Due to the pandemic we had to cut short an experiment (that was done mainly outside) leaving us with less than 30 Vis-NIR samples, and now we are thinking of generating new synthetic data from the spectra collected. Methods like SMOTE might be useful, but they keep the balance between clases, and are not a way to directly generate "new" samples, I am thinking of a poisson bootstrap after a (non-linear) regression. Any ideas?
Thanks