Support vector machine classification of woody patches in New Zealand from synthetic aperture radar and optical data, with LiDAR training

Abstract

A light detection and ranging canopy height model (CHM) was used as training data for a segment-based classification of woody patches. The classifier is accurate (∼92%∼92%) and suitable for use at the national scale. Height thresholds and percentage cover of vegetation from the CHM were used to produce larger quantities of reliable training data compared to other, mostly point or plot-based, ground-truthing approaches. It was found that the regional-scale differentiation between woody and nonwoody vegetation might be achieved by a combination of L-band dual-polarized Phased Array type L-band synthetic aperture radar data (HV) with multispectral optical data that include a short-wave infrared band. The application of a support vector machine algorithm to these data proved successful. The versatility of these algorithms regarding the discrimination function and their ability to solve classification problems with multiple output classes were critical factors for success. The identified and classified woody patches constitute a valuable addition and enhancement of the national land cover database.