Search

This study aims to assess the contribution to prediction performance of dynamically downscaling seasonal precipitation for the Brazilian Northeast region (NEB) produced in February with a global climate model for the austral autumn (March-April-May). This season includes the peak precipitation period of a large portion of this region. To achieve this aim, 32 years of hindcasts for the period 1979–2010 were produced with a regional model known as the Brazilian developments on the Regional Atmospheric Modeling System (BRAMS) run at 30 km spatial resolution, which was unidirectional nested to a global model run at approximately 200 km spatial resolution. The used global model was the Atmospheric General Circulation Models of the Brazilian Center for Weather Forecast and Climate Studies (CPTEC/AGCM), which provided initial and lateral boundary conditions for BRAMS. Hindcast ensembles of 10 members were produced with both the global and the regional models, simulating an operational prediction environment using persisted sea surface temperature (SST) anomalies observed in the previous January of each of the 32 hindcast years as oceanic boundary conditions. A comparative deterministic and probabilistic performance assessment of the two models was conducted. The deterministic assessment included the investigation of the mean error (bias) of the produced seasonal precipitation hindcasts, the degree of linear association between the observed and predicted precipitation anomalies measuring phase error through the correlation of the ensemble means with the corresponding observations, and amplitude error investigated through the ratio between the hindcast ensemble mean and observed variances. The probabilistic assessment was performed investigation the reliability, resolution, and discrimination characteristics of tercile probability predictions for the above normal and below normal categories, and also for probabilistic predictions for the occurrence of the symmetric event positive (or negative) precipitation anomaly. The two investigated modeling approaches showed similar performance, with an indication for dynamically downscaled predictions to have reduced amplitude error and slightly improved reliability and resolution compared to the predictions produced with the global model.