Near-real-time insight delivery pipeline for earth observation satellite images

Tao, Bill

Permalink

https://hdl.handle.net/2142/129256 Copy

Description

Title

Near-real-time insight delivery pipeline for earth observation satellite images

Author(s)

Tao, Bill

Issue Date

2025-04-28

Director of Research (if dissertation) or Advisor (if thesis)

Vasisht, Deepakv

Doctoral Committee Chair(s)

Vasisht, Deepakv

Committee Member(s)

• Gupta, Indranil
• Nahrstedt, Klara
• Kumar, Swarun

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Satellite Networking
• Remote Sensing

Abstract

The emergence of the new generation of small yet computation-capable low-earth-orbit (LEO) satellites for earth observation (EO) has created opportunities for humanity to monitor events that happen in every corner of the globe and even use the emerging Computer Vision (CV) models to generate insights from these satellite images. However, the current pipeline suffers from multiple challenges, including the limitation on both computing and energy capacity of the satellite hardware, and the limited network capacity in the whole pipeline. This dissertation mitigates these challenges by delivering a cross-layer end-to-end optimization solution. It focuses on answering this key question: what can be done in EO satellite systems to help it satisfy the real-time requirements from users as well as the tight energy, compute and networking budgets? To answer this question, this dissertation built 2 major systems: Umbra system for optimally using satellite link capacity to avoid congestion, Serval system for distributed edge computing on EO satellite systems to reduce latency as well as reduce computation costs. This combination of systems greatly enhances the quality of the satellite image & insight delivery system, while greatly boosting the efficiency within the whole pipeline end to end. It showcases the great benefit of a detailed understanding of the underlying physical dynamics as well as unique patterns when designing novel systems in unseen environments. By modeling the orbit dynamics of satellites, Umbra and Serval were able to predict the future activities of satellites and proactively schedule compute and network transmission. As a result, Umbra can reduce the P90 latency for all network traffic by 3.5x, while Serval can reduce the median latency for the critical data from 3 days to 2 minutes.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129256

Copyright and License Information

Copyright 2025 Bill Tao

Owning Collections