Urban tree failure probability prediction based on dendrometric aspects and machine learning models

Abstract

Urban forests provide many benefits for municipalities and their residents, including air quality improvement, urban atmosphere cooling, and pluvial flooding reduction. Monitoring the trees is one of the tasks among the several urban forest assessment procedures. Trees with a risk of falling may threaten the locals and the infrastructure of the cities, thereby being an immediate concern for forestry managers. In general, a set of measures and aspects are collected from field survey analysis to estimate whether the trees represent a risk to the safety of the urban spaces. However, gathering the tree's physical measures in fieldwork campaigns is time-consuming and laborious considering the massive number of trees in the cities. Therefore, there is an urge for new computational-based methodologies, especially those related to the latest advances in artificial intelligence, to accelerate the assessment of trees in the municipality areas. In this sense, this work aims at using several machine learning-based methods in the context of tree condition inspection. Particularly, we present the prediction of the tree failure probability by using several aspects collected over time from fieldwork campaigns, with a special focus on external physical measures of the trees. Further, we provide the samples with their respective tree failure probability values as a new open dataset for further investigations on tree status monitoring. We also present a novel dataset composed of images of trees with bounding boxes delineations of the tree, trunk, and crown for automating the tree monitoring tasks. Regarding the tree failure probability estimation, we compared several regression algorithms for estimating the tree failure likelihood. Moreover, we propose a stacking generalization approach to enhance forecast accuracy and minimize prediction errors. The results showed the viability of the proposed method as an auxiliary tool in tree analysis tasks, which attained the lowest average Mean Absolute Error of 5.6901±1.1709 yielded by the stacking generalization model.