---

Live links

	URL
🛰️ Landing page	https://interactiveintel.github.io/edge-ai-satellite-triage/
🎛️ Live dashboard	https://interactiveintel-edge-ai-satellite-triage.streamlit.app/
📦 Source	https://github.com/interactiveintel/edge-ai-satellite-triage

---

Architecture

Image Tile ─► ImageIngestor ─► QuantizedInferencer ─► ObjectDetector ─► EdgeAgent ─► TriageResult
                │                       │                     │              │              │
            normalize/tile      MobileNetV3-Small      YOLOv8 / SAR CFAR   ReAct loop   audit + provenance
            any format →        INT8 via TRT/ONNX     "items of interest"  optional SLM HMAC-SHA256 log
            (C,H,W) f32         auto-build pipeline   dark-ship detection  <2W, 3 steps

Pipeline stages

1. Ingest — load TIFF / JPEG / numpy, tile-split, normalize to (C, H, W) float32
2. CNN Inference — MobileNetV3-Small backbone, INT8 quantized via ONNX / TensorRT. Outputs cloud_fraction, anomaly_score, value_score
3. Object Detection — YOLOv8-nano (optical) or classical CFAR (SAR). Detects vessels, vehicles, aircraft, fires
4. Agentic Reasoning — ReAct loop (pure Python, <2 W) activated on high-value tiles or when items of interest are present
5. Triage Decision — KEEP or FILTER with human-readable explanation, bandwidth savings estimate, and full provenance

Safety & hardening

• NaN/Inf guard at both inference and triage layers — bad sensor data defaults to safe FILTER
• SIGALRM-based inference timeout ceiling (MAX_INFERENCE_MS)
• Error recovery with safe fallback results at every pipeline stage
• Power budget enforcement via AgentPowerGuard
• HMAC-SHA256 authenticated audit log — tamper-evident

---

Quick Start

# Install (lightweight — no GPU required for dev/CI)
pip install -e ".[dev]"

# CLI demo (uses simulation stub)
PYTHONPATH=src python -m edge_triage

# Streamlit dashboard (recommended)
pip install -e ".[all,dev]"
PYTHONPATH=src streamlit run app.py

# Run tests (119 tests)
make test

---

Jetson Deployment

# Build Docker image (uses NVIDIA L4T base)
docker build -t edge-triage:latest .

# Or on Jetson directly
pip install -e ".[ml,jetson]"

# The model auto-downloads on first run:
# 1. MobileNetV3-Small pretrained weights from torchvision
# 2. ONNX export with dynamic axes
# 3. TensorRT INT8 engine cached via trtexec (Jetson only)

# Set power mode
sudo nvpmodel -m 1          # 15W mode
sudo jetson_clocks           # lock clocks for benchmarks

---

Project Structure

src/edge_triage/
  __init__.py            — package exports
  __main__.py            — 5-tab Streamlit dashboard + CLI entry point
  config.py              — power budgets, thresholds, model paths
  data_ingest.py         — image loading, tiling, normalization
  inference.py           — quantized CNN inference (TensorRT > ONNX > PyTorch INT8 > stub)
  detection.py           — YOLOv8 object detection (items of interest)
  ship_detector.py       — SAR ship detection (classical CFAR + AIS cross-reference)
  live_data.py           — Sentinel-1, Sentinel-2, NOAA GOES, NASA FIRMS feeds
  secrets_store.py       — local API-key vault with env-var fallback
  reasoning_loop.py      — pure-Python ReAct agent loop
  agent.py               — EdgeAgent with optional SLM support
  triage.py              — main pipeline engine (EdgeTriageEngine)
  audit.py               — HMAC-SHA256 authenticated JSON Lines audit log
  metrics.py             — bandwidth / power / TOPS-per-Watt tracking
  model_registry.py      — versioned model management with SHA-256 integrity
  utils.py               — PowerMonitor, AgentPowerGuard

scripts/
  train_cloud_mask.py    — EuroSAT + BigEarthNet training (MobileNetV3-Small)
  retrain_pipeline.py    — continuous retraining from analyst feedback or audit logs

docs/
  index.html             — public landing page (GitHub Pages)
  og-image.png           — social preview card
  favicon.png            — site favicon

tests/
  test_triage.py         — 119 tests across 22 test classes

app.py                   — Streamlit Cloud entry point
.streamlit/config.toml   — dark mission-control theme

