Data Dimension Reduction and Band Selection using Canopy Spectral Invariants (CSI) Concept

In this paper, we carried out a case study with a multidirectional and hyperspectral data, CHRIS (the Compact High Resolution Imaging Spectrometer) on board platform PROBA (Project for On Board Autonomy). After orthocorrection and atmospheric correction, we analyzed the directional reflectance (HDRF) of a dense conifer forest to get two CSI, i.e. effective value of re-collision probability, pr, and escape probability R₁ at different angles. These two invariants at all five angles were regressed to LVIS_H100 data, which is relevant to canopy height, and demonstrated a comparable statistics to the regression with spectral reflectances. In this sense, CSI method can act as an effective tool to reduce the dimension of hyperspectral remote sensing data by a ratio of N/2 (N is the number of spectral bands). Re-collision probability, p_r, and escape probability R₁ can be understood as two principle components in PCA, however, they are superior to principle components because the transformation is based on radiation transfer physics and reduction result have explicit interpretation. In similar consideration to reduce the redundancy as above, we also calculated the deviation of each spectral band to the fitted value from p_r and R₁ and make a statistics for all the sampling pixels. We picked out the best fitted bands, as most of them locate in NIR range, we employed statistics to determine which to supplement in visible range, and discussed the potential of these bands to be selected as optimal VNIR bands combination of future vegetation sensor design.