Archived remote sensing imagery provides valuable data for assessing long-term biodiversity trends in urban environments. However, accurate delineation of individual tree crowns–a prerequisite for reliable biodiversity assessment–remains challenging due to factors such as overlapping canopies, heterogeneous backgrounds, limited spectral information, and diverse tree crown sizes. In this study, we propose Urban-TreeSeg, a deep learning framework specifically designed for urban tree crown delineation from archived aerial RGB imagery. The framework integrates three key components: (1) a CBAM Residual UNet (CRUNet) architecture for enhanced multiscale feature extraction, (2) a classification branch to improve separation of tree crowns from urban backgrounds, and (3) star-convex polygon representation to better capture irregular crown shapes. Urban-TreeSeg was evaluated in two contrasting urban landscapes: a residential neighborhood and a densely wooded park, achieving F1-scores exceeding 85% and outperforming the widely used Mask R-CNN model by over 4%. We further validate that tree crown delineation in the Fall 2020 imagery can be improved by reusing knowledge learned from Summer 2008 imagery. These results demonstrate the potential of the proposed approach for consistent and robust extraction of urban tree crown information from archived imagery, supporting long-term urban vegetation monitoring and biodiversity assessment.
Tong et al. (Fri,) studied this question.