Ten technologies changing weather observation and forecast speed

Four imagers, two sounders and a 2.5-minute regional scan define EUMETSAT’s Meteosat Third Generation constellation, which can produce up to 50 times as much data as traditional meteorological satellites. These systems extend a forecast chain already built from weather balloons, satellites, radar and numerical prediction models.

By Sophie Lindqvist

What do new satellite and balloon observations contribute?

ESA and JAXA’s EarthCARE satellite is entering ECMWF’s global forecasting system, the first forecasting center to use this form of data. Its Cloud Profiling Radar (CPR) returns radar reflectivity describing vertical cloud structure and precipitation intensity. ECMWF combines those profiles with millions of other observations, giving the forecast a vertical view of cloud layers rather than a horizontal snapshot.

EUMETSAT’s geostationary Meteosat Third Generation (MTG) constellation combines four imagers with two sounders for severe-weather forecasting. It can generate up to 50 times as much data as traditional satellites and scan Earth’s full face every 10 minutes. But it can target a smaller region such as Europe in 2.5 minutes.

Lightning imagers provide day-and-night surveillance of cloud-to-cloud and cloud-to-ground flashes, while infrared sounders map atmospheric temperature and humidity. Ultraviolet, visible and near-infrared spectrometers on the sounding satellites monitor trace gases and air quality.

WindBorne’s WeatherMesh extends global observation with sounding balloons that can remain airborne for days or weeks. The balloons add multi-day persistence to pressure, temperature, humidity and wind measurements over oceans and remote regions, unlike traditional radiosondes (instrumented weather balloons). WindBorne says its deep-learning model assimilates the data and generates forecasts within seconds rather than hours; no test conditions accompany that claim.

Across these projects, richer or more persistent observations enter the forecast chain earlier. The practical gain is input quality and coverage, although most timing and performance statements remain agency or vendor claims rather than shared benchmarks.

How are AI and high-performance computing changing forecasts?

NOAA’s experimental, AI-powered HRRRCast Version 3 (High Resolution Rapid Refresh) trains on three years of HRRR model data. Its March 2026 release raises regional resolution to 3 km and expands the training set from three years to nine.

New diagnostics include wind shear and helicity, plus neural-network predictions for low-, mid- and high-cloud cover. NOAA intends the system for National Weather Service (NWS) testing; HRRR has served as NWS’s short-term forecast model for the past 10 years.

NOAA’s new global AI suite contains three models intended to improve speed, accuracy and efficiency. AIGFS (Artificial Intelligence Global Forecast System) aims to deliver forecasts faster and more efficiently; AIGEFS (Artificial Intelligence Global Ensemble Forecast System) generates a range of probable outcomes. HGEFS (Hybrid-GEFS) combines AIGEFS with NOAA’s Global Ensemble Forecast System as a hybrid grand ensemble.

NOAA says HGEFS “consistently outperforms both the AI-only and physics-only ensemble systems.” No scores, forecast periods or evaluation setup accompany the statement, so teams should treat it as an agency claim until they reproduce the comparison.

Tomorrow.io’s ICGen applies machine learning to data from a growing constellation of microwave sounders. These satellites collect temperature and humidity profiles even in cloudy conditions, and the machine-learning stage produces improved atmospheric initial conditions for existing numerical weather prediction (NWP) models. The design replaces massive processing with learned processing; no error reduction or compute-time benchmark is stated.

The European Commission’s Destination Earth (DestinE) has entered its third phase and builds a digital twin of Earth’s system for climate-change adaptation and weather-induced extremes. It combines ESA and EUMETSAT satellite observations, ECMWF forecasting models and European supercomputers rather than producing conventional forecasts.

Meteorologists can run what-if scenarios involving floods, hurricanes, heatwaves and storms. The platform operates at highly scaled resolution, but no numerical resolution or performance result accompanies that description.

Swiss company Meteomatics generates EURO1K forecasts across Europe at 1-km resolution. Incorporating NASA’s digital terrain model and low-altitude atmospheric measurements from Meteomatics’ Meteodrones scales specified landscapes to 90 m. The model targets sudden thunderstorms, hail, fog and local wind channeling, with the same micro-weather capability across national borders.

Across these projects, AI shortens computation while higher-resolution observations and terrain data reshape the forecast itself. The practical choice depends on whether agencies need a regional grid, a global ensemble, a digital twin or a local forecast.

How much faster can radar identify severe weather?

NOAA’s new mobile radar fleet places research-grade instruments at the front lines of tornadoes, wildfires, hailstorms, flash flooding and severe winds. Every truck carries an X-band radar with a 3-cm wavelength, and one truck also carries a C-band unit at 5 cm. NOAA expects the deployment to provide real-time insight into threats as they develop.

The wavelength choice sets a practical trade-off. Because the wavelength is shorter, X-band detects smaller drops and particles more efficiently, at the expense of imaging large drops or dense concentrations that can cause data-quality issues.

Robin Radar Systems’ PHARA (PHased Array Radar for Atmospheric research) is a 3D Ku-band phased-array platform with polarization diversity. It tracks the movement and growth of cloud particles as well as large-scale weather fronts. Robin says traditional radar can take “up to five minutes to scan the sky,” while a transmit/receive panel array completes the same job “in under a minute.”

Robin has developed the first prototype transmit tile in a 14-tile array that will ultimately form PHARA’s complete transmit phased-array antenna. The stated timing comparison supplies no sky volume, range or weather-condition details, which limits direct comparison with conventional radar.

Together, the mobile fleet and PHARA target earlier recognition of rapidly evolving hazards through placement and scan speed. Radar remains an observation layer feeding decisions; the projects present gains in detection, not verified reductions in warning errors.

What should agencies validate before adoption?

The roundup links faster models, better observations and more detailed atmospheric data to earlier warnings for severe thunderstorms, hurricanes and flash flooding. It also acknowledges that forecasting will retain some uncertainty. Most projects supply specifications or institutional claims rather than common forecast-error, latency and detection metrics, so none supports a universal performance ranking.

Adoption therefore depends on matching each system to the environment and warning task: global coverage, regional rapid refresh, local terrain effects, vertical cloud structure or hazard-adjacent radar coverage. Which combinations should agencies validate against their existing forecasts, and which can enter operational service first?