Predicting the content of heavy metals, such as copper (Cu) and zinc (Zn), based on spectral modeling through Vis-NIR-SWIR spectroscopy is a challenging task, especially when the database comprises soil samples with high variations in Cu and Zn content, associated with pedological, geologic and climate diversity. The aim of this study was to assess whether DB-global stratification, which is based on physiographic region criteria, can improve the accuracy of stratified regional models to predict soil available Cu and Zn content, in comparison with nonstratified global models (GM). Furthermore, we tested the applicability of regional models (RMs) for accurately estimating Cu and Zn in vineyard soils in southern Brazil in comparison with global models. We used a DB-Global model with 1,454 samples derived from 3 different physiographic regions in Rio Grande do Sul State, Brazil. The prediction models were developed via random forest models with spectra subjected to smoothing based on Savitzky-Golay's 1st derivative. DB-Global stratification based on physiographic regions has shown that grouping the most homogeneous samples increases the prediction accuracy of regional models when they are applied to samples from the specific regions for which they are calibrated. The most accurate predictions were recorded for models calibrated with data in databases with the largest number of samples and with the lowest standard deviations of the Cu, Zn, organic matter and soil clay content. The definition for the calibration and application of a GM, in comparison with an RM, must consider soil pedological diversity in a given region.