Multivariate Ensemble Sensitivity Analysis for an Extreme Weather Event Over Indian Subcontinent

Abstract

<p>Ensemble forecasts have proven useful for diagnosing the source of forecast uncertainty in a wide variety of atmospheric systems. Ensemble Sensitivity Analysis (ESA) uses ensemble forecasts to evaluate the impact of changes in initial conditions on subsequent forecasts. Ensemble sensitivity is often used as a simple univariate regression as it approximates the analysis covariance matrix with diagonal elements on predicting the response to an initial perturbation. On the contrary, the multivariate ensemble sensitivity computes the ensemble sensitivities by incorporating the contribution from the full covariance matrix. In this work, the precipitation forecast responses of an extreme rainfall event from both univariate and multivariate ensemble sensitivity methods are analyzed. The ensemble forecasts and analyses are generated using an ensemble Kalman Filter (EnKF) coupled with the Advanced Research version of the Weather Research and Forecasting (WRF) model. Based on the results obtained the univariate ensemble sensitivity analysis shows broadly distributed sensitivity patterns, while the multivariate sensitivity analysis generally exhibits organized sensitivity patterns. The perturbation initial condition experiment applied to both methods proves that the ensemble sensitivity estimated by multivariate is more realistic compared to univariate ensemble sensitivity. In the presence of added model error, using the Stochastic Kinetic Energy Backscatter Scheme (SKEBS) it is found that the forecast response estimated by multivariate sensitivity compares better with the actual model response. Further, it is identified that if the localization used in multivariate is not sufficient its performance is contaminated by the occurrence of spurious correlations. The impact of lead times on both methods shows that multivariate provides better performance than univariate mainly at longer lead times when nonlinearity becomes important. The use of convection-permitting ensemble forecasts reveals that the multivariate ensemble sensitivity with localization ameliorates the sensitivity estimates in convective scales.</p>