Leaf spectra aggregate over plant structure and chemistry, which are produced from the interaction of environmental and the genetic variation at the population and phylogenetic scales. Uncovering the signatures of phylogeny in leaf spectra is especially important for biodiversity monitoring applications and functional analyses at broad evolutionary and spatial scales. Yet, no model for linking spectra to phylogenies has been proposed. Challenges to modeling the evolution of spectra include the pervasive autocorrelation among spectral regions and the high dimensionality of the spectral data. Our goal was to develop a model that estimates which spectral regions are more strongly associated with deep evolutionary history.
Results/Conclusions
We first constructed a multivariate brownian motion model of trait evolution that estimates the within spectrum covariance as well as the rates of evolution of different spectral regions given a fixed phylogeny. Simulations showed that the performance of this model degrades rapidly as spectral data resolution increases. That is, it does not have enough power to cope with the high dimensionality of raw spectra. To mitigate the high dimensionality problem, we extended the model above with a first stage gaussian mixture model (GMM) that approximates the spectrum with N latent gaussians. We then estimate evolutionary rates and covariances on the latent gaussians using the multivariate model of evolution described first. Our results show that the best fitting Ns are generally much smaller than the size of the raw spectra. Preliminary results using simulated data also indicate that the full model that couples the GMM and the brownian motion models can recover the true evolutionary rates of different spectral regions.