Urban tree extraction from UAV Imagery using deep learning techniques for urban management applications

Abstract

This paper proposes a method that leverages deep learning techniques to extract urban trees from high-resolution UAV imagery. UAVs play a crucial role in capturing both nadir and oblique images with high geometric detail, facilitating effective modeling of the three-dimensional spatial characteristics of urban vegetation. The core of the proposed approach lies in its ability to automatically identify and segment trees based on learned visual features, particularly the shape and size of tree canopies, which often exhibit irregular boundaries and natural spatial distribution, distinct from man-made objects. The resulting output is a spatial dataset that accurately reflects the location and geometric structure of trees, effectively supporting urban green space management, analysis, and planning. The method demonstrates high feasibility, automation capability, and scalability, contributing to the development of intelligent and sustainable urban management systems.