High-stakes regions still lack frequent, automated insight into what's growing — and the pipeline that should deliver it is constrained at every step from acquisition to analyst.
High-stakes regions lack real-time insight on farmlands — making operational decisions always reactive rather than proactive. The data exists; the timing doesn't.
A single multispectral tile is ~1 GB over X-band. Bandwidth forces a permanent trade-off between coverage, resolution, and timeliness.
Ground-side cloud processing adds hours to a full day from capture to analyst — too slow for disease, storm damage, or sudden-stress response.
Raw multispectral imagery → actionable agricultural insight before the next ground contact. Three engineering commitments hold the proposal together.
Each module cuts the data before handing off to the next. ~1 GB enters; ~5 KB leaves. The satellite becomes a filter.
LTAE asks "what crop?" — order-agnostic. CNN-LSTM asks "where in season?" — direction matters. Concurrent CUDA streams on Orin Nano · combined ~3 ms / point TensorRT INT8.
FP32 → FP16 → INT8 staged compression. Lean uplink AND lean inference — 3.5× reduction, validated at every step on full and hard-confidence subsets.
Three constraints ground-based ML never faces — power, radiation, and bandwidth — wrapped by a coordinated defense posture and a measured resource fingerprint.
| Resource | Peak | Note |
|---|---|---|
| System RAM | ~500 MB | 6.3% of 8 GB |
| CPU (6-core A78AE) | 60–80% | Feature-engineering bound |
| GPU (Ampere) | 15–25% | Sub-second burst · headroom |
| NVMe storage | ~12 GB | of 64–256 GB available |
| ROM | ~15 MB | Radiation-hardened |
| Active power | 5–8 W | within 7 W TDP1 |
| Stage | Time |
|---|---|
| Radiometric calibration | 2.0 s |
| Cloud screening | 0.5 s |
| AOI crop + DEM ortho | 3.0 s |
| Point extraction | 200 ms |
| Feature engineering | 30.0 s |
| Dual inference | 0.5 s |
| JSON packaging | 99 ms |
Hundreds of bit-flips/day in LEO2 · three independent layers wrap every inference cycle.
Same pipeline, swap weights — Haumea generalizes across mission domains. Three decision categories where same-orbit insight changes the outcome, plus three integration patterns into real-world decision pipelines.
| Mission | Adaptation | Pipeline modules |
|---|---|---|
| Forestry | Re-train weights | Unchanged |
| Water resources | Re-train + spectral indices | Modules 1–4, 6 unchanged |
| Urban observation | Re-train + AOI DB | Modules 1–4, 6 unchanged |
| Disaster response | Re-train + trigger logic | Modules 1–4, 6 unchanged |
Footprint differentiator — Three-Body Computing Constellation3 runs an 8B-param foundation model (~600 MB) on flagship platforms. Haumea is 1.2 MB · 2,000× smaller · fits every CubeSat in a constellation.
Three decision categories where same-orbit insight changes the outcome.
Disease, drought stress, storm damage, pest invasion. Flagged on the pass that observed it — queued HIGH-PRIORITY for the next ground contact.
Days → hours latencyIrrigation, fertilizer windows, harvest readiness. Per-pass phenology tracking triggers on the actual transition — not on the calendar.
+3–7 day lead timeCrop type maps, yield forecasts, regional productivity. Continuous in-orbit classification at constellation scale.
Annual → continuousAnchored in real demand — NE China, US Midwest, Ukraine wheat belt, Brazil soybean face the same challenges, the same need. The 1.2 MB footprint makes per-region deployment economical.
The orbital filter lets downstream systems act directly on insights — saving time and effort.
Georeferenced insights via GNSS (BeiDou, GPS, Galileo) at cm accuracy. Sprayers, tractors, irrigation — zero manual routing.
Precision ag deployments globallyInsights feed markets, governments, and insurers. Fits any structured-data pipeline.
USDA · EU CAP · yield forecastingAnomalies routed to field teams in near-real-time. Disease, storm, stress — rapid intervention.
FAO locust watch · crop insurance · ERNsThree-Body demonstrated that real-time on-orbit EO analysis is operationally viable — identifying stadiums and bridges through heavy snow cover across NW China.4 Haumea extends the same workflow class, multi-region deployable, but 2,000× smaller — tuned for specific high-value tasks, and fits every CubeSat in a constellation.
From validated INT8 deployment to persistent agricultural intelligence in orbit.
INT8 on Orin Nano. Hardware-in-the-loop testing — power, thermal, radiation, pipeline timing. Bench-validated radiation strategy.
Payload reference design for CubeSatHosted payload on AI-capable CubeSat. Demo over target agricultural region. Telemetry-validated accuracy and latency.
Published in-orbit performance metricsMulti-constellation, multi-region, multi-mission. Agencies, platforms, insurers — daily-revisit cadence.
Continuous crop intelligence as a serviceFour parallel directions for technique evolution.
Distill from models like Prithvi-EO5 / RemoteCLIP into the compact dual architecture. Improved zero-shot generalization to new geographies.
Add Sentinel-1 SAR for cloud-penetrating coverage — critical for regions like NE China during summer cloud season.
If needed, push quantization further to maximize efficiency and improve resource requirements on Orin Nano Tensor Cores.
Each satellite contributes regional gradient deltas via inter-satellite laser links — models updating weights in orbit.
Five external anchors underpin every measured claim and external comparison in this proposal.