Visual Analysis of the Temporal Evolution of Ensemble Forecast Sensitivities

Ensemble sensitivity analysis (ESA) has been established in the atmospheric sciences as a correlation-based approach to determine the sensitivity of a scalar forecast quantity computed by a numerical weather prediction model to changes in another model variable at a different model state. Its applications include determining the origin of forecast errors and placing targeted observations to improve future forecasts. We—a team of visualization scientists and meteorologists—present a visual analysis framework to improve upon current practice of ESA. We support the user in selecting regions to compute a meaningful target forecast quantity by embedding correlation-based grid-point clustering to obtain statistically coherent regions. The evolution of sensitivity features computed via ESA are then traced through time, by integrating a quantitative measure of feature matching into optical-flow-based feature assignment, and displayed by means of a swipe-path showing the geo-spatial evolution of the sensitivities. Visualization of the internal correlation structure of computed features guides the user towards those features robustly predicting a certain weather event. We demonstrate the use of our method by application to real-world 2D and 3D cases that occurred during the 2016 NAWDEX field campaign, showing the interactive generation of hypothesis chains to explore how atmospheric processes sensitive to each other are interrelated.

1 Introduction

Ensemble sensitivity analysis (ESA) has been established in meteorology as an ensemble-based approach to estimate the sensitivity of a scalar forecast quantity \$J\$ (in meteorology referred to as “forecast metric”) computed from numerical weather prediction (NWP) model output to changes in the same or another model variable at an earlier state. ESA has been introduced by Ancell and Hakim [1] and Torn and Hakim [42]. It is a correlation-based approach which considers the variations of physical quantities at different domain locations relative to each other.