This study proposes a method which can be used to provide guidelines to aircraft reconnaissance for hurricane predictions. The method combines numerical weather prediction (NWP) model with statistics approach to target adaptive observations to the areas to where the hurricane predictions are very sensitive in the NWP model. A control experiment is performed from regular initial analysis, while 50 other experiments are performed from slightly randomly perturbed initial analysis. Under the perfect model assumption, the control experiment is served as the truth of atmosphere. The forecast error variances at certain forecast hour, e.g. hour 48, can then be computed and maximum variances can be easily located. After the locations of the of the maximum forecast error variances are known, various correlations of different variables between these maximum variance points and the perturbation fields at target hour, e.g. hour 12, are calculated to find out the locations of the perturbations at target hour which cause the maximum forecast error variances at forecast hour. Statistically, these correlation fields indicate where the most sensitive areas are at target hour, that is where the additional observations are needed.
The hurricane Fran of 1996 is used to examine the proposed method. The reason to choose this case is that, for the first 48 hour forecast, the track forecast from NWP is very close to the best track. Two experiments are designed to examine the method. One experiment updates predicted variables at target hour (12h) over the areas, to where the proposed method indicates the forecast will be sensitive, by combining truth and first guess fields. Another experiment also modifies predicted variables at target hour (12h), but over the areas where the method indicates the forecast is less correlated. The results show that the modification has greatly reduced the forecast error variances at forecast hour (48h) in the first experiment, while has very little impact on the variance fields at forecast hour (48h) in the second experiment. It is very clear from our experiments, the proposed method is able to target the sensitive areas, over where additional observations can help to reduce hurricane prediction from NWP model.