Machine Learning-based Probabilistic Precipitation Estimation with the GOES-16 Advanced Baseline Imager

This study introduces a new machine learning-based probabilistic quantitative precipitation estimate (PQPE) retrieval that uses observations from the GOES-16 Advanced Baseline Imager (ABI) across the CONUS at 5 min temporal resolution and ~2 km spatial resolution. It is developed and evaluated using the Ground Validation Multi-Radar/Multi-Sensor (GV-MRMS) system as a benchmark, and features Convolutional Neural Network (CNN) machine learning. Key advances include (1) the design of a three-dimensional CNN model to retrieve the distribution of precipitation rate instead of a single deterministic value; (2) a comprehensive set of predictors based on spatio-temporal information from infrared ABI channels and complemented by environmental conditions from Numerical Weather Prediction (NWP) models; and (3) using probabilities of occurrence for different precipitation type (e.g., convective and stratiform), retrieved from a separate machine learning  model as predictors in the CNN model. Precipitation type predictors allow a single model to be used seamlessly for all precipitation types. The analysis reveals that combining predictors based on satellite and NWP data leads to improved performance, with the greatest improvement in the stratiform precipitation type. The use of probabilities of precipitation type as predictors contributes significantly to the improved performance of the quantitative precipitation retrievals. Furthermore, improvements in conditional biases are demonstrated for all precipitation rates when compared to a deterministic CNN model.