Chapter 6 - Field-Scale Sensitivity of Vegetation Discrimination to Hyperspectral Reflectance and Coupled Statistics

Abstract

Remote sensing of land covers utilizes an increasing number of methods for spectral reflectance processing and its accompanying statistics to discriminate between the covers' spectral signatures at various scales. To this end, the present chapter deals with the field-scale sensitivity of the vegetation spectral discrimination to the most common types of reflectance (unaltered and continuum-removed) and statistical tests (parametric and nonparametric analysis of variance). It is divided into two distinct parts. The first part summarizes the current knowledge in relation to vegetation discrimination spectral analysis at field scale. The results clearly show that (1) an increasing number of studies demonstrate distinct spectral signature for many plants, given the subtle characteristics of the vegetation signature and (2) the studies handle the complexity of the vegetation spectral signature—primarily its shape—using various processing methods, although the use of unaltered and continuum-removed reflectance predominates. The second part of the chapter aims to disentangle whether there is a general trend for preferred statistical test and reflectance type that eases vegetation spectral discrimination or discrimination analysis is always case sensitive. Hence, it conducts the first systematic monofactor sensitivity analysis of vegetation spectral discrimination under different spectral settings and biogeographical regions. The results overall point to continuum-removed reflectance as being a more powerful input to a nonparametric analysis for discrimination at the field scale, when compared with unaltered reflectance and parametric analysis. However, the discrimination outputs interact and are very sensitive to the number of observations—an important implication for the design of the field campaigns and the utilization of the spectral libraries for discrimination and classification purpose. All in all, this work contributes to better understand the uncertainty of the field-scale discrimination results due to the type of reflectance, the nature of the statistical test, and the number of observations in the spectral libraries, to improve estimation of land cover from hyperspectral remote sensing data, and it also facilitates the upscaling of the results to larger scales used in land cover delineation and environmental management.