DEMETER
A time machine for NDVI and physical features.
Etherion / Messis / DEMETER
Product direction / historical IoT → physical projection
Train the model on a field’s historical IoT measurements of physical features. With the World Indexes, the aim is short- to long-term projection of NDVI and physical features — their interconnectivity, evolution, variations and response to perturbations. A time machine for the field’s measured physical world, not another standalone NDVI score.
API business model on standby. Later: case-by-case negotiated contracts, planned against the historical IoT data your fields already hold, to train and deliver a fine-tuned model on that history.
From field history to projected futures
IoT-trained direction- 01
Existing physical history
The measurements already recorded by the field's IoT sensors.
- 02
A model trained on that history
Fine-tune the universal physical base around the customer's measured channels and cadence.
- 03
World Indexes through time
Project NDVI and physical features together: their relationships, evolution and response to changed conditions.
2
World Indexes
17
physical features in the base
69
held-out location/year histories
52
common rollout origins
00 / From satellite averages to physical history
From an NDVI calculator to coupled physical projection.
The current satellite-trained model is an average NDVI calculator. The next step is to train on historical in-situ IoT measurements, so the model learns from a field's measured physical history rather than satellite averages alone.
The objective is not just a future NDVI value. It is the joint projection of vegetation condition and physical features — soil moisture, temperature, rainfall, radiation and water balance — with the World Indexes exposing their modelled relationships and how they evolve under changed conditions.
01 / World Indexes — the product
Interconnectivity, evolution, variations and perturbations — projected through time.
World Indexes are the product. Paired with a model trained on historical IoT measurements, the aim is short- to long-term projection of NDVI and physical features: how they connect, how they evolve, and how variations and perturbations change the projected state.
Stability / Instability
How the projected state responds to changes, where its stability shifts, and how model uncertainty changes along a scenario. Read this alongside the physical trajectories, not as a standalone score.
Attention / Influence
Keep what the model attends to separate from what its projection depends on. Attention describes what is read; influence describes which inputs move the projected state. The distinction matters when inspecting interconnectivity.
Read the physical world together
02 / World operations
One physical world, seven ways to explore it.
These are the planned operations of the IoT-trained model: explore a field's physical state through time and read its changing relationships with the World Indexes.
Prediction
Project NDVI and physical features together from observed history and forcing.
Simulation
Compare trajectories under different rainfall, heat or management scenarios.
Observation
Inspect measured variations and where modelled relationships change.
Speculation
Explore alternative physical futures consistent with a changed scenario.
Windback
Explore past states or disturbances consistent with the observed history.
Intervention
Explore action sequences against a desired future physical state.
Multi-head inference
Read vegetation condition, water balance and uncertainty on one projected trajectory.
03 / Completed physical-base benchmark · 7 October 2026
The base is prepared. The measurements are in.
Completed evaluation on 10,713 next-day transitions across 69 held-out location/year histories and 24 locations. These are satellite/reanalysis results for the universal physical base, not customer IoT fine-tuning results.
Next-day performance is close to the source model.
NDVI and solar-radiation RMSE changed little; soil-moisture RMSE rose while its typical absolute error fell. The eight-feature aggregate did not show a clear improvement: the paired uncertainty interval spans zero. This is a physical starting point for training on field history, not a claim that retraining alone delivered the IoT product.
60 / 69
healed histories below their persistence error
0.468211
healed median eight-feature score
0.506807
same-window persistence median
16.17%
pooled squared-error skill vs persistence
A new score contract
04 / Next-day physical accuracy
Native units. Matched observations.
RMSE is shown for the source model, the prepared physical base and the same-window persistence forecast. Lower is better. Correlation describes the prepared base's prediction against observed state.
| Feature / unit | Source RMSE | Physical base RMSE | Persistence RMSE | Base correlation |
|---|---|---|---|---|
| NDVI / index | 0.021782 | 0.021763 | 0.021758 | 0.973 |
| Soil moisture / m³/m³ | 0.140850 | 0.144091 | 0.145582 | 0.562 |
| Temperature / °C | 1.772 | 1.789 | 1.919 | 0.977 |
| Rainfall / mm/day | 5.606 | 5.628 | 7.124 | 0.376 |
| ET0 / mm/day | 0.846960 | 0.846449 | 0.947874 | 0.914 |
| Humidity / % | 7.990 | 8.046 | 8.551 | 0.931 |
| Wind speed / m/s | 0.925414 | 0.927499 | 1.000 | 0.711 |
| Solar radiation / MJ/m²/day | 3.611 | 3.610 | 4.339 | 0.828 |
Typical error and tail error are different
05 / Evolution through time
Follow the trajectory, not just the next point.
