Quantitative assessment of urban tree fall risk using ensemble learning: integrating structural, defect, and meteorological features

Urban trees provide critical ecosystem services yet pose significant public safety risks when structurally compromised. Existing risk assessment frameworks rely primarily on on-site arborist inspection, which is resource-intensive and difficult to scale across large municipal inventories. This study develops a quantitative fall-risk classification framework using 5,000 New York City Forestry Tree Points with recorded professional risk ratings, structural attributes, species, borough assignment, spatial blocks, and date-aligned NASA POWER meteorological variables. A proportional borough × principal-species/Other × DBH-class sampling design preserved the eligible population structure while supporting transparent model evaluation. Five-fold spatial-block validation compared Logistic Regression, Random Forest, XGBoost, LightGBM, and Stacking models for distinguishing high-risk trees from low and moderate risk trees. The Random Forest model provided the strongest structural-only ROC-AUC (0.809; 95% bootstrap CI: 0.773–0.842) and PR-AUC of 0.122. Coarse meteorological variables yielded only a small and uncertain incremental change in Random Forest discrimination (Delta ROC-AUC = 0.006; 95% CI: −0.011 to 0.025), emphasising the need for finer storm, wind-canyon, and root-zone measurements. Sensitivity analyses showed modest performance change after excluding Manhattan (ROC-AUC = 0.801), a larger decrease when species was excluded (ROC-AUC = 0.783), and minimal change when borough was excluded (ROC-AUC = 0.804). The framework is best interpreted as a decision-support tool for inspection prioritisation rather than a replacement for field arborist assessment.