The Limitations of Using Satellite and Radar Observations in Artificial Intelligence Models for Convective Initiation Nowcasting

Artificial intelligence models for nowcasting convective initiation often rely on satellite and radar observations to determine the location of convection that is likely to initiate a thunderstorm in the near-term. Such models often exhibit high false alarm ratios, impacting the models’ reliability. To investigate the extent of false positive (< 30 dBZ at -10°C) and true positive (≥ 30 dBZ at -10°C) predictions in a thunderstorm nowcasting model called ThunderCast, an object-based method is used to identify and track predicted storm objects from ThunderCast’s output. The resulting predicted storm tracks are identified as true or false positive and further classified by the presence (or lack thereof) of lightning from the Geostationary Lightning Mapper (GLM) and Earth Networks, Inc.’s (ENI’s) total lightning product. A comparison of the predicted storm tracks to their corresponding Multi-Radar Multi-Radar Multi-Sensor (MRMS) radar reflectivities at -10°C reveals 67.8% of tracks were true positive and over half (59.3%) of these did not produce lightning. Advanced Baseline Imager (ABI) data from the Geostationary Operational Environmental Satellites (GOES) indicated non-electrified true and false positive tracks contained similar visual signatures during initiation such as cloud-top glaciation. This demonstrates defining convective initiation with a radar reflectivity threshold is imperfect and additional predictors may be needed to supplement satellite data in convective initiation models.