Remote sensing techniques for forest parameter assessment: Multispectral classification and linear spectral mixture analysis

Abstract

One of the most common applications of remote sensing in forestry is the production of thematic maps, depicting e.g. tree species or stand age, by means of image classification. Nevertheless, the absolute quantification of stand variables is even more essential for forest inventories. For both issues, satellite data are attractive for their large-area and up-to-date mapping capacities. This study followed two steps, and at first a supervised parametric classification was performed for a German test site based on a radiometrically corrected Landsat-5 TM scene. There, eight forest classes were identified with an overall accuracy of 87.5%. In the following, the study focused on the estimation of one key stand variable, the stem number per hectare (SN), which was carried out for a number of Norway spruce stands that had been clearly identified in the multispectral classification. For the estimation of SN, the approach of Linear Spectral Mixture Analysis (LSMA) was found to be clearly more effective than spectral indices. LSMA is based on the premise that measured reflectances can be linearly modelled from a set of so-called endmember spectra. In this study, the endmember sets were held variable to decompose pixel values to abundances of a vegetation, a background (soil, litter, bark) and a shade fraction. Forest structure determines the visible portions of these fractions, and therefore, a multiple regression using them as predictor variables provided the best SN estimates. LSMA allows a pixel-by-pixel quantification of SN for complete satellite images. This opens the view to exploit these data for an improved calibration of large-scale multi-parameter assessment strategies (e.g. statistical modelling or the kNN method for satellite data interpretation).