Four methods of downscaling daily rainfall sequences from general circulation model (GCM) simulations are intercompared over Senegal, using a 13-station network of daily observations during July-September 1961-98. The local scaling method calibrates raw GCM daily rainfall at the closest grid point to a given station so that the climatological distribution of rainfall matches the observed one. The k-nearest neighbor and weather classification schemes resample historical station rainfall observations according to the similarity between the daily wind fields from an ensemble of GCM simulations and a historical library of reanalysis [from 40-yr ECMWF Re-Analysis (ERA-40)] daily wind fields. The nonhomogenous hidden Markov model uses a small set of hidden states to describe the relationship between daily station rainfall observations and low-pass-filtered simulated winds to simulate stochastic sequences of daily rainfall. The four methods are assessed in terms of seasonal statistics of daily rainfall, including seasonal amount, rainfall frequency, and the mean length of wet and dry spells. Verification measures used are mean bias error, interannual anomaly correlation, root-mean-square error, and ranked probability skill score.The k-nearest neighbor and weather-type classification are shown to perform similarly well in reproducing the mean seasonal cycle, interannual variability of seasonal amount, daily rainfall frequency, as well as the mean length of dry and wet spells, and generally slightly better than the nonhomogeneous hidden Markov model. All three methods are shown to outperform the simple local scaling method. This is due to (i) the ability of the GCM to reproduce remarkably well the mean seasonal cycle and the transitions between weather types defined from reanalysis and (ii) the GCM's moderate-to-strong skill in reproducing the interannual variability of the frequency occurrence of the weather types that strongly influence the interannual variability of rainfall in Senegal. In contrast, the local scaling exaggerates the length of wet and dry spells and reproduces less accurately the interannual variability of the seasonal-averaged amounts, occurrences, and dry/wet spells. This failure is attributed primarily to systematic errors in the GCM's precipitation simulation.
256CFTimes Cited:1Cited References Count:40