A Generalized Deep Neural Network for Estimating Severe Hail Likelihood from Satellite Infrared Cloud Top Patterns and Microwave Radiances

Deep convection signatures in satellite infrared (IR) and passive-microwave imagers can provide a method of severe weather detection around the globe where ground-based observations are otherwise inconsistent or unavailable. Distinct patterns in these remote-sensing datasets, such as cold overshooting cloud tops and prominent local brightness temperature depressions, can be leveraged to analyze the occurrence of costly storms hazards like hail. When analyzed alone these satellite measurements suffer from key limitations, but otherwise complement one another when collocated to combine high-resolution identification of storm cores with cloud penetrating retrievals. As such, this work highlights the use of a deep neural network (DNN) for unraveling the interrelationships of MODIS IR and AMSR-E passive-microwave signatures to thereby characterize potentially severe hail defined by ground-based weather radar. On an unseen validation set, the generalized DNN identifies better than 77% of potentially severe maximum expected size of hail (MESH) signatures with false alarm occurrence under 30%. Because the model is based on satellite imagery alone, it is well suited for global application compared to methods that rely on environmental reanalysis, although the contributions of environmental data are also, nevertheless, quantified. This generalized approach serves to benefit those interested in better understanding hail frequency and severity in regions with insufficient storm reporting.