Background error covariance (B) matrix is critical for variational data assimilation as it greatly affects the analyses of three-dimensional variational assimilation. The National Meteorological Center method was used to estimate the B matrix using the forecasts from the Advanced Research Weather Research and Forecasting regional model. To further understand and evaluate the newly generated regional B matrix, its characteristics were compared with the global B estimated from the Global Forecast System model. Sensitivity experiments were undertaken by changing the horizontal length-scales and standard deviations of the B matrix, and its impacts on the typhoon forecast were also examined. Verification against radiosonde observations showed that the varying horizontal length-scale has a significant positive impact on the 24-h forecast of temperature, specific humidity, u-wind, and v-wind. On the other hand, changing standard deviations of the B matrix has a slight influence only on the specific humidity and wind (u-component) forecast. Compared with the global B, the tuned regional B showed improvements in temperature forecasts. In addition, using the tuned regional B also led to a positive impact on the typhoon (Saola, Damrey, and Haikui) track forecasts as compared with the untuned B and global B.