Abstract ID: 250

Removing vegetation canopy bias from SRTM elevation data using an objective sampling technique

The height reported in Shuttle Radar Topography Mission (SRTM) elevation data is a complex function of forest canopy density and structure as well as land surface characteristics. Abrupt vegetation discontinuities from clear-cutting produce errors that are clearly visible upon close inspection, and present problems when attempting to use the data for modeling and hydrological analysis. Here we present an objective, robust sampling technique for removing the vegetation bias from SRTM data. The test area for the development of this technique was the 84,000 hectare Fazenda Tanguro in the Upper Xingu River basin. A Landsat TM image from August 7, 2000 was selected (closely corresponding to the SRTM acquisition date), and after correcting a geometric misalignment between the two data layers, an unsupervised classification was performed on the TM image to generate enough classes to capture the height diversity of forests in the region of interest. Classes were put into bare soil and vegetated groups. Forest edges were defined by the border between these groups, and buffers were generated along both the inside and outside of forest edges to define sampling zones. Pairs of adjacent sample points generated across this edge were used to extract vegetation class values for forested pixels, and SRTM heights corresponding to the vegetated and non-vegetated areas. After eliminating sample pairs near stream channels and other areas of localized terrain transition, statistics on height-difference between points inside and outside the forest were extracted for each vegetation class. The average height difference for each class was then subtracted from the SRTM data, and the result smoothed with a low-pass filter. Stream channel and basin delineations produced from the corrected SRTM data gave much better results than those conducted prior to correction . This technique lends itself to partial automation for application to much larger areas, with operator input needed only for checking registration between the two data layers and grouping classes to define vegetation boundaries.

Session:  LCLUC and Human Dimensions - Current and future trends of land-use/land-cover change and agricultural intensification.

Presentation Type:  Poster