Storms, seen before they strike.

Challenge

Convective storms are the forecasting problem numerical weather models are worst at. They form, intensify and dissipate in tens of minutes, below the resolution and update cadence of operational NWP. Yet the zero-to-two-hour window is exactly where the operational decisions live: grounding flights, closing yards, dispatching crews, protecting equipment. The classic answer, extrapolating radar echoes forward, degrades quickly because storms do not move linearly; they grow and decay.

The question this project set out to answer: how far can nowcasting skill be pushed using only public data, without proprietary sensor networks?

Solution

Antenor built RADARES, a machine learning nowcasting system for convective weather over Argentina that fuses five open data streams: the SINARAME national weather radar network, GOES-19 geostationary satellite imagery (ABI infrared and derived Level-2 products), Geostationary Lightning Mapper detections, surface METAR reports and open numerical-model point forecasts. Radar sees precipitation structure at high resolution; the satellite and lightning data see the pre-convective environment radar coverage misses. The system treats latency as a first-class scientific variable and validates every field directly from the rendered public imagery.

The work is published as an EarthArXiv preprint, From Public Weather Images to House-Scale Convective Nowcasting, including data preparation, georeferencing, validation and evaluation against standard baselines.

Stack

• SINARAME weather radar network (rendered PNG products, georeferenced to 0.3–1.1 km per radar)
• GOES-19 ABI infrared, Level-2 products and Geostationary Lightning Mapper detections
• Surface METAR reports and open numerical-model point forecasts
• Machine learning models scored against operational optical-flow baselines (S-PROG)

Results

• Critical Success Index of 0.39 at +60 minutes, against 0.32 for the S-PROG infrared optical-flow baseline
• 80,444 radar echoes validated against ground truth with 0% loss of real-rain echoes
• Forecast horizons tested from +10 to +120 minutes, on a radar-to-product latency of 10–15 minutes
• Methods, validation and results published openly as an EarthArXiv preprint
• Demonstrates the approach Antenor applies to client forecasting problems: public or in-house data, measurable skill, documented methods