Testing the woodcock-harward image segmentation algorithm in an area of southern california chaparral and woodland vegetation

Vegetation maps were produced by applying a region-growing segmentation algorithm to Landsat Thematic Mapper (TM) data, and labelling the resulting segments or map polygons by overlay of a per-pixel classification and applying a plurality rule. Thus, each segment was assigned a vegetation class label based on the most frequently occurring pixels in the segment. The segmentation improved overall map accuracies by an average of 10 per cent relative to the underlying per-pixel classification for three subimages within a southern California montane watershed based on a comparison with photointerpreted maps. While it was hypothesized that including transformed slope aspect and image texture as input to the segmentation would improve map accuracy by creating segments corresponding more closely to vegetation stands, our results did not support these hypotheses. Further, performing the segmentation on principal components bands, or a vegetation index, did not improve results over the segmentation based on TM bands 2, 3, and 4.