The Impact of Image Spatial Resolution and Machine Learning Algorithm on Urban Vegetation Classification: Focus on Data Loss and Misclassification

Urban vegetation classification is challenging due to the heterogeneous nature of urban environments. Accurate mapping of urban vegetation, which plays a crucial role in regulating microclimates, mitigating the urban heat island effect, and supporting environmental sustainability, is essential. High-resolution remote sensing imagery is widely used for urban vegetation classification, but the influence of spatial resolution on mapping accuracy remains underexplored. This research investigates the classification efficiency of various satellite resolutions and machine learning algorithms, assessing the impact of spatial resolution on urban vegetation classification using WorldView-2 imagery resampled from 0.5 to 5 and 10 m. The classification was performed with support vector machine (SVM), random forest (RF), decision tree (DT), and Gaussian Naïve Bayes (GNB) algorithms to examine data loss across resolutions and algorithms. Findings show a significant decrease in accuracy as resolution becomes coarser, with classification accuracy declining from 93%–98% at 0.5 m to 80.31%–91.39% at 5 m, and 78.29%–87.02% at 10 m, with misclassification reaching nearly 22%. SVM and RF achieved the highest performance across resolutions (AUC up to 0.99), followed by DT and GNB. Among endmember types, shrubs demonstrated superior mapping consistency across classifiers, followed by nonvegetation, trees, and grass-based on F-score results. Analysis of pixel transitions revealed substantial losses at coarser resolutions, especially for trees and shrubs, reaching up to 68% in all classes and 50% in vegetation pixels only. These findings highlight the sensitivity of classifiers to resolution-based pixel transitions in urban vegetation, emphasizing the need for careful selection of imagery resolution for accurate urban vegetation mapping.