Abstract
Urban tree species classification is essential for green space management but is challenged by the fragmented spatial distributions of trees. To address this issue, this study employs a hyperspectral imagery-based classification method to improve the accuracy of urban forest identification. By leveraging the red-edge spectral reflectance characteristics and optimizing band combinations, this study constructs an optimized vegetation index, NDVI (680, 748), which significantly enhances vegetation extraction in urban environments, achieving an overall accuracy of 97.82% and a Kappa coefficient of 0.95. Using this index, recursive feature elimination with cross-validation was applied to select highly discriminative features, which were then integrated into multiple machine learning models. Among these, the random forest model-incorporating characteristic bands, textural features, and vegetation indices-achieved the best test set performance, with an overall accuracy of 86.91% and a Kappa coefficient of 0.85. This method enables high-precision urban forest mapping, facilitating sustainable ecosystem management.