For model training within different regions, gages should be weighted by the probability that they belong within that region.
Then, the final prediction for each site should be a weighted combination of the models trained in each region for which the reach has > 0 probability.