Frost Risk Intelligence — Argamasilla de Alba

Case Study: A frost forecast system based on on-site sensor data + weather models.

We join local sensor data with global weather forecast models to give orchards accurate, verified frost forecasts up to 5 nights ahead: across two frost seasons the forecast error stays between 1.4 °C (next night) and 1.8 °C (five nights out), frost/no-frost discrimination holds at AUC 0.96–0.99, and 12 of 14 frost-night forecasts in the archive raised the alert in time.

The service includes:

• On-site sensor installation — temperature and humidity in your orchard, the local truth every forecast is calibrated against.
• Calibrated forecast system — three global ensemble models, bias- and spread-corrected for your site and blended into one verified probability.
• Real-time monitoring — this dashboard, refreshed every forecast cycle.
• Alert system — WhatsApp/SMS the moment frost probability crosses your cost/loss threshold, hours before the minimum.

Case study site: Argamasilla de Alba (SIAR CR07) — grapevine , critical temperature -1.5 °C.

The model reproduces the observed values within its confidence band

A perfect point forecast is impossible — so we forecast the honest range instead. The blue band below is the model's own 90% confidence interval, verified against two seasons of data: nine times out of ten the observed temperature falls inside it. Frost alerts are computed from exactly these probabilities, not from a single number that can silently miss.

Reliable up to 5 nights before a frost comes

Forecast error grows from 1.4 °C the night before to 1.8 °C five nights out (CRPS 0.78 → 1.00 °C), and frost discrimination stays at AUC 0.96–0.99 across all five leads — enough time to plan protection, not just react.

Beyond five nights only a single frost night is scoreable in the archive (the horizon limit) — the practical service window is 1–5 nights ahead.

Every frost night, caught

Raw weather forecast models fail to predict local conditions. Our calibrated system instead issues alerts on all analysed frost nights!:

• Performance — every reliability metric, measured on ~200 nights.
• The model — how the calibration and model blending work.
• The service — what we install, operate and answer for, orchard by orchard.