The Center for Data Innovation recently spoke with Michael Stanway, CEO of The Compression Company, a San Francisco–based startup developing AI-powered software to speed up data transmission from Earth‑observation satellites. Stanway explained how the company’s system runs directly on satellites, using trained AI models to detect patterns and flagging relevant features in captured images, and a compression engine using those signals to reduce the volume of data that needs to be sent to the ground.
David Kertai: What role does your technology play in the satellite data pipeline?
Michael Stanway: Earth observation satellites constantly capture images of the planet to monitor weather, infrastructure, and natural disasters. But today’s high-resolution cameras generate enormous data files, and satellites get only short communication windows to transmit them. That imbalance causes data to pile up on the satellite, limiting their availability to gather new information, and delaying access to time sensitive data that organizations, such as weather agencies and emergency response teams rely on to make fast decisions about storms, wildfires, and other dangerous conditions.
We address this problem by reducing data size before transmission. Our AI-powered software—integrated into the satellite’s existing computer system—analyzes each image in real-time to identify which regions contain essential detail and which areas are repetitive or predictable. A dedicated compression engine then uses those AI-generated signals to encode the data more efficiently while preserving the information needed for downstream analysis. This improves how much data the satellite can send without requiring upgrades to onboard hardware like radios or processors, and a paired ground decoder reads the compressed data and converts it back into a standard image format once it reaches the ground.
Kertai: How does your compression technology work?
Stanway: Traditional compression methods use one‑size‑fits‑all formulas that treat every image the same, breaking them into patterns to shrink file size but failing to adapt to the unique characteristics of different sensors, such infrared cameras, radar systems that see through clouds, or multispectral cameras that record light our eyes can’t see.
Our approach learns directly from this data. We train our AI models on large sets of past images from specific sensors so they can recognize recurring patterns, much like a person learns to identify features after seeing many examples. When a new image is captured, the model identifies which regions contain important details. For instance, broad areas of ocean or cloud cover often contain similar visual information, which the model can encode more efficiently than complex regions like cities. The compression engine then uses the models’ guidance to encode those repetitive regions more compactly while preserving critical details where it matters. This improves rate‑distortion performance, meaning that it reduces file size while maintaining image quality.
Kertai: How do you ensure compression efficiency speed while avoiding increased computational cost?
Stanway: Satellites have limited computing power, so efficiency is crucial. We design our models to run on existing onboard hardware—such as compact NVIDIA processors—and ensure they operate fast enough to keep pace with the satellite’s camera in real-time. To avoid overloading the satellite’s computer system, we split the workload: a lightweight encoder on the satellite compresses data quickly, while a more powerful decoder on the ground handles the heavier reconstruction. This minimizes onboard computing demands while preserving the details analysts need.
Kertai: What industries benefit most from your solution?
Stanway: Satellite operators and imagery providers benefit first because they work with massive image files and need them delivered quickly for tasks like crop monitoring and disaster response.
But the same problem—too much data and too little bandwidth—shows up in many other sectors. Industries that run large sensor networks, track activity in real-time, or build detailed 3D maps often face similar bottlenecks; for example, systems that watch for vibrations along fiber‑optic cables, defense networks that follow moving objects, or LiDAR tools that scan terrain. Anywhere organizations need to move high‑quality data through a limited connection, this kind of compression can make the whole system faster and more efficient.
Kertai: What is the biggest challenge you have faced so far?
Stanway: Convincing satellite operators and data-service providers to adopt new onboard software is our biggest hurdle. Even with clear performance gains, organizations are often cautious about changing core systems that handle critical data. They worry about risks such as data loss, compatibility issues, or whether new software will hold up in harsh space conditions. That’s why we focus on demonstrating reliability in real‑world environments and have planned an in‑orbit demonstration in March 2026 to show that our software performs reliably on an active satellite and delivers measurable improvements in data transmission and quality.
