City managers are responsible for ensuring a diverse and well-functioning urban forest. To accomplish this feat, planners rely on public tree inventories that are often incomplete and outdated due to time-consuming and costly surveys. Remote sensing techniques offer a way to increase the speed and scope of tree inventories while also detecting some private trees. Mobile terrestrial LiDAR and street view imagery are two well-adapted methods for urban tree mapping though they are rarely used in combination despite their potential for synchronous application. We classified 14 of the most common tree genera in Southern Quebec’s urban environments to test whether such a combination of mobile data sources can improve tree classification. Data spanned 3625 trees from four cities in Southern Quebec. We (1) transformed 3D LiDAR scans into 2D images to reduce computation costs, (2) extracted street view images, and (3) applied a multiview technique on both datasets whereby multiple photos from the same tree were stacked. We constructed three Residual Neural Network models, two were fed solely LiDAR or street view images, and the third incorporated both data sources using a late fusion approach. Combining data sources improved classification for most genera. The three conifer genera ( Picea , Pinus , and Thuja ) were among the most successfully classified (F1-score 0.8). Genera dominated by a single species, like Celtis , Syringa , and Tilia , were well classified, while multi-species/cultivar genera such as Acer and Ulmus were a point of confusion for the models. Overall, these results represent advances towards quicker and more complete urban tree surveys using emerging land-based and easily implemented technologies.
Poirier et al. (Fri,) studied this question.