Identification of Leucaena leucocephala in Urban Landscapes Using Street-Level Images and Deep Learning

Mapping urban tree species supports green infrastructure planning. An essential issue refers to the monitoring of exotic species that may become invasive. Street-level imagery provides a complementary perspective to aerial images for species identification that are difficult to distinguish from above. In this context, our study aimed to evaluate deep learning-based object detection and image segmentation approaches to identify a potentially invasive tree species known as Leucaena leucocephala in an urban environment in Brazil, using 422 street-level images acquired from Google Street View (SV) and mobile phones (MPs). Object detection models (YOLOv8 and DETR) and a foundation segmentation model (SAM, zero-shot) were applied to assess how deep learning paradigms perform under heterogeneous urban imaging conditions. YOLOv8 achieved detection performance with mAP50 above 0.83 and recall up to 0.76. DETR showed domain sensitivity, with mAP50 of 0.45 in SV images and 0.84 in MP imagery. For segmentation, SAM zero-shot achieved 0.92 accuracy and 0.93 F1-score in SV images, decreasing to 0.63 accuracy and 0.66 F1-score in MP images. Overall, this study demonstrates that combining detection and segmentation approaches provides complementary information for urban vegetation monitoring, supporting decision-making related to invasive species management and sustainable urban landscape planning.