Quantitative detection of sediment dust analog over green canopy using airborne hyperspectral imagery

Abstract

A smart unmixing approach for quantitative detection of small amounts of dust that settle on the vegetation canopy using hyperspectral (HRS) airborne imagery data is proposed. A dust analog composed of Alumina (Aluminum Oxide Al2O3) powder was artificially spread over vegetation that covered 4 × 4 pixels of the AISA-Dual sensor. The alumina spectral signal could not be extracted using ordinary methods such as supervised classification (e.g. SAM or MTMF), unsupervised classification (Maximum Likelihood or Minimum Distance), and linear unmixing (e.g. MESMA or VCA). Considering the limitations of the above methods for extracting endmembers in a nonlinear domain, we developed a new approach that is capable of detecting the alumina powder from HRS imagery covering the VIS-NIR-SWIR (400–2400 nm) spectral regions. This step wised approach is based on a sequence merge between a decision tree algorithm, several spectral indices and a flexible constrained nonlinear unmixing method. The endmember vectors and abundances are obtained through a gradient-based optimization approach. Ground-truth examination of the results showed that the method is reliable and that it may represent a new frontier for assessing sediment dust contamination on a dark background via airborne sensors.