Plane tree risk assessment in urban space using Artificail Neural Network

Introduction:
As one of the most valuable parts of a municipal system, the green space has long been praised. For that matter, all the data related to street-side trees must be assessed and recorded. A large part of old plane trees (Platanus orientalis) planted in industrial, dense populated cities are subjected to all kinds of air and water pollutions, frequent droughts and various physical stresses, which make them less likely to sustain. Identification of the most vulnerable tree individuals can be prioritized by a variety of statistical methods. A less applied statistical model in this field is the Artificial Neural Network (ANN). In this study, we present the results of applying ANN model in risk assessment of plane trees planted along Kuala Lumpur Avenue in Isfahan in 2018.
Material and methods:
In the current study, the risk level of keeping plane trees in Kuala Lumpur Ave. Isfahan was studied using data acquired by a full survey method, and using quantitative tree body proportions and few risk factors (qualitative or imperfect properties). Following coining the share of each of seven hazard criteria and their trio importance indices, a Kruskal-Wallis test compared the number of trees in different risk levels. Then, all the trees were scored via the biased levels of their risk level. Accordingly, based on the weighted scores, they were divided into five hazardous categories. To develop an understanding of the quantitative variables, risk factors, weight parameters, and hazard classes, we carried out a principle component analysis (PCA) and a multi-layer perceptron (MLP) network procedure.
Results and discussion:
The results of the proportion of each hazard index revealed the importance of the trunk and root wounds (83%), the structural tree weakness (61%), root problems (54%), and branch and twig dieback (50%). Also, results of Kruskal-Wallis test showed that the risk levels of the planted trees can be significantly classified into four classes of: with no risk or healthy, low, moderate, and high-risk classes, at one percent error level. The results of Duncan's mean test showed that the number of trees in no risk and low-risk classes were significantly higher than the other classes at one percent error level. The results of the PCA indicated that the first and second components explained 44.69 percent of the total variation. The risk and weighting parameters of the branch and twig dieback, the tree diameter, advanced decay, and wound on the trunk and root were highly and positively correlated. In general, the two variables of the trunk and root wounding, as well as branch and twig dieback, were among the most important variables in terms of risk assessment of the plane trees. The high coefficient of determination values of training, validation, verification, and finally, all neural network data (0.999, 0.949, 0.996, and 0.991) and the least mean square error values (training data = 0.052, verification 0.114, and validation = 0.044) indicated the accuracy and desirability of the ANN in the prediction of the risk classes for street side trees.
Conclusion:
According to the results of risk assessment diagnostic criteria and main components analysis, the two variables of trunk wound and root and branch mortality of plane trees should be regularly reviewed. Moreover, decision-makers may use the neural network method to identify and detect the risk possibility of planted trees. Therefore, this approach can be proposed as a suitable and useful solution in urban green space management and preventive measures.