Single Tree Detection in Very High Resolution Remote Sensing Data.

Tree detection is a major focus in the field of (semi-) automatic extraction of forest information from VHR data. Many existing tools require a set of seed pixels with which to start segmentation. In this study, different methods of obtaining seeds (semi-) automatically from both orthophotos and digital surface models (DSM) derived from stereo imagery are tested and compared. The evaluation is performed based on field measurements and visual aerial photo interpretation. The results for 3D seed generation using the local maximum approach (LMA) are very poor; apparently, the smoothing effect of the DSM is too strong to model single trees. Seed generation based on orthophotos performed better: for a dense, natural forest, the “morph” algorithm detected 64% of the trees visible in the aerial photos by an error of around 25% both for commission and omission. Compared to the field measurements, the results (correct 47%) are worse, as suppressed trees cannot be detected in optical data. The LMA based on orthophoto generally led to slightly lower accuracies and more multiple hits than the morph algorithm. Further studies are needed to analyze in detail the dependence of successful tree detection on different tree parameters such as height, crown diameter or tree species. Another important task is the use of the derived seed points in the region growing process in order to evaluate their applicability and accuracy concerning tree species segregation and timber volume estimation.