Analysing urban trees on verges and slopes along a highway using machine learning methods

Abstract

In arboricultural research, data analysis is important to the understanding of the characteristics of urban forest. This study attempted to apply machine learning techniques on a relatively small data set. This research aimed at exploring the biodiversity and structure of tree stands on verges and slopes along a highway, and analysing the influences of habitat characteristics on the tree stands with the aid of machine learning techniques. 53 slopes and 52 verges along San Tin Highway, Hong Kong were surveyed. 7209 trees belonging to 23 species were found. Dimension reduction proved successful in the concise characterisation of urban forest by a biodiversity component and an abundance component. The biodiversity component score of the slopes (0.625) was higher than that of the verges (−0.637). However, the abundance component scores of slopes (−0.059) and verges (0.060) showed slight difference, reflecting comparable tree abundance. A 75–25 train/test split was applied on a data subset consisting of slopes registered under a scheme called Systematic Identification of Maintenance Responsibility of Slopes in the Territory for regression analysis. The scores of the two components were regressed on several slope geophysical variables. Slope height and slope area served as significant predictors explaining biodiversity. Boosting improved the explanatory power and predictive accuracy of the regression model on the biodiversity component, as evidenced by an increase in adjusted R2 by 0.23 and a decrease in RMSE by 0.40. This research proved that component scores can serve as inputs for regression models for the explanation of urban forest characteristics by habitat-related variables. In future, small data sets from tree surveys can be analysed using the workflow demonstrated in this study for the generation of more management insights.