Using Deep Neuron Network to improve the Performance of NUCAPS Profiles in Lower Atmosphere

High temporal resolution brightness temperature (BT) measurements from Advanced Baseline Imager (ABI) over CONUS are used together with short-range forecasts from NWP (numerical weather prediction) to develop a local severe storm (LSS) prediction model. This model is capable of identifying potential occurrence of LSS systems in pre-convection environment and predicting their respective locations and intensity classification (weak, medium and severe) at an interval of 5 minutes. The Random Forests (RF) technique is used to train the statistical prediction model. The LSS cases over CONUS from March to October 2018 are identified and tracked for the training. The ABI BT measurements from 10 infrared (IR) bands, together with spatially-temporally collocated atmospheric environment parameters from NOAA GFS (global forecast system, and later FV3) short-range forecast data are collocated and used as predictors. CMORPH2 precipitation analysis data is used as the benchmark of storm intensity to build a predictive classification model. Sample-balance technique is utilized during the training to enhance the accuracy of the model by adjusting the sample size of different classes to minimize the bias from the over-fitting of any class. Validation studies are being carried out with independent storm dataset from April to September 2019, and the relative importance and impact of the different predictors, including ABI’s BT observations from different channels, their temporal variations, NWP parameters, and their temporal variations, are investigated. Results and findings will be analyzed and presented. The objective is to develop a statistical prediction model for LSS occurrence in pre-convection before the radar observations can identify the system.