A new statistical method for downscaling is introduced and is based on application of neural networks models and kriging algorithm to estimated atmospheric variables in three-dimensional space. Artificial neural networks (ANN) were used to interpolate atmospheric data in the vertical domain while the kriging algorithm was used to perform horizontal interpolation. This methodology was used to covert a very limited radiosonde data set for the Caribbean basin to a gridded set of data at 11 vertical levels. The proposed estimation scheme can be used to establish the initial and boundary conditions to run a regional atmospheric model. Estimation of atmospheric variables can also be used to identify short- and long-term correlations between upper-air variables with special climate events, such as hurricane intensity and tracking. Atmospheric variables estimated on solely radiosonde observations were compared with National Center for Environmental Prediction (NCEP) data. Sixteen days were randomly selected from two dry and two wet months in an arbitrary year, 1995. Estimation based on radiosonde observations was in agreement with NCEP data for the eleven mandatory pressure levels. Furthermore, cross-validation techniques with five years (24,000 soundings) of radiosonde observations were used to test the proposed methodology and satisfactory results were obtained. Cross-validation results indicate that the proposed methodology is a reliable tool to obtain estimation of atmospheric variables.

The proposed estimation scheme can be used to predict atmospheric variables 24-hours ahead with a high resolution based on radiosonde and NCEP data. Furthermore, the suggested scheme can be used to perform almost real time estimation of atmospheric variables. For instance, sounding information obtained from a ship, or information from airplane recognition can be added to the regular sounding and NCEP data to derive a better initial and boundary conditions, and consequently, regional numerical models will increase accuracy in hurricane tracking prediction.