Comparison of rainfall structures between NCEP/NCAR reanalyses and observed data over tropical Africa

A comparison is made between modelled (NCEP/NCAR reanalysis) and observed (CRC and CRU dataset) annual and monthly precipitation over tropical Africa during the period 1958-1997. The split moving-windows dissimilarity analysis (SMWDA) is used to locate abrupt changes in rainfall time series. In the NCEP reanalysis data, we identify a main abrupt shift, which occurs in 1967 and concerns more than 50% of grid points. In the observation, this shift is only found over parts of tropical North Africa. Three other NCEP abrupt shifts (1975, 1983 and 1987) in the reanalysis shown by the SMWDA, each concerning about 20% of tropical Africa, are not identified in the observation. One hypothesis concerning the 1967 marked abrupt shift is a problem of data assimilation in the NCEP/NCAR model which generated artificial shifts in the time series. In view of this result, further comparisons have been restricted to the period 1968-1997. On a continental scale, while the CRC and NCEP mean seasonal rainfall patterns are almost the same, however, some regional features are not well reproduced. Using five regional rainfall indexes, the mean seasonal rainfall cycle is correctly reproduced, but the NCEP reanalysis generally underestimates the amounts during the rainy season. The NCEP reanalysis rainfall is closer to the observation when the region shows a single rainy season. The correlation values between NCEP and CRC interannual rainfall variations over the period 1968-1997 are very low and seldom significant. The NCEP four main structures of rainfall variability as deducted from rotated principal component analysis are not realistic at all and the associated time series are systematically dominated by a marked low-frequency variability not present in the observation. However, the main teleconnections between ENSO and African rainfall variations are fairly well reproduced, with a correct location of the main structures, but with lower correlation values than those found in the observation.