From Public Weather Images to House-Scale Convective Nowcasting: A Latency-Aware, Multi-Source Fusion and Alerting System Built Entirely from Open Data

Jose Antonio Velazquez Bustamante

Short-range convective warning systems are usually evaluated as gridded meteorological products, but many real decisions are point decisions: whether a specific place will be affected in the next minutes to hours, whether the event is electrically active, and when conditions will improve. This paper presents RADARES, a low-cost operational nowcasting and alerting system built entirely from public data: rendered Argentine weather-radar imagery (SINARAME PNG products), GOES-19 ABI infrared and derived Level-2 products, Geostationary Lightning Mapper (GLM) detections, surface METAR reports, and open numerical-model point forecasts. The system runs 24/7 on consumer hardware, targets a fixed point in northwest Greater Buenos Aires, and pushes phone alerts that answer four questions: when does precipitation arrive, how intense, how long, and when does it improve. Three design principles distinguish the system. First, latency and source hierarchy are treated as scientific variables: the nearest radar (10-15 min latency) triggers urgent alerts while slower context is cached and never blocks the urgent path. Second, every field recovered from rendered imagery is validated before use: colorbar inversion carried a silent +5 dBZ bias; ground-control-point georeferencing reached 0.3-1.1 km per-radar accuracy. Third, multi-source physics replaces single-source heuristics: radio-frequency interference indistinguishable from rain bands by geometry alone is removed by physical cross-source vetoes, validated on 80,444 echoes with 0.0% loss of real-rain echoes. Retrospective verification shows infrared optical-flow advection beats persistence at all leads from +10 to +120 min, with S-PROG selected by bake-off (CSI 0.39 vs 0.32 at +60 min); radar nowcasting requires regime-adaptive method selection, with a recent-skill gate that never collapses across four contrasting events. Pixel-level convective-initiation detection reproduces the literature's pattern of high detection probability with unusable false alarms, supporting an object-level future path. Documented operational episodes include a same-day double outage with honest degradation onto satellite rainfall, and an advection-versus-propagation disagreement in which radar and satellite trackers reached opposite, both correct, conclusions -- the concrete argument for hierarchical fusion without inter-source motion vetoes.