Utilizing a PLSR-Based Band-Selection Procedure for Spectral Feature Characterization of Floristic Gradients

The study introduces a new approach for the characterization of floristic gradients by hyperspectral features in a partial least squares regression (PLSR) framework. As ecological factors influence the composition of vegetation, our study is aimed to reveal related effects on spectral signatures. For this purpose, the variation of plant species in an open dryland area was projected into a three-dimensional ordination space using nonmetric multidimensional scaling (NMDS). Subsequently, ordination axes score rotations were performed in 180. semicircles and the waveband-specific correlation to spectral field measurements of reflectance, continuum removed, and first-derivative spectra were extracted. A bootstrapped PLSR modeling was applied over the entire rotation space using varying numbers of correlated spectral variables as input samples. On that basis, a new PLSR model suitability term was defined by isosurfaces that are spanned over ordination regions where PLSR latent vector (LV) number and PLSR R-2 variance is minimized. It incorporates model performance evaluation with feature characterization using weighted frequencies of spectral variable input in suitable ordination areas. Final PLSR suitability surfaces were transferred to image spectra to prove feature stability and model performance. Our investigation supports the assumption that spectral features are separable to distinct ordination space regions that can be related to individual species gradients. Thereby, the selection of an optimal PLSR model crucially depends on the spectral transformation technique. We further show that stable PLSR models can be derived in multiple ordination directions whereby an appropriate variable selection using suitability surface optimization reduces feature mismatch between field and image spectra.