Performance and regime of ensemble perturbations in sampling initial and forecast errors at multiple spatial scales in the NCEP global ensemble forecast system

One of the key to ensemble forecasting lies in the effective generation of initial ensemble perturbations, which aim to represent the probability distribution of analysis errors. Most previous studies focus on assessing the predictive skill of ensemble forecasts, offering only indirect assessments of the quality of initial ensemble conditions. This study evaluates the performance of ensemble perturbations in sampling errors at both the initial and forecast stages in the operational ensemble forecast system at NCEP. In particular, the ensemble sampling behavior across different spatial scales is assessed quantitatively, with both the magnitude and the direction considered. The primary results indicate that, at the initial time, smaller scales exhibit a higher spatial correlation between ensemble spread and error magnitude compared to larger scales. However, ensemble perturbations at larger scales demonstrate superior performance in capturing the spatially coherent structures of analysis errors, exhibiting a higher perturbation-error correlation than those at smaller scales. These differences in sampling performance across scales persist throughout the entire forecast lead times. The contrasting trend in sampling behavior of magnitude and direction across various scales suggest that the keys for effective ensemble generation vary depending on spatial scales.