The past informs. The future stays out.
We investigate backward-only lags and site-local calendars to prevent temporal leakage in evaluation.
LOCAL CALENDAR / BACKWARD-ONLYScientific honesty is part of the product. HydroFlux reports where the evidence is strong, where local verification is required and where it will not produce an estimate.
SCIENTIFIC ARCHITECTURE
The difference is how we connect observations, time and uncertainty.
We investigate backward-only lags and site-local calendars to prevent temporal leakage in evaluation.
LOCAL CALENDAR / BACKWARD-ONLYWe combine Sentinel-2 optical signals, Sentinel-1 radar and Landsat surface temperature with meteorological context.
OPTICAL + RADAR + THERMALWe evaluate conformal intervals and observed coverage by domain. A nominal 90% level does not guarantee coverage in every crop or region.
SPLIT-CONFORMAL / EVALUATIONA support-aware AI system for agricultural evapotranspiration, combining satellite observations, weather, terrain and field context with training evidence from eddy-covariance measurements.
Support is evaluated for every pixel and request. A crop name or country alone never guarantees support.
North American annual crops and forage when climate, land cover, sensor freshness and required context pass the support checks.
Typical examples: corn, soybean, wheat, forage and comparable annual cropping systems with usable observations.
Agricultural pixels outside the strongest evidence envelope, including specialty or perennial systems, may be returned with limited support and wider uncertainty.
Examples: vineyards, orchards, new climates, weak land-cover confidence or missing phenology and irrigation context. Verify locally.
HydroFlux returns null when critical evidence is stale, missing or contradictory, or when the pixel is outside the supported agricultural domain.
Examples: no usable optical and radar evidence, a clear observation older than 32 days, unresolved domain conflicts or non-agricultural surfaces.
SUPPORTED describes resemblance to the established evidence and input contract. It is not a claim of field-specific calibration or ground-truth certainty.
The public contract is intentionally simple: inspect the evidence first, estimate only where support is defensible, and expose the reasons.
The system evaluates sensor availability, crop and land cover, climate, terrain, irrigation, phenology and recent field history.
Support-aware evapotranspiration engine
The system prefers a transparent null over unsupported extrapolation.
Estimate inside the strongest evidence envelope.
Estimate supplied with a clear verification warning.
No value supplied when evidence is insufficient.
Every pixel includes uncertainty, observation age, support status and human-readable reasons.
This diagram documents the product contract without exposing private implementation details. Support can vary within the same field.
A daily map combines the latest valid multi-sensor HydroFlux anchor with daily weather and reference ET. It always declares how final the result is.
Multi-sensor crop state over the 16-day observation window
Radiation, temperature, humidity, wind and rainfall
Uncertainty · observation age · support status · temporal status
Four fields decide how much weight a number deserves.
Whether the pixel sits inside the validated domain. Limited support is a signal to verify locally, not a defect to hide.
An operational range that widens as evidence weakens. Treat it as a heuristic band, not as validated probabilistic coverage.
Days since the last clear Sentinel-2 or Landsat observation for that pixel. Older evidence caps support and widens uncertainty.
A pixel with no usable estimate returns null. An abstention is information, not a gap to fill with a candidate value.