---

Key Modules

Live satellite data (live_data.py)

Four free public sources wired in — no paid API required:

• Sentinel-1 SAR — global all-weather radar, vessel detection over open water
• Sentinel-2 L2A — global 10 m optical, 2-5 day latency
• NOAA GOES-18 — Americas near-real-time, ~10 min latency
• NASA FIRMS — global active-fire detections (needs free MAP_KEY)

SAR ship detection (ship_detector.py)

Classical CFAR + connected components — no ML training needed. Ships appear as bright scatterers on dark water in Sentinel-1 VV polarization. Cross-references with simulated AIS to flag “dark ships” (vessels with no transponder broadcast — the actual intel signal for maritime ISR).

Object detection (detection.py)

Pluggable backend: TensorRT/ONNX YOLOv8n → Ultralytics → heuristic stub. Detects vessels, vehicles, aircraft, smoke, fires. Detection results feed the agent’s ReAct reasoning.

Audit trail (audit.py)

Every triage decision is logged as HMAC-SHA256 authenticated JSON Lines. Tamper detection is built in — verify any log file with AuditLogger.verify_log(path). Key sourced from EDGE_TRIAGE_AUDIT_KEY env var or machine-derived fallback.

Model registry (model_registry.py)

File-based version registry (models/registry.json) with SHA-256 checksums, activation/rollback, and training metadata. Every deployed model is traceable to a training run and dataset.

make registry-list      # List registered models
make registry-verify    # Verify all model checksums

Continuous retraining (scripts/retrain_pipeline.py)

Ground-station feedback loop: load analyst corrections or audit-log pseudo-labels, fine-tune, validate against hold-out set, auto-register if improved.

python scripts/retrain_pipeline.py --feedback corrections.csv
python scripts/retrain_pipeline.py --from-audit logs/triage_audit.jsonl

---

CI/CD

GitHub Actions pipeline (.github/workflows/ci.yml):

1. Lint — ruff check + format
2. Test — pytest with coverage (JUnit XML + coverage XML artifacts)
3. Security — bandit source scan + pip-audit dependency scan
4. SBOM — CycloneDX Software Bill of Materials
5. Docker — build + Trivy vulnerability scan (main branch only)

---

Dashboard

The Streamlit dashboard provides five tabs:

Tab	Contents
Triage Pipeline	Mission scenarios, file upload, or synthetic tiles → per-tile result cards
Live Feed	Real Sentinel-1 / Sentinel-2 / GOES / FIRMS imagery with on-the-fly triage
Analytics	Score bar charts, decision distribution, power budget gauge, detail table
Audit Trail	Browse audit log entries, HMAC integrity verification
System Status	Backend detection, model registry, configuration dump

---

Dependencies

Group	Packages	Purpose
core	numpy, Pillow	Always installed — enough for stub inference and CI
[ml]	torch, torchvision, onnxruntime, opencv	Production inference
[dashboard]	streamlit, pandas, pydeck	Web UI
[jetson]	jetson-stats, tensorrt	NVIDIA Jetson hardware
[dev]	pytest, ruff, bandit, pip-audit	Development and security

---

Performance Targets

Metric	Target	Status
Total inference power	<15 W	✓ design
Agent reasoning power	<2 W	✓ pure Python
TOPS/Watt (INT8 DLA)	>40	requires Jetson hardware
Bandwidth reduction	>85%	✓ verified on 24-tile scenarios
Per-tile latency	<50 ms	✓ benchmarked
Test coverage	119 tests, 22 classes	✓ all passing

---

License

MIT — see LICENSE.

---