Characterization of wetland vegetation distribution using satellite imagery.

Abstract

Distribution of wetland vegetation along an altitudinal gradient has been investigated using various ground, aerial and space-based techniques, yielding results with varying accuracies. The advent of space-based hyperspectral satellite imagery, however, provided an opportunity of adding to a pool of resources used by environmental managers in understanding wetland vegetation. This research investigates the use of Hyperion, a satellite hyperspectral imagery, to characterize wetland vegetation along an altitudinal gradient in terms of the hydrological regime determined indirectly by means of elevation change. Hyperion imagery was captured over the Mfabeni wetland of the Isimangaliso World Heritage Wetland Park. A vegetation map of the Mfabeni wetland that served as a secondary source of ground truthing data, along with a topographic base map were then acquired. A 5m interval contour map and spot-heights for use to generate and assess the accuracy of a Digital Elevation Model (DEM) of the Mfabeni wetland were also acquired. A vegetation map of the Mfabeni wetland was georeferenced to a topographic base map and transformed to the spatial reference of the Hyperion image. The georeferenced vegetation map was then superimposed over the Hyperion image and the vegetation assemblages’ boundaries digitized into a shapefile. Vegetation assemblage classes were then defined. The Hyperion image was radiometrically calibrated to apparent surface reflectance values. Pixel based spectra were then randomly extracted from the Hyperion image for each of the vegetation assemblages and used to compute class spectra means and to identify optimal wavebands, using Random Forest and variable elimination. Class spectra means were used to create a spectral library with class spectra means as endmembers. Regions of Interest (ROIs) representing each of the vegetation assemblages were then extracted. Spectral Angle Mapper (SAM) algorithm was then used to classify the Hyperion image using the ROIs, class spectral means held in the spectral library and Hyperion image spectral subset corresponding to the identified optimal bands. The classified images were then assessed for accuracy. DEMs were also generated using various interpolation techniques and assessed for accuracy. The best classified image and best DEM were then overlain to create a composite image used for characterization of vegetation assemblages in terms of elevation. The results showed that it is possible to characterize wetland vegetation assemblages in terms of elevation along an altitudinal gradient.