On-Orbit Image Classification

About This Project

Deploy a TinyML model on the satellite's onboard processor to classify captured images (clouds, land, ocean, fire) before downlink. Reduce data transmission by only sending images that meet classification criteria.

Overview

Satellites generate far more data than they can transmit to the ground. A typical low Earth orbit pass over a ground station lasts only five to twelve minutes, creating a severe bandwidth bottleneck. On-orbit image classification addresses this by letting the satellite decide which images are worth sending before it ever opens the downlink. A lightweight machine learning model runs directly on the satellite's processor, examining each captured frame and scoring it for scientific or operational value — is this clouds or clear sky? Land or ocean? Does this frame contain a feature of interest, or is it blurry and useless? Only the highest-scoring images get queued for downlink, potentially reducing transmitted data volume by half or more. This concept was proven at professional scale by the European Space Agency and is now one of the fastest-growing capabilities in the commercial satellite industry. A student version uses simplified models and lower-resolution imagery but demonstrates the same fundamental principle: intelligent data triage at the edge. The project spans hardware integration, model training, embedded deployment, and statistical validation — making it genuinely interdisciplinary.

Technical Details

Use TensorFlow Lite Micro on an ESP32-S3 (~$10, 0.5W, dual-core 240 MHz, native vector instructions) with ArduCAM OV2640 (2MP, ~$15, SPI/I2C). Train models using Edge Impulse no-code pipeline — classify telemetry patterns (power bus anomalies, temperature spikes) or run quantized MobileNet-v2 for image classification. INT8-quantized models fit in ESP32-S3 512 KB SRAM. Watchdog timers and periodic cold-restarts mitigate radiation-induced latch-ups. Stretch goal: compare on-orbit inference accuracy against ground-truth processing.

Research & Notes

ESA PhiSat-1 (2020) demonstrated cloud-detection CNN on Intel Movidius Myriad 2, filtering hyperspectral imagery at ~1W and reducing downlink volume by over 50%. PyCubed ATSAMD51 showed latch-up susceptibility during 2024 solar maximum — independent watchdog timers essential for COTS processors. Edge Impulse platform makes ML accessible to freshmen with no-code training pipeline. Entire payload draws under 2W, masses under 100 g. Cost: $200-$1,500. On-board AI inference is the fastest-growing capability in the CubeSat industry. Complexity: low-to-medium thanks to Edge Impulse and Arduino ecosystem. Publishable results if comparing on-orbit vs ground-truth accuracy. Tier 1 recommendation — high industry relevance, low hardware barrier.

Required Disciplines

This project spans 3 disciplines, making it suitable for interdisciplinary student teams.

Available At

This project is available at the following Blackwing chapters:

Next Steps

Ready to take on this project? Here's a general roadmap that applies to most CubeSat missions:

1. Build your foundation: Complete the core modules in the CubeSat Academy to understand spacecraft subsystems, mission design, and development workflows.
2. Form a team: Recruit students across the required disciplines and identify a faculty advisor. Plan for knowledge transfer between graduating and incoming members.
3. Write a mission concept: Draft a 1–2 page document outlining your objectives, target orbit, payload requirements, and success criteria.
4. Connect with a chapter: Join a Blackwing chapter for mentorship, shared resources, and access to the platform ecosystem.
5. Explore the developer tools: Visit the Developer Portal for platform documentation, SDKs, and hardware specs.
6. Plan your timeline: Map milestones to your academic calendar. Most projects align well with a 2–4 semester capstone or research sequence.
7. Reach out: Contact us to discuss your project goals, platform selection, and path to orbit.