Spatial Transferability of Machine Learning Models for Urban Tree Species Classification using Sentinel-2 Multispectral Imagery

Urban tree species classification supports urban forest management, biodiversity mapping, and assessments of the urban environment. Machine learning models have shown promise for urban species classification using remote sensing data, but their spatial transferability remains unexplored in urban environments. This study evaluates the spatial transferability of Extreme Gradient Boosting (XGBoost), a classical machine learning algorithm, for urban species classification using Sentinel-2 imagery. Fifteen tree species common to New York, New York; Portland, Oregon; and Seattle, Washington were selected from tree inventories. Time-series Sentinel-2 Level-2A surface reflectance imagery and NDVI data from 2024 were used to classify samples. Random oversampling and a unified sample size were both tested. Using random oversampling, on-site models were moderately successful, achieving overall accuracies of 0.50 to 0.58 and kappa values of 0.40 to 0.51, although accuracy may be attributed to class imbalance. When models were transferred to use testing data from alternate study areas, overall accuracy and kappa values decreased significantly, to 0.10-0.24 and 0.04-0.12, respectively. Using balanced data, models achieved overall accuracies of 0.21 to 0.43, and kappa values of 0.16 to 0.39. Overall accuracy decreased to 0.07 to 0.16, and kappa values decreased to 0.01 to 0.10. The results show the limited spatial transferability of classical machine learning models for urban tree species classification when using Sentinel-2 data, due to severe class imbalances, amplified through model transfer, phenological and data availability differences, and spatial context. This study emphasizes the benefits of localized training data for urban tree species classification and highlights the decrease in accuracy for transferred urban tree species classification models using Sentinel-2 imagery.