Estimation of height eucalyptus tres with... ESTIMATION OF HEIGHT OF EUCALYPTUS TREES WITH NEUROEVOLUTION OF AUGMENTING TOPOLOGIES (NEAT)

1 Received on 01.06.2015 accepted for publication on 16.03.2017. 2 Universidade Federal de Viçosa, Doutorado em Ciência Florestal, Viçosa, Minas Gerais, Brasil. E-mail: <danielhbbinoti@gmail.com>. 3 Universidade Federal de Lavras, Programa de Pós-Graduação em Tecnologia da Madeira, Lavras, Minas Gerais, Brasil. Email: <pauloduarte.floresta@gmail.com>. 4 Universidade Federal do Espírito Santo, Departamento de Engenharia Florestal, Espírito Santos, Brasil. E-mail: <mayrabinoti@gmail.com> and<fernandes5012@gmail.com>. 5 Universidade Federal de Viçosa, Departamento de Engenharia Florestal, Viçosa, Minas Gerais, Brasil. E-mail: <hgleite@gmail.com>. 6 Universidade Federal do Espírito Santo, Departamento de Engenharia Florestal e da Madeira, Jerônimo, Espírito Santo, Brasil. E-mail: <ribeiroflorestal@yahoo.com.br>. 7 Universidade Federal de Lavras, Doutorado em Engenharia Florestal, Viçosa, Minas Gerais, Brasil. E-mail: <vclandradeuft@gmail.com>. 8 Universidade Federal de Viçosa, Programa de Pós-Graduação em Ciência Florestal, Viçosa, Minas Gerais, Brasil. E-mail: <aepifania16@gmail.com>. *Corresponding author.


INTRODUCTION
The estimation of the volume of trees in forest inventories provides information about the wood stock and the productive potential of the forests.This is important information for the planning of the sector.The most commonly used estimation method is the forest inventory, which is based on the allocation of representative forest stands and then estimation of the results for the total area (Husch et al., 2003;Burkhart and Tomé, 2012).
In forest inventories, height and diameter are dendrometric variables essential for estimating the volume of wood in trees (Campos, 1983).The relationship between these variables is referred to as hypsometric.This relationship is affected by environmental factors and characteristics of the population, such as age, genetic material, productive capacity and qualitative characteristics (Curtis, 1967;Zhang et al., 1997;Fang and Bailey, 1998).The inclusion of forest stands characteristics in hypsometric models such as those cited above has advantages including obtaining more accurate estimates and greater biological realism, making the equation adaptable to various locations (Campos and Leite, 2013).
According to Leite and Andrade (2004), height measurement is considered a significant component in the calculation of forest inventory.Obtaining it involves performing the measurement indirectly, and so the method is neither fast nor accurate, particularly when compared to dbh.In addition, problems such as wind conditions, and lack of visibility from the top of the tree in dense forests, may make measurement infeasible in forest plantations.Within this context, the common practice is to measure the height of some trees in the forest inventory plots, and estimate the heights of the others with hypsometric models.
ANN is a mathematical representation of the biological neural networks found in the human brain.In practice, they consist of a parallel computer system consisting of several simple processing elements (artificial neurons) linked together in a specific way to perform a given task.The ANN is presented with several examples of a possible problem, from which it extracts characteristics during the learning process (adjustment of the synaptic weights), and then applies the knowledge acquired to new data from similar problems (generalization) (Braga et al., 2007;Silva et al., 2010).
Despite the increasing development and application of ANN in the area of forestry (in both academic and business environments), one of the biggest challenges to ANN users is to define the topology and the appropriate architecture.Several factors including the number of layers, the number of neurons per layer, and type of activation function, make up the topology.The architecture refers to the direction of the connections between the neurons (feedforward or feedbackward).Generally, these parameters are defined empirically, testing different configurations until an acceptable level of error is reached (Braga et al., 2007;Silva et al., 2010;Binoti, 2012;Diamantopoulou, 2012).
The neuroevolution of augmenting topologies (NEAT) method was developed by Stanley (2004) to address the problem.NEAT allows for the simultaneous adjustment of the synaptic weights of an artificial neural network and the appropriate topology.This method uses genetic algorithms (GAs).NEAT initiates the topology with input and output layers, increasing the complexity of the topology with each GA generation.In a NEAT network, an intermediate neuron does not necessarily constitute an intermediate layer because it can connect to only a few neurons with unidirectional or bidirectional connections (Stanley and Miikkulaien, 2002;Stanley, 2004).
The general objective of this work is to evaluate the NEAT method for use in adjusting the weights and topology of artificial neural networks in the estimation of the height of clonal eucalyptus trees, as well as to comparing the values with estimates obtained by a hypsometric regression model.
Estimation of height eucalyptus tres with...

Data Description
Data from forest plantations located in the southern region of the state of Bahia were used.The total height of 19,560 eucalyptus trees, from a single source of genetic material, with an age range from 21.3 to 136.8 months, was considered for this analysis.For the estimation of the total height of the trees (Ht) we used quantitative variables, with the diameter value set to 1.3 m height (dbh) and dominant height (Hd) used for both RNA and regression model.The measured sample trees had Ht values between 7.0 and 39.0 m, dbh values between 5.1 and 28.3 cm, and the Hd of the plots ranged between 8.6 and 39.0 m.
The data obtained from the sample trees were divided into two groups: training (adjustment) and generalization (validation).In the selection of these sets, an equivalent proportion of data was considered in each diametric class, so that the sets were representative of the data.For the artificial neural networks, the training set consisted of the first three trees with the total height measurement of each plot, while the generalization set consisted of the other trees with total height measurement (10 to 15 per plot) in each forest inventory.To adjust the hypsometric regression models, the adjustment set consisted of all heights measured in the plot (10 or 15 trees with height measured per plot).As such validation was not performed, as this is the traditional process in use in most companies in the Brazilian forestry sector.

Neuroevolution of Augmenting Topologies (NEAT)
The NEAT method simultaneously adjusts synaptic weights and ANN topology using a GA, constituting an area known as neuroevolution (Stanley and Miikkulaien, 2002;Stanley, 2004).GA are heuristic techniques inspired by the biological process of natural evolution.Each ANN is represented by a set of parameters called the genome or chromosome.The genome is composed of genes that can take on different values called alleles.
The coding of neural networks in genome structure (chromosomes) consists of a list of genes referring to the artificial neurons and a list of genes referring to the connections between these neurons.Figure 1 shows the genotype (lists of neurons and connections) and the phenotype (network) of an ANN with two input neurons, two hidden neurons, and one output neuron.The list of neurons identifies the neuron number and type (input, hidden, or output).The list of connections identifies the connection sense, the synaptic weight value, whether the connection is enabled or not (whether or not it is expressed in the phenotype), whether it is recurring or not, and the number of the innovation.
Initially the NEAT creates a population of single layer networks, but with different synaptic weights.Through successive generations of GA, genetic operators increase the structure of networks by adding neurons and/or connections, as well as adjusting the synaptic weights until a network with satisfactory results is obtained (Figure 2).

Hypsometric Model
For comparison purposes, a hypsometric model