Regional mapping of spekboom canopy cover using very high resolution aerial imagery

Abstract

Very high resolution canopy cover maps of spekboom are required to assist with the restoration of degraded habitat in the Little Karoo, a large semiarid region in South Africa. Variations in habitat and level of degradation, in addition to radiometric variations in the imagery, make spekboom mapping at a regional scale a challenging problem. We present a per-pixel classification approach for canopy cover mapping of spekboom using multispectral 0.5-m resolution aerial imagery. The imagery is radiometrically homogenized with a technique that uses satellite data to convert digital numbers to estimated surface reflectance values. A feature selection procedure that is robust to redundancy is applied in order to select an informative feature subset from a typical set of spectral, textural, and vegetation index features. Support vector machine, random forest, decision tree, k-nearest neighbor, and Bayes normal classifiers are evaluated against labeled pixel data and canopy cover ground truth acquired at 20 field sites. The results show that all the classifiers, except the Bayes normal classifier, perform well. The decision tree produces the best results (mean absolute canopy cover error of 5.85% with a standard deviation of 4.65%).