The fluorescence absorption and emission spectra I'm familiar with are fairly smooth, and look like a superposition of several bumps with varying heights and widths. This suggests that their intrinsic dimensionality is much lower than the number of densely sampled wavelengths at which they're typically measured.
A simple choice would be to use PCA. It's not specific to spectra, but is a tried and true method that works well for many signals. Reduce the dimensionality by discarding the lowest variance components, which will probably be noise.
Along the lines of the 'bumps' description, you might try performing a basis expansion using a set of Gaussian (or other bump-like) basis functions with different positions and widths. This expresses each spectrum as a linear combination of basis functions. The weights would be used as your features. I suspect the spectra would be relatively sparse in this basis--that is, only a few basis functions would needed to approximate each spectrum fairly well. Regularization can be used to enforce sparsity (i.e. force most weights to zero) and de-noise the signal. A good method to use here is called 'basis pursuit denoising' in the signal processing literature and 'lasso' in the statistics literature. Depending on the nature of your spectra, it might be possible to find a small, shared set of basis functions for all spectra, which would let you reduce the dimensionality. More about this approach here.
The emission spectrum of a simple mixture of fluorophores is a linear combination of the spectra of the individual fluorophores. If all of your spectra are from mixtures of known fluorophores (and assuming you're interested in emission spectra), a good thing to do would be to measure reference spectra for the individual fluorophores. Then, treat these as basis functions and fit the spectrum of each mixture as a linear combination of basis functions (e.g. using nonnegative least squares, since the weights can't physically be negative). The weights would be used as your features.