How CubeSats Use Edge AI to Filter Cloud-Covered Images

Earth-observation CubeSats generate vast amounts of image data, much of which is rendered useless by clouds blocking the view of the ground. Historically, these small satellites transmitted every captured photograph to Earth, wasting limited bandwidth, time, and onboard power. Today, CubeSats utilize onboard edge computer vision artificial intelligence to analyze photographs immediately after capture. By identifying and discarding cloud-obscured images directly in orbit, satellites only transmit clear, valuable data down to ground stations, drastically improving mission efficiency.

CubeSats are miniature satellites built in standard, compact units. Because of their small size, they have strict limits on battery power and antenna strength. Furthermore, a CubeSat can only transmit data when it passes directly over a ground station, giving it a communication window of just a few minutes per orbit.

Earth is covered by clouds roughly 67% of the time. Without onboard filtering, up to two-thirds of the data transmitted during these brief communication windows consists of useless photos of clouds. Downlinking this unusable data wastes electricity and clogs transmission pipelines.

How Onboard Edge AI Works

Edge computing solves the downlink bottleneck by processing data right where it is gathered—on the satellite itself—rather than sending it to Earth for analysis. The process follows four main steps:

  1. Image Capture: The CubeSat's optical sensors capture a raw photograph of the Earth's surface.
  2. Preprocessing: The onboard computer quickly formats the raw sensor data, often shrinking the image or separating it into specific light bands (such as visible and infrared light) to make it easier for the AI to read.
  3. AI Inference: The image is fed into a pre-trained computer vision model, typically a lightweight Convolutional Neural Network (CNN). This AI model has been trained on thousands of previous satellite images to recognize the distinct textures, brightness, and thermal signatures of clouds compared to land, water, and snow.
  4. Scoring: The AI assigns a "cloud cover score" to the image, calculating the exact percentage of the photograph obscured by clouds.

Hardware Behind the AI

Running artificial intelligence in space requires specialized hardware. Standard computer processors consume too much power and can be damaged by space radiation. CubeSats rely on low-power, radiation-tolerant Vision Processing Units (VPUs) and Field Programmable Gate Arrays (FPGAs).

These microchips are designed specifically to run neural networks while consuming only a few watts of electricity—less power than a standard household lightbulb. They allow the satellite to process high-resolution images in milliseconds without draining the satellite's solar batteries.

The Discard Decision

Once the AI assigns a cloud cover score, the satellite's flight software makes an automated decision based on rules set by mission control.

If an image has a cloud cover score above a certain threshold—for example, 80% cloud cover—the onboard computer immediately deletes the file from its memory. If the image falls below the threshold, it is compressed, stored, and placed in a priority queue for the next available downlink window.

Benefits for Earth Observation

By discarding cloudy images in space, CubeSats maximize their operational value. Ground stations receive only actionable data, which speeds up the response time for critical applications like tracking wildfires, monitoring agricultural health, and mapping flood zones. Additionally, the freed-up bandwidth allows satellites to capture and send more images of clear targets, effectively multiplying the scientific return of a single CubeSat mission.