An entirely optical, dynamic thermometry technique based on the temperature dependence of a fluorescence spectrum is presented. Different from conventional intensity-based fluorescence thermometry, in this work, neural network recognition is employed to extract the sample temperature from the magnitude and shape of recorded fluorescence spectra. As a demonstration to determine the depth profile of dynamical temperature variations and of the thermal and optical properties of semitransparent samples, in-depth photothermally induced periodical temperature oscillations of a rhodamine B and copper chloride dyed glycerol sample were measured with an accuracy of 4.2 mK·Hz−1/2 and fitted well by a 1D thermal diffusion model.