Airborne multi-seasonal LiDAR and hyperspectral data integration for individual tree-level classification in urban green spaces at city scale

Accurate identification and classification of tree species on a large scale are crucial for the effective management of urban green spaces; however, previous research combining airborne sensors such as LiDAR and hyperspectral imaging for tree classification has generally focused on smaller areas or specific sites, with limited studies applying this approach at the city-wide scale. This study focuses on the utilization of multi-temporal airborne light detection and ranging (LiDAR) and hyperspectral imaging (HSI) data for the classification of 10 species of urban trees at the city scale, which collectively cover over 95 % of the tree-covered areas within the city. Our objective is to evaluate the utility of metrics and indices derived from LiDAR (leaf-on/leaf-off) and HSI (peak growing season/autumn senescence) data in a 35.86 km² urban green space in Gwacheon, Republic of Korea. A comprehensive set of 15 independent variables was extracted from preprocessed and calibrated airborne LiDAR data (footprint size: 0.46 m, density: 42.7 points/m²) and HSI data (127 bands, 400–970 nm range, spatial resolution: 0.68 m) to train seven machine learning classifiers. The model was trained on a stratified random sample of 21,826 tree crown polygon samples collected from individual trees surveyed. The results showed that the combination of airborne LiDAR and HSI data from two seasons achieved the highest classification accuracy with the light gradient boosting machine (LGBM) classifier (90.6 %; Kappa: 0.895) for all 10 major tree species across the entire city, especially for Ginkgo, American sycamore, and Yoshino cherry. Among all variables, the maximum tree height (Hmax) and the intersection symmetric difference ratio index (ISDRI) were among the top influential factors for tree species classification accuracy. Hmax, with an importance value of 0.490, is particularly effective due to the characteristics of urban green spaces. ISDRI, with an importance value of 0.336, highlights seasonal leaf volume differences, aiding in species differentiation. The spectral indices acquired during the autumn leaf senescence showed a cumulative shapley additive explanations (SHAP) importance score that was 0.374 points higher than that of the leaf-on period, highlighting the enhanced significance of hyperspectral data from the leaf senescence phase in classifying tree species. The synergistic integration of airborne LiDAR, HSI, and seasonal data gathered during key phenological periods, along with relevant indices, will contribute significantly to urban forest management at the city-wide level.