Seminar: Physical Interpretation and Uncertainty Quantification of Deep Learning for Convective Initiation Nowcasts

Speaker

Da Fan, The Pennsylvania State University

Bio

Da Fan is a PhD student from the meteorology department in The Pennsylvania State University

Abstract

Convective initiation (CI) nowcasts remains a challenging problem for both numerical weather prediction models and existing nowcast algorithms. In this study, object-based probabilistic deep learning models are developed to predict CI over the next 10 minutes based on infrared GOES-R satellite observations. The data comes from patches surrounding potential CI objects identified in Multi-Radar Multi-Sensor Doppler weather radar products by an objective radar-based approach from 1 June to 30 July 2020. The deep learning models significantly outperform existing nowcast algorithms in terms of accuracy and false alarm rate. To understand the features important for CI occurrence, deep learning models are interpreted through attribution methods including SHapley Additive exPlanations (SHAP) and gradient times input (GI). The interpretation through these methods reveals that signals of growing cumulus clouds and ice production are important attributions behind the initiation of convection, consistent with our understanding of CI. To further explore how trustworthy the results are, the uncertainty of the results is quantified and connected to the spatial distribution of model performance and CI frequency. The uncertainty is further separated into aleatoric and epistemic uncertainty to measure the relative importance of the uncertainty from data and model structure.

Add this event to your calendar (internal only).