Georeferencing and Classification of Urban Trees Using Aerial and Street-Level Images, and Machine Learning

This article discusses using machine learning and remote sensing to identify and classify tree species. Unmanned Aerial Vehicles (UAVs) captured aerial images, while a GPS-based system ensured accurate georeferencing. Ground-level photos, taken by a 360º camera, provided a thorough view of the study area. A large dataset of aerial and ground images was created to train models like Random Forests, Support Vector Machines, and Convolutional Neural Networks (CNNs), focusing on the classification of tree species. Challenges include intraclass variability (differences within the same species) and interclass similarity (similarities between species), complicating the classification process. Color balance issues in cameras can also introduce unwanted hues. By integrating aerial and ground images with advanced machine learning algorithms, classification accuracy is expected to improve. Studies using CNNs, such as U-net, have demonstrated up to 84% accuracy in vegetation classification with high-resolution images. U-net’s proven ability in image segmentation makes it a reliable tool for geospatial data analysis, and applying it here could significantly enhance the classification and mapping of tree species.