The completed test follows 52 common origins for 1, 7, 30 and 90 days, feeding predictions back without inserting later observations. It uses supplied future ERA5 weather: a reanalysis-conditioned scenario, not weather known at forecast time.
NDVI
Endpoint RMSE / index
| Days | Source | Base |
|---|---|---|
| 1 | 0.020901 | 0.020559 |
| 7 | 0.048172 | 0.050858 |
| 30 | 0.132404 | 0.156315 |
| 90 | 1.025 | 1.425 |
Soil moisture
Endpoint RMSE / m³/m³
| Days | Source | Base |
|---|---|---|
| 1 | 0.133189 | 0.140394 |
| 7 | 0.268961 | 0.159847 |
| 30 | 3.286 | 1.291 |
| 90 | 19.913 | 12.020 |
Solar radiation
Endpoint RMSE / MJ/m²/day
| Days | Source | Base |
|---|---|---|
| 1 | 4.245 | 4.096 |
| 7 | 10.675 | 10.116 |
| 30 | 34.153 | 27.024 |
| 90 | 279.004 | 85.077 |
Raw trajectories remain visible
06 / Measured interconnectivity response
Change one condition. Read what moves with it.
The completed suite includes changes to rainfall, temperature, soil moisture and radiation. It reports how NDVI and other physical features respond across time, separately from forecast accuracy.
A one-day rainfall change
A +6.754173 mm/day rainfall change in the observed history produces these median absolute NDVI responses in the prepared base, with future weather unchanged.
0.001111
Day 1 / NDVI index
0.003648
Day 7 / NDVI index
0.046039
Day 90 / NDVI index
Combined conditions do not simply add
Combined rainfall and temperature changes produce a non-additive NDVI response: median absolute interaction is 0.001863 at day 1, 0.005705 at day 7 and 0.090113 at day 90. This is a measured property of the modelled scenario, not a causal field-trial result.
The index readout was evaluated at eight anchors. It keeps attention distinct from influence and scenario spread distinct from realised error.
What these responses mean
07 / Evaluation contract
One physical base. Explicit evidence.
The completed benchmark is kept separate from the historical release and from future customer IoT delivery.
- 01
Next-day evaluation: 10,713 transitions across 69 location/year histories at 24 locations; fitting 2017–2023, evaluation 2024–2026.
- 02
Both models are scored on the same 17 physical targets, with separately fitted readouts and the same physical-unit observations. The summary uses eight physical features only.
- 03
Temporal evaluation: 52 common origins across all four horizons; predicted state is fed back, future weather is supplied, and raw range violations remain visible.
- 04
The benchmark compares the complete base-preparation change. It is not an isolated test of removing static features and not evidence of customer IoT fine-tuning.
Completed 7 October 2026 · native errors · matched scope
08 / Universal physical base — completed
The starting point for each field's measured history.
The universal physical base has completed retraining and evaluation on corrected native-unit data, with all static features stripped. Customer IoT fine-tuning is the next engagement-specific step.
Prepared physical inputs
The retrained base uses corrected soil moisture, solar radiation and the other physical inputs. It is a physical-feature starting point, not a per-crop model.
No static crop constants
All static features are stripped from the base. Its measured performance is reported above, rather than inferred from a training loss.
Cumulative features are optional
Cumulative features, such as a 30-day NDVI mean, may be retained but are optional during IoT fine-tuning. The customer's actual sensor history determines the included features.
09 / Public provenance
Benchmark observations. Customer history next.
The base benchmark uses satellite observations and reanalysis. They are not presented as in-situ measurements of a customer's field.
NASA SMAP
Satellite soil-moisture observations, measured in native volumetric units.
Sentinel-2
Vegetation and canopy observations used in the benchmark history.
ERA5
Copernicus reanalysis weather; future forcing is explicitly supplied in the rollout tests.
10 / The engagement model
Your existing history. A model fine-tuned on it.
The API business model is on standby. Later, Messis will operate through case-by-case negotiated contracts: plan against a customer's existing historical IoT data, train on that history, and deliver a fine-tuned model.
Plan against existing data
Start with the physical channels already measured on the customer's fields, their cadence and historical coverage. Define the engagement around that data.
Train and deliver the model
Fine-tune the universal physical base on that history, then deliver the fine-tuned model under the negotiated contract.
Project the measured physical world
Use the World Indexes to inspect NDVI and physical-feature interconnectivity, evolution, variations and perturbations through time.
A negotiated engagement, not a packaged API launch