Derivation of Forest Parameters from UltracamD Data
To guarantee sustainability in forests, accurate information on forest parameters is needed for their proper management. In this thesis, UltracamD digital frame camera data is investigated for the derivation of these forest parameters. Given the typically very high geometrical resolution of Ultracam data, the forest parameters investigated in this thesis are tree detection, individual tree height, species recognition, and tree crown extent. Because the derivation of these parameters is a complex task comprising different pre-processing and processing steps, this thesis examines which sequence of steps is most suitable. Throughout the study, special regard is given to the integration of optical and height data. In order to investigate the transferability of the methods, two testsites were chosen in Austria and one in Finland, the latter one also being used in the EuroSDR & ISPRS Tree Extraction Test.
Altogether, twelve different methods of pansharpening are tested on their suitability for the subsequent forest parameter derivation. Through a four step evaluation procedure, YIQ, Brovey and Texture-based weighted fusion were identified to be equally well suited. A new method to generate highly precise digital surface models from multi image matching is presented and the resulting DSMs are compared to traditional stereo matching results. It is demonstrated that by using multiple overlapping images, the mean height error of individual trees can be reduced from 134 cm to 77 cm.
Multitemporal data from one vegetation period is used to investigate the influence of the acquisition date on the separability of tree species classes. It is shown that there is a clear impact on the separability of tree species; images acquired in spring are to be preferred over ones acquired in summer. Furthermore, some methods of tree detection and tree crown delineation based on both 2D and 3D data are tested.
To work out the best sequence of processing steps, two testcases were selected: Burgau in Austria and Espoonlahti in Finland. The results in test case Burgau show a large variation in accuracy depending on the type of forest. For evenly aged coniferous dominated stands, the results are much better than for younger, denser and/or deciduous dominated stands. Regarding tree delineations, the automated algorithms reach correct hit rates similar to those obtained by visual interpretation, however with much higher commission errors. The results in test case Espoonlahti show that with the selected processing chain, 75% of the well matched reference trees were correctly detected. In this test case, the combined use of height and optical data decreased the mean location error from 150 cm to 80 cm.