HAUMEA
SPACE INTELLIGENCE · ZERO HUNGER

Crop intelligence,
on-orbit.

Two neural networks · Sentinel-2 · 1.19 MB · real-time

0.000 Crop F1
0.000 Phenology F1
0.00 MB · INT8
<10 ms · CPU
5–8 W · active
10⁶× BW reduction
§ 01 · PROBLEM

Three converging pressures on food security.

90%

of space-generated data never reaches an analyst.

A single Sentinel-2 tile is ~1 GB. Bandwidth is the bottleneck — most of what a satellite sees never makes it back in time to act on.

715Mt

China's 2025 grain output — 2nd consecutive record.2

NE China provides 70% of national grain growth. Real-time monitoring is an explicit 14th/15th Five-Year Plan priority.

  • Japonica rice · NE China30 – 50%
  • Soybean · national41%
  • Corn · national34%
§ 02 · SOLUTION

Small, fast, satellite-deployable.

Two specialized networks → one compact JSON per inference. Insight, not imagery.

Most current systems

Ground-based, full-precision

Where it runsCloud / data-centre
Model size~ 100 – 850 MB
Per-tile downlink≈ 1 GB raw imagery
Latency to insightHours to days
Power per inferenceCloud-tier GPU
Our baseline

FP32 dual architecture

Where it runsEdge accelerator
Model size4.28 MB
Per-inference downlink< 1 KB JSON
Latency to insight~ 8 ms / sample
Crop F1 / Stage F10.994 / 0.961
Our deployment · INT8

Quantized dual architecture

Where it runsOn-orbit, satellite payload
Model size1.19 MB
Per-inference downlink< 1 KB JSON
Latency to insight~ 8 ms / sample
Crop F1 / Stage F10.994 / 0.961
02.A · MODULAR

Crop and stage independently deployable

  • LTAE alone = ~190 KB (crop-only)
  • Each model updates via separate uplink
02.B · HETEROGENEOUS

CPU handles I/O, accelerator handles math

  • Forward passes route to GPU/DLA when available
  • Falls back to CPU-only; no output lost
02.C · FAULT-TOLERANT

Missing bands, sparse passes — still valid output

  • Missing TIFFs handled at extraction time
  • Single-model fallback if one checkpoint corrupts
§ 03 · MODELS

Two specialists. One decision.

LTAE: order-agnostic attention for crop type. CNN-LSTM: bidirectional recurrence for phenology. Click any layer to inspect.

§ 04 · COMPRESSION

4.28 MB → 1.19 MB. Zero accuracy loss.

FP32 → FP16 → INT8 staged compression. KL ≈ 0.002, prediction flip rate 0.02%.

§ 05 · ARCHITECTURE

Acquisition to insight, on-orbit.

01 · Input
Sentinel-2
12 bands · multi-temporal acquisition
02 · Preproc
Feature Extraction
70 spectral cols · indices · DOY
03 · Dual AI
LTAE + CNN-LSTM
Crop · phenology stage
04 · Encode
JSON Result
< 1 KB compact payload
05 · Downlink
Ground Station
Insight, not raw data

Modular · Heterogeneous · Fault-tolerant — each stage can be swapped or skipped under hardware constraint. Target: Jetson Orin or equivalent edge accelerator.

§ 06 · FEASIBILITY

Resource footprint, measured.

Validated on conservative CPU. Jetson Orin INT8 Tensor Cores yield 2–4× additional speed-up.4

FOOTPRINT TABLE
METRIC CURRENT EO OUR BASELINE OUR DEPLOYMENT
Model size100 – 850 MB4.28 MB1.19 MB
LTAE latency · CPUN/A · ground2.42 ms1.72 ms
CNN-LSTM latency · CPUN/A · ground5.72 ms6.16 ms
Crop F1 / Stage F1≈ 0.95 / 0.920.994 / 0.9610.994 / 0.961
KL divergence vs FP32—00.002
Uplink for model update100 – 850 MB4.3 MB1.2 MB

Latency on Colab CPU is conservative — no AVX-512 / VNNI. Jetson Orin Tensor Cores are expected to deliver 2 – 4× speed-up on the INT8 path.4

ON-ORBIT PROFILE
POWER
5–8 W
active inference
Jetson Orin 7W mode
LATENCY
<10 ms
LTAE + CNN-LSTM
CPU baseline
MODEL UPLINK
4.7 s
1.2 MB @ 256 KB/s
S-band
MODEL SIZE
1.19 MB
INT8 deployed
vs. 850 MB ground
SIZE VS. ALTERNATIVES
Foundation model ~600 MB
Generic RS-CNN ~50 MB
Ours — INT8 1.19 MB
§ 07 · RADIATION

Bit-flips happen. Hardened against them.

07.A · LAYER-AWARE

RedNet selective redundancy

  • SEU-sensitive layers protected; others run lean5
  • Validated on Chaohu-1 SAR · Jetson Xavier NX
≈ 0 residual error rate
07.B · PASSIVE DEFENSE

Quantization-induced robustness

  • INT8 weights = ¼ the bit-width of FP32 — 4× fewer vulnerable bits
  • Compression pipeline doubles as radiation defense at zero added cost
4× fewer vulnerable bits
07.C · WATCHDOG

Periodic weight reload from ROM

  • Reference checkpoint in radiation-hardened ROM
  • Checksum every inference cycle; reload on mismatch
100% ROM-verified weights
§ 08 · INNOVATION

Three things done differently.

08.A · DUAL ARCHITECTURE

Crop and phenology need different inductive biases

  • LTAE: order-agnostic — cloud gaps don't break inference
  • CNN-LSTM: bidirectional — Greenup vs. Dormancy differ by temporal slope, not NDVI
Decoupled loss +0.5% accuracy
08.B · COMPRESSION DESIGN

Heavy training, lean inference — engineered for orbit

  • FP32 → FP16 → INT8, validated at each step
  • Same model class as ESA's Φ-sat-26
3.6× smaller · zero F1 loss
08.C · GENERALIZATION

Every decision made knowing validation is unseen regions

  • REP channel fix: +1.0% accuracy
  • Window-aware augmentation: stage F1 lift
  • Ensemble & TTA measured to hurt on unseen — dropped
Single-checkpoint · simpler · more robust
§ 09 · SHIFT

Satellite as filter, not sensor.

Today · satellite as sensor

Stream everything, interpret on the ground

1 GB
raw tile downlinked per acquisition
  • Bandwidth-limited — minutes per tile
  • All processing happens on the ground
  • Latency-bound: hours to days, observation → insight
Ours · satellite as filter

Only insights leave orbit

< 1 KB
JSON insight downlinked per inference
  • ≈ 10⁶× downlink reduction per useful inference
  • Onboard inference — insights ready before the downlink window
  • Same orbit pass produces actionable output

Precedent: Three-Body Constellation ran an 8B-param model in orbit, infrastructure census across 189 km² NW China (Nov 2025).8 Our system is 2,000× smaller.

§ 10 · VALUE

Three tiers of impact.

TIER 01 · DIRECT

Real-time grain monitoring

  • 82 Mt from NE China alone in 20252
  • Same-orbit detection of stress, drought, disease
  • Feeds 286 emergency centres + 716 response teams
TIER 02 · COORDINATED

Satellite-terrestrial intelligence loop

  • Onboard AI filters observations → BeiDou geo-references
  • 33M BeiDou terminals9 deliver advisories to tractors
TIER 03 · TRANSFER

Cross-mission generalization

  • Forestry · water resources · urban observation
  • Disaster response · Belt & Road agricultural cooperation
Grain fields at dawn — Sentinel-2 monitoring target
MONITORING SCALE
82 Mt of grain from NE China alone.
Detected per orbit pass.

Satellite-terrestrial coordination loop

Sentinel-2 + onboard AI

Filters which observations matter

BeiDou navigation

Precise geo-reference link

Farm equipment

33M BeiDou terminals · advisories

Ground analysts

Closes seasonal action loop

§ 11 · ROADMAP

Three horizons. BeiDou 2027. 15th FYP.

2026 – 27

Distillation & HIL

Deploy INT8; complete hardware-in-the-loop validation on a Jetson Orin reference platform.

2027 – 29

In-orbit demonstration

Partner with ADA Space, Three-Body Constellation, or Tiansuan for a first orbital demo.

2030 +

Operational service

Daily-revisit grain monitoring across NE China; integrated with 15th-FYP infrastructure.

Technical improvement paths

11.A

INT4 group-wise quantization

  • ~0.99 MB total (23% of FP32)
  • Requires Jetson Orin Tensor Cores
11.B

Foundation-model distillation

  • Distill from Prithvi-EO, RemoteCLIP10
  • Pre-train upstream · deploy downstream
11.C

Multi-modal fusion · S1 + S2

  • Add Sentinel-1 SAR for cloud-penetrating coverage
  • Critical for NE China summer cloud cover
11.D

Federated constellation learning

  • Gradient deltas shared via inter-satellite laser links
  • Validated on Three-Body Constellation8

From HAUMEA.

§ 13 · REFERENCES

References.

  1. ESA · "Φ-sat-1 on-board AI for Earth observation" — esa.int
  2. NBS China · "2025 Grain Output Bulletin" — stats.gov.cn
  3. State Council PRC · "14th / 15th Five-Year Agricultural Modernization Plans" — english.www.gov.cn
  4. NVIDIA · "Jetson Orin INT8 Tensor Core Technical Brief" — developer.nvidia.com
  5. Wei et al. · "RedNet: radiation-tolerant on-orbit inference" — arXiv:2407.11853
  6. ESA · "Φ-sat-2 onboard AI demonstrator" — esa.int
  7. Planet Labs · "Pelican-4 onboard inference demonstration, March 2026" — planet.com
  8. ADA Space · "Three-Body Constellation, NW China census, Nov 2025" — adaspace.com
  9. CSNO · "BeiDou Navigation Satellite System Annual Report" — en.beidou.gov.cn
  10. IBM & NASA · "Prithvi-EO geospatial foundation model" — huggingface.co
  11. ESA · "Sentinel-2 mission, Copernicus Open Hub" — sentinels.copernicus.eu