Modeling the Risk Possibility of Trees Falling with Artificial Neural Network and Logistic Regression Application in Urban Green Space

Document Type : Research Paper


1 Department of Natural Engineering, Faculty of Rangeland and Watershed Management, University of Shahrekord, Shahrekord, Iran

2 Department of Forestry, Faculty of Natural Resources and Earth Sciences, University of Shahrekord, Shahrekord, Iran


Urban green space plays an important role in city functioning and its lacking can lead to city life disruption. Urban green space trees include a wide range of environmental, aesthetic, social, physiological and economic benefits. Tree falling due to various factors, especially storms and high-speed winds, can cause severe economic damage, even endangers public health and safety, which is very important due to their widespread presence, especially in the urban areas. It is attracted by municipal green space and urban crisis management to identify and quantify the severity of the hazardous trees and only can provide risk management and suitable preventive measures.
Hazardous trees referres to absolutely dried or drying trees, dead components or highly unstable living trees that may be come from structural damage or other factors. These trees pose to a high risk of threatening citizen lives or their property in the urban environment. Normally, healthy trees are more resistant to applied forces and damaged while defective trees are more prone to fracture and eradication in the crown, trunk and root. Hazardous trees management inspects tree risks in the natural and synthetic green space by modeling the behavior and measuring tree conditions in the urban green space.
There are several ways to correlate the quantitative and qualitative variables (defects) of trees with their probability of tree hazardous, including artificial neural networks and multivariate regression that have less constraints and assumptions than statistical methods for modeling process. The purpose of logistic regression is describing the relationships between quantitative and qualitative variables as a set of independent variables and risk intensity as a dependent variable. This model maximizes the probability of an event will occur and does not necessarily have a linear relationship between the independent and dependent variables. The neural network model is able to detect the relationship between a set of inputs and output data to predict the output corresponding to arbitrary inputs regardless of the parameters.
Recently, studies have been performed on the risk possibility of the trees. Elm tree (Ulmuscarpinifolia var. Umbelifera) belongs to the family Ulmaceae. This genus has various species that are distributed in most parts of the world, especially in Asia, Europe and North America.These trees have relatively dense distribution in the natural forests of Iran, especially the northern and southern slopes of Alborz.They are also used in the most urban green spaces due to their beautiful appearance and wide shade as an ornamental and shaded tree. According to available information, in our country, despite the great extent of the urban green space and its importance and the need to be aware of street tree dangers that pose to a high risk to the life-threatening safety of citizens or their property, on the other hand, there has been limited applied research to predict the risk possibility of the fallen trees. Therefore, this study aimed to determine the most important independent quantitative and qualitative variables affecting the probability of falling trees and developing multilayer perceptron neural network models and logistic regression to predict the probability of falling of elm hazardous trees and also to compare these two models in predicting the probability of falling in a tree in Shariati has done Shahrekord.

Material and methods
The study area is located in Chaharmahal anf Bakhtiari province, Shahrekord, between 49 degrees and 22 minutes to 50 degrees and 49 minutes and latitude 32 degrees and 20 minutes to 33 degrees and 31 minutes, respectively. The study area was a part of Shariati Boulevard between Basij Square and the intersection of North Ferdowsi Street with a length of 584.13 meters. In order to compare the ability of neural network model with logistic regression to predict the risk possibility of elm trees, the study was performed on 129 elm trees in Shariati St. Boulevard. For this purpose, the variables including tree diameter, tree height, dry branches and woods, cracks, structural and physical weakness (vertical deviation), root and wound and root problems as independent. Quantitative and qualitative variables were classified and the risk classification of elm trees were measured as dependent variables. A complete survey method used to measure and record the quantitative and qualitative variables of elm trees on both sides of the street.
Firstly, Normalization of the data was performed and then Principal Component Analysis was used to select the main variables.
If KMO (Kaiser-Meyer-Olkin) index is less than 0.5, the data are not suitable for principal factor analysis and if the value is between 0.5 and 0.69, it should be used with higher caution. Also this method used for binary logistic regression models that is a step-by-step method were included in the model. By choosing this method it is possible to isolate significant independent variables related to the probability of elm falling trees (dependent variable). In this study, 70%, 15% and 15% of the total data were allocated to multi-layer perceptron network training, validation and testing, respectively. Qualitative and quantitative variables are considered as input variables and dependent variable of risk intensity class in each tree as output. In order to predict neural network model for prediction of the risk of hazardous elm tree two hazard severity classes were used based on weighting method. According to general approximation theorem, a neural network with a hidden layer and a sufficient number of neurons per layer can approximate any arbitrary continuous function. This study aimed to train neural network from Multilayer Perceptron network of artificial neural networks. Neural network training was used with learning rate reduction 10 (BDLRF). The activity functions in the hidden layers for all networks were considered as sigmoid function and in the output layer linear transfer function. Accordingly, multilayer perceptron network with 5 neurons in input layer, one hidden layer with 20 neurons and one neuron in output layer were used to modeling the risk possibility of elm trees. Surface area under ROC curve and Kappa curve were used to evaluate model prediction accuracy.

Results and discussion
Results showed that three criteria for branches, root and trunk problems were 87%, 79% and 77%, respectively, and criteria for contact with power lines had the least rate (29%) in the risk of elm trees. Regarding the analysis of principal components tree height, tree diameter, dried branches, structural status or physical weakness (vertical deviation), root problems and advanced decay are the most important variables affecting logistic regression and neural network models. Logistic regression results indicated that with 65.9% confidence iccluding 5 independent variables of tree height, tree diameter, branch and dried branches, structural weakness or physical weakness, advanced decay are predictor variables in probability of tree falling in the present study.The results of the NIMBUS tests to exam the model's explanatory power and efficiency show that the fit of the model is acceptable at the error level of less than 0.01. Independent variables used have moderate explanatory power regarding variance and variations dependent on the probability of falling trees. In fact, these variables were able to explain between 0.33 to 0.456% of tree fall variations.

Diagnostic criteria for hazardous trees showed that three criteria of dried branches, root problems and trunk cracks had the highest rate in the risk possibility of elm trees. Using the principal component analysis reduced the number of input variables to the models. It also eliminated the correlation between the input variables to the model and easier interpretation of the models. According to this analysis, the variables of tree height, tree diameter, twigs and branches, structural or physical weakness (vertical deviation), root problems and advanced rot were identified as the most important variables. Therefore, two criteria for woods and shoots and root problems in both analyzes are selected the most important variables.