| Abstract [eng] |
The European Commission has adopted a nature restoration law that requires green spaces to expand to a satisfactory level, and the area covered by these spaces as well as tree canopy should not decrease by 2030, compared to the year the regulation came into force. In the Republic of Lithuania, the inventory of trees and green areas is regulated and the change in the law allows the work to be done with remote methods and the municipalities can choose their preferred method. It is expensive and time-consuming to cover a large city with inventory data using human resources, hence this paper attempts to assess how remote sensing data can be used for inventorying urban trees. Since cities like Vilnius can afford to update and reuse remotely collected data like LiDAR or orthophotos, the possibility of extracting physical tree parameters by different methods using LiDAR and R programming language libraries, and tree species identification using hyperspectral imagery is explored. The study compares the two methods for extracting physical parameters of trees, the required data granularity and the software used. R and Python programming language solutions are used to automate parts the work. Once the more robust method for tree parameter extraction has been chosen, the most optimal parameters are selected to process a larger sample. The data obtained were suitable for object-based classification. It was chosen to identify the five most frequent tree species in Kaunas city within the available hyperspectral image coverage. Tree tops and crowns are extracted more reliably using the tree height model than by directly segmenting the point cloud. Preparing the data for machine learning reduces the number of trees that can be used by about a factor of two, but it speeds up the process. The best accuracy achieved was 75,26% for species identification using the polynomial SVM algorithm. The data analysis showed that the results should improve with an increase in the number of trees to train the model, but there is still room to investigate the relevance of the additional features extracted from the spectral images to improve the model. |
