Evaluation of survey flight parameters for the accuracy of single tree extraction based on high resolution aerial images

Kurzfassung

Digital Surface Models (DSM) generated with pixel-sharp Semi-Global Matching (SGM) in full image resolution are more and more accepted to set the worldwide standard for photogrammetric image matching. For many mapping and classification tasks the combination of a DSM and a True-Ortho Mosaic (TOM) is better suited than a traditional orthophoto since TOM and DSM are perfectly co-registered and geometrically more correct. Due to the fact that the combination of TOM and DSM has been established as standard for large scale aerials surveys in the last years, TOM and DSM are more and more used as input for the extraction of forest parameters on the stock- and individual tree level. The limits of the used methods are directly linked to the constraints of the used data source. For example understory in coniferous and deciduous forest cannot be extracted or trees with a branch tree trunk cannot be extracted as a single tree. The methods used for the extraction of single trees are the focal statistics analysis and the sink analysis for the detection of tree tops which are implemented in a GIS model. In order to extract the crowns of trees a region growing segmentation and basin analysis are used. In a previous study the achieved completeness was >90 percent with a correctness >85 percent for coniferous trees. The completeness in a deciduous forest came up to 87 percent with a correctness of the extracted trees close to 89 percent. In this paper the extraction accuracies for individual trees and the minimum requirements for an efficient single tree extraction from TOM and DSM are analyzed as function of the geometric properties of the sensor system (e.g. FOV, geometric and radiometric resolution), flight parameters (altitude, overlap along / across track) and the corresponding qualities of DSM and TOM. Aerial images from UltraCam, DMC and the DLR MACS systems with systematically varying flight parameters including those for standard aerial surveys (overlap 80/60, 80/40, 60/40) are used as data base. Depending on the forest structure the best geometric resolution for an efficient single tree extraction in the GIS model ranges from 30 to 50 cm. Further increase of the geometric resolution from 30 to 10 cm does not improve the extraction accuracies. However our analyses show that the flight parameters have a significant influence on the extraction accuracy. For example from a flight altitude of about 7000 meters a nominal geometric resolution of 50 cm from an UltraCamX System. Due to the high flight altitude strong atmospheric effects occur, which have a strong negative influence on the radiometry and image homogeneity and subsequently the image matching (DSM). Hence the classification and the extraction accuracy decrease. Therefore higher resolution aerial images between 20 and 30 cm or better are important as data base for DSM and TOM generation. For the GIS based single tree extraction further resampling of the DSM is often necessary. The second important flight parameter is the overlap along and across track. The overlap depends on the FOV of the camera systems as well as the method of image matching. For example due to FOV the recommended minimum strip side lap for a DSM generation with SGM and DMC1 system is 40 percent. For an UltraCamX system 30 percent strip side lap are sufficient. If the minimum quality requirements to the flight parameters and the resulting TOM and DSM are met, the type of the used sensor system has a fairly low influence on the extraction accuracy because of the robustness of the extraction methods. Commercial camera systems as well as the DLR camera systems lead to comparable results. More detailed results of the ongoing study will be presented at the conference.