Assessing the accuracy of UAV-captured images used for individual trree crowns delineation in different structures of an urban forest

Advances in computer vision and the development of remote sensing instruments have made indirect measurement of tree features possible. Individual tree crown delineation is an important step towards information collection and mapping trees in an urban area. This information is then used to help planners design strategies for optimization of urban ecosystem services and adapt to climate changes. Common methods of Individual tree crown delineation (ITCD) were based on very high-resolution satellite or Light Detection and Ranging (LiDAR) data. However, satellite data are usually covered by clouds and thus, cannot be appropriate for the measurement of individual trees. Aerial Laser scanning is also relatively expensive. Remote sensing with unmanned aerial vehicle (UAV) captures low altitude imagery and thus, is potentially capable of mapping complex urban vegetation. Automatic delineation of trees with UAV data makes collection of detailed information from trees in large geographic and urban regions possible. Therefore, a multirotor UAV equipped with a high-resolution RGB camera was used in the present study to obtain aerial images and delineate individual trees.

The present study has compared the performance of Inverse watershed segmentation (IWS) and region growing (RG) algorithms using point clouds derived from Structure from Motion (SfM) algorithm and UAV imagery captured with the aim of tree delineation in Fateh urban forest located in Karaj. Region growing (RG) is used to separate regions and distinguish objects in an image. It starts at the initial seed points and determines whether the neighboring pixels should be added to the growing region. If the neighboring pixels are sufficiently similar to the seed pixel, they are labeled as belonging to the seed pixel. To implement the algorithm, the window size and the growing threshold were set for all resolutions. In order to obtain the most appropriate size for the search window, we examined different window sizes with a growing threshold of 0.5 for each resolution. Individual trees delineation was performed for each CHM resolution in the three different sites using "itcSegment" package of R software. Watershed segmentation algorithm is also similar to RG algorithm. The only difference is that it sets the growing seeds at the local minima. In other words, the local maxima in this algorithm change into local minima and vice versa. Inverse Watershed Segmentation (IWS) method was implemented in ArcGIS 10.3 because of its capability in delineation of distinct tree entities. In the summer of 2018, three sites with different structures including a mixed uneven-aged dense stand (site 1), a mixed uneven-aged sparse stand (site 2), and a homogeneous even-aged dense stand (site 3) were surveyed and photographed, and a 3D point cloud was extracted from the images. Then, the performance of algorithms was tested using a series of different canopy height models (CHM) with spatial resolutions of 25, 50, 75, 100, and 120 cm. To generate these models, digital surface model (DSM) was subtracted from digital terrain model (DTM). Results of individual tree delineation were validated using data collected in field observation of the aforementioned sites.

Results indicated that both RG and IWS algorithms yielded their best performance in the dense homogeneous structure. Moreover, the number of segments resulting from CHMs with low resolution was often much more than the actual number of trees. This was due to the occurrence of several peaks within an individual tree crown especially in low resolution images. With an F-score of 0.88, homogeneous even-aged dense stand (site 3) showed the highest overall accuracy in RG algorithm with a pixel size of 75 cm. Generally, results indicated that RG is an appropriate approach for individual tree delineation due to its flexibility in delineation of varying crown sizes. Furthermore, this method does not assume a circular shape for tree crowns and thus, is capable of detecting and segmenting irregular crowns. Generally, delineation of trees in urban forests using CHMs obtained from UAV-captured aerial imagery was highly accurate in homogeneous sites, while such models lacked efficiency in heterogeneous sites.