Use of Machine Learning Techniques for Urban Tree Detection: A Systematic Literature Review and a Synthesis of Operational Workflows

Monitoring urban tree cover is essential for quality of life and thermal regulation, but the detection of individual trees still faces high costs and limited scalability in field methods. This systematic review, conducted in accordance with the PRISMA protocol, analyzed 38 original articles to identify how remote sensing and artificial intelligence automate the detection, segmentation, and classification of trees in urban areas. The results show a predominance of aerial platforms (drones and airplanes), often combining RGB and LiDAR sensors to capture spectral and three-dimensional information from tree canopies. Deep learning techniques, particularly those from the YOLO family, are predominant, achieving accuracy exceeding 95% under optimal conditions—specifically, low-to-moderate canopy density with isolated or regularly spaced trees—with performance comparable to manual methods, yet with significant cost reductions and gains in scale. Building on this evidence, a five-phase Synthesis Workflow—spanning data acquisition, pre-processing, modeling, post-processing, and evaluation—is proposed as a methodological synthesis of the practices recurring across the 38 studies, with post-processing recommendations drawn from the high-performing ones, rather than an independently validated framework, offering researchers and practitioners an evidence-based reference for implementation. It is concluded that the integration of remote sensing and computer vision already offers mature and operational solutions for urban tree inventory. However, significant challenges remain: the standardization of evaluation metrics, the generalization of models across different cities, and the reliance on manual annotation for training, all of which still require significant advances for large-scale implementation.