Analysis Verification Experiments with a Statistical interpolation System

Abstract

An analysis performance verification (APV) system was developed using the 1989 version of the National Meteorological Center global data assimilation system (GDAS). It allows comparison of the GDAS analysis to observations that were withheld from the interpolation. APV was compared with another methodology, the HL method, based on some results of Hollingsworth and Lonnberg, where the analysis verification is done to non-withheld observations. GDAS analysis sensitivity to the ratio of the observation and prediction errors ? and to horizontal correlation length scale L variations was studied. Theoretical analysis experiments show greater analysis sensitivity to ? variation than to L. APV results for the GDAS 1989 version show that GDAS overestimates prediction errors in data-sparse regions and underestimates them in data-dense regions. These results indicate that the prediction error growth rate assumed in GDAS should have a regional variation. Optimal combinations of ? and L values were obtained for different regions. Over data very dense areas the results have shown that the GDAS height horizontal correlation function is inefficient. Thus, other correlation functions should be tested. The comparison between the results obtained with both verification methods shows that APV results are more sensitive to the proposed parameter variation than the other methodology results.