Logo Logo
Switch Language to German
Polewski, P.; Yao, W.; Heurich, M.; Krzystek, P.; Stilla, U. (2015): DETECTION OF SINGLE STANDING DEAD TREES FROM AERIAL COLOR INFRARED IMAGERY BY SEGMENTATION WITH SHAPE AND INTENSITY PRIORS. In: Pia15+Hrigi15 - Joint Isprs Conference, Vol. Ii, Vol. 2-3, No. W4: pp. 181-188


Standing dead trees, known as snags, are an essential factor in maintaining biodiversity in forest ecosystems. Combined with their role as carbon sinks, this makes for a compelling reason to study their spatial distribution. This paper presents an integrated method to detect and delineate individual dead tree crowns from color infrared aerial imagery. Our approach consists of two steps which incorporate statistical information about prior distributions of both the image intensities and the shapes of the target objects. In the first step, we perform a Gaussian Mixture Model clustering in the pixel color space with priors on the cluster means, obtaining up to 3 components corresponding to dead trees, living trees, and shadows. We then refine the dead tree regions using a level set segmentation method enriched with a generative model of the dead trees' shape distribution as well as a discriminative model of their pixel intensity distribution. The iterative application of the statistical shape template yields the set of delineated dead crowns. The prior information enforces the consistency of the template's shape variation with the shape manifold defined by manually labeled training examples, which makes it possible to separate crowns located in close proximity and prevents the formation of large crown clusters. Also, the statistical information built into the segmentation gives rise to an implicit detection scheme, because the shape template evolves towards an empty contour if not enough evidence for the object is present in the image. We test our method on 3 sample plots from the Bavarian Forest National Park with reference data obtained by manually marking individual dead tree polygons in the images. Our results are scenario-dependent and range from a correctness/ completeness of 0.71/0.81 up to 0.77/1, with an average center-of-gravity displacement of 3-5 pixels between the detected and reference polygons.