Predicting Urban Tree Falls: A Machine Learning-Based Approach Using Low-Cost MEMS

This study proposes a technique for predicting tree-falling hazards using data collected by low-cost MEMS (Micro-Electro-Mechanical Systems) accelerometers and a machine learning algorithm. In Hong Kong, about 8,000 sensors were installed in urban trees, recording roll, pitch and the calculated tilt angles during 2021 and 2024. A temperature compensation algorithm based on Random Sample Consensus (RANSAC) was applied to correct the temperature drafting. 17 and 28 fallen trees during the typhoon periods were detected from August 20, 2021, to August 31, 2023. A tree selection algorithm that considered factors such as distance between healthy and fallen trees, diameter at breast height (DBH), tree height, and crown spread was implemented to select healthy trees as the control sample for model training. Features such as tilt angle, absolute tilt angle difference, z-score and angle variance over intervals of 7, 15, 30, 60, and 90 days were used for model construction. Fallen trees and healthy trees as controls were trained using logistic regression with L2 regularization and an Extra Tree Classifier (ETC), respectively with Leave-One-Out Cross-validation (LOOCV) and Stratified 5-Fold Cross-validation with 10 Repeats. Models showed the best fit, with McFadden's Pseudo R2 ranging from 0.5993 to 0.7286 and binary log loss from 0.0806 to 0.1110. The ETC model using the typhoon dataset performed best, with absolute tilt angle change being a significant contribution. Probability predictions were calculated during the testing period and fitted to probability density functions, with Generalized Extreme Value (GEV) and Pareto functions being the best fits. A color-coded warning system based on percentile thresholds (PR) of these PDFs was developed, potentially providing alerts up to two months in advance, helping maintenance professionals to prioritize trees with higher falling risk and take precautionary actions.