conference paper. During this project, I forked the popular ranger library for random forests and extended it to work for object detection in images. I open-sourced my implementation.

Controls

Since June 2020, I have been working on the attitude determination and control system for BeaverCube, an underactuated (magnetorquer-only) 3U CubeSat with onbord propulsion designed and built by MIT STARLab. Along with one other student, I wrote all of BeaverCube’s attitude determination and control flight software, including an Estimated Kalman Filter (EKF) in Python for attitude determination, a nonlinear optimization program to plan underactuated slews in C++, and a state machine in Python that calculates specific pointing goals for BeaverCube in different pointing modes. BeaverCube is the first satellite to use nonlinear optimization for magnetorquer-only control, and BeaverCube has more onboard autonomy than most other CubeSats. BeaverCube launched in July 2022 to ISS orbit.

In 2020, I contributed to a conference paper on BeaverCube’s overall design and engineering approach.

In 2022, I began working on BeaverCube-2’s attitude control system, which inherits most components and software from BeaverCube, but adds a single reaction wheel to improve attitude control. I contributed to a conference paper (manuscript not yet available online) comparing attitude control systems with three magnetorquers and one reaction wheel to magnetorquer-only systems.

Also in 2022, I did an unpublished class project on using sums-of-squares optimization to prove stability of planned slews for magnetorquer-only, magnetorquer + one reaction wheel, and fully actuated satellites. My final presentation and final paper are publically available.

Collocating Satellite Data

In 2021, I began working on a project that leverages satellite orbital dynamics to efficiently collocate data from nadir-scanning satellites (such as satellites with nadir-sounding microwave radiometers) with data from other satellites. In 2022, I presented on my work at the Remote Sensing Data Management Technologies in GeoScience Workshop, where I won the Best Presentation award. I’m currently preparing a first-author paper on this project.