Automatic Detection and Classification of Orographic Precipitation using Machine Learning

Ground-clutter is a major cause of large detection and underestimation errors in satellite-based (e.g. Global Precipitation Measurement Dual Polarization Radar, GPM DPR) precipitation radar retrievals in complex terrain. Here, an Artificial Intelligence (AI) framework consisting of sequential precipitation detection and vertical structure prediction algorithms is proposed to mitigate these errors using machine learning techniques to uncover predictive associations among satellite- and ground-based measurements aided by Numerical Weather Prediction model analysis, specifically the High-Resolution Rapid Refresh (HRRR) model. The framework is implemented and tested for quantitative estimation of orographic precipitation in the Southern Appalachian Mountains (SAM). Precipitation detection relies on a Random Forest Classifier to identify rainfall based on GPM Microwave Imager (GMI) calibrated brightness temperatures (Tbs) and HRRR mixing ratios in the lower troposphere (~ 1.5 km above ground level). The vertical structure of precipitation prediction algorithm is a Convolution Neural Network trained to learn associations among GPM DPR Ku-band reflectivity profiles, GMI Tbs, and orographic precipitation regimes in the SAM including low level light rainfall, shallow rainfall with low-level enhancement, stratiform rainfall with bright band, and deep heavy rainfall with low- and mid-level enhancement. Vertical structure classes corresponding to the distinct orographic precipitation regimes were isolated through k-means clustering of ground-based Multi-Radar/Multi-Sensor radar reflectivity profiles. The AI framework is demonstrated for automatic retrieval of warm season precipitation in the SAM over a 3-year period (2016-2019) achieving large reductions in false alarms (77%) and missed detections (82%) relative to GPM Ku-PR precipitation products, and significant rain-rate corrections (up to one order of magnitude) by using a physically-based model to capture the microphysics of low-level enhancement (i.e. seeder-feeder interactions).