Compact PyTorch → ONNX → INT8 pipeline for onboard inference on EO small-sat/CubeSat platforms. Includes an Earth Observation tile triage example that mirrors a PhiSat-2 style onboard selection step. Runs locally on a laptop or a small SBC. See the runnable demo in examples/phi2-eo-tile-filter.

• Train a tiny CNN on small tiles.
• Export FP32 ONNX and quantize to static INT8 (QDQ) with ONNX Runtime.
• Benchmark latency and memory on the ORT CPU Execution Provider.
• Calibrate a confidence threshold to hit a target recall.
• Filter tiles for event-triggered downlink to save bandwidth.
• Ship assurance hooks: watchdog, telemetry log, rollback, and a run summary.

• End-to-end path from training to a compact INT8 artifact suitable for onboard execution (PyTorch → ONNX → ORT INT8).
• Calibrated confidence gate to meet a target recall, letting you trade downlink volume against science yield.
• Assurance hooks: watchdog, fallback-to-downlink on low confidence, rollback to last-known-good, and JSONL telemetry (latency + hashes) for audit.
• Deterministic, CI-checked demo with fixed seeds and small scripts; easy to port to HIL or ARM/aarch64 ORT.
• Produces measurable stats (accuracy, confusion, latency, bandwidth saved) to support requirement verification.

• Minimal dependencies (PyTorch, ONNX, ONNX Runtime; a few small extras).
• Clear scripts: train, export_onnx, quantize_ptq, infer_onnx, bench_onnxruntime, bandwidth_filter.
• Self-contained synthetic dataset for quick runs; swap to Sentinel-2 crops and recalibrate.
• CI workflow that smoke-tests the pipeline on push and PR.
• Seeded runs and JSONL logs for reproducibility and audit.
• Laptop or small SBC ready (ORT CPU EP) with a path to ARM/aarch64.
• Assurance utilities in assurance/ (watchdog, telemetry, summarizer, rollback).

This repository reflects workflows used in missions with onboard processing, such as ESA’s PhiSat-2 CubeSat for AI-enabled EO. The example filters tiles onboard and prioritizes downlink of useful data. Mission page: https://earth.esa.int/eogateway/missions/phisat-2

cd examples/phi2-eo-tile-filter
python -m venv .venv && source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt onnxscript==0.1.0 scikit-learn

# data → train → export → quantize → eval
python -m data.synth --out ./tiles --n 200 --bands 3 --size 64
python -m src.train --data ./tiles --epochs 12 --base 32 --lr 0.003
python -m src.export_onnx --weights runs/tinycnn.pt --out models/tinycnn_fp32.onnx --bands 3 --size 64 --base 32
python -m src.quantize_ptq --onnx models/tinycnn_fp32.onnx --calib ./tiles/val --out models/tinycnn_int8.onnx --size 64 --bands 3
python -m src.infer_onnx --onnx models/tinycnn_int8.onnx --data ./tiles/val --size 64 --bands 3

# calibrate → filter → telemetry → report
python -m src.calibrate_threshold --onnx models/tinycnn_int8.onnx --data ./tiles/val --target_recall 0.95 --out calibration.json
mkdir -p logs
python -m src.bandwidth_filter --onnx models/tinycnn_int8.onnx --data ./tiles/val --calibration calibration.json --downlink_out downlink --log logs/downlink.jsonl
THR=$(python -c "import json; print(json.load(open('calibration.json'))['threshold'])")
python ../../assurance/telemetry_log.py --onnx models/tinycnn_int8.onnx --data ./tiles/val --out logs/val.jsonl --threshold "$THR"
python ../../assurance/summarize.py --val_log logs/val.jsonl --downlink_log logs/downlink.jsonl --val_dir tiles/val --calib calibration.json --out_dir reports
cat reports/summary.md

• logs/downlink.jsonl decisions used by the bandwidth filter
• logs/val.jsonl per tile telemetry with probabilities and latency
• reports/summary.md and reports/metrics.json summary for quick review
• models/tinycnn_fp32.onnx and models/tinycnn_int8.onnx artifacts

Numbers from the latest synthetic run in examples/phi2-eo-tile-filter.

See assurance/.

• watchdog.py restarts a failing inference command a few times
• telemetry_log.py emits one JSON line per tile with hashes and timings
• summarize.py converts logs into a small Markdown and JSON metrics report
• rollback.sh swaps back to the last good FP32 model

.
├─ assurance/                   # watchdog, rollback, telemetry, summary
├─ examples/
│  └─ phi2-eo-tile-filter/
│     ├─ data/                  # synthetic tiles
│     ├─ src/                   # train, export, quantize, infer, bench, filter
│     ├─ logs/ and reports/     # created by the quick start
│     └─ models/                # TinyCNN
└─ .github/workflows/ci.yml     # smoke test pipeline

• Swap synthetic tiles for Sentinel-2 crops and recalibrate the gate for the same recall target.
• Add OpenVINO or TensorRT export paths for specific hardware.
• Log exact downlink bytes and confidence histograms for fuller telemetry.
• Replace TinyCNN with a stronger model and keep the PyTorch → ONNX → INT8 interface stable.

• Python 3.12
• torch ≥ 2.2, onnx 1.19.1, onnxruntime 1.23.2
• numpy, Pillow, scikit-learn, psutil, pytest

If you use or adapt this repository, please cite

Kaczmarek, S. (2025). PhiSat-2 Trustworthy Onboard AI. Zenodo. https://doi.org/10.5281/zenodo.17567181

BibTeX

@software{Kaczmarek_2025_PhiSat2_Onboard_AI,
  author    = {Sylvester Kaczmarek},
  title     = {{PhiSat-2 Trustworthy Onboard AI}},
  year      = {2025},
  publisher = {Zenodo},
  url       = {https://github.com/sylvesterkaczmarek/phisat2-trustworthy-onboard-ai},
  doi       = {10.5281/zenodo.17567181}
}

MIT. See LICENSE.

© Sylvester Kaczmarek · https://www.sylvesterkaczmarek.com