Influence of Image Segmentation Parameters on Positional and Spectral Quality of the Derived Objects

With the launch of very high spatial resolution sensors, a good positional quality can be expected from spatial imagery. Furthermore, efficient segmentation algorithms can increase the productivity and reduce the subjectivity of manual delineation. However, segmentation algorithms rely on a combination of parameters. This study used two indices to assess the quality of a segmentation in a forested landscape for 32 parameter combinations. The study area was a forest of Southern Belgium covered by an IKONOS image. A manual delineation based on a 1 : 10 000 vector database was used for the validation. The segmentation algorithm from the eCognition ® software was used to produce the image objects. These image objects were then evaluated in terms of (1) positional accuracy and (2) feature consistency prior to classification. The positional quality assessment quantified the errors of location along the edges thanks to estimates of the positional accuracy and precision, that is bias and range of errors. Beside this, the consistency of object features was compared using a distance measure in a 6 dimensional feature space. The general tendency was an increased class-by-class discrimination and a lower edge precision when the scale parameter increased. Furthermore, when using a shape parameter, the compactness often improved the overall segmentation quality. However, the separation of deciduous and coniferous forests, which was the lowest of all classes, did not significantly increase with scale parameter. A small scale parameter seems therefore to be advisable. * Corresponding author