Using Wasserstein distance and conditionality for stable and selective data augmentation in pavement crack analysis