“…The statistics of innovations, the differences between model forecasts and data, are an important component for data assimilation and evaluation of performance of numerical weather prediction (NWP) models (Rutherford, 1972;Hollingsworth and Lönnberg, 1986;Lönnberg and Hollingsworth, 1986;Thiebaux et al, 1986Thiebaux et al, , 1990Shaw et al, 1987;Daley 1991Daley , 1992Daley , 1993Bartello and Mitchell, 1992;Mitchell et al, 1993;Devenyi and Schlatter, 1994;Cohn, 1997;Xu and Wei, 2001a,b;Xu et al, 2007). Rawinsonde data have been the main source of innovation statistics and the rawinsonde observation errors are assumed to be spatially uncorrelated which produces simultaneous estimates of the total observationand model forecast-error covariance Lönnberg and Hollingsworth, 1986).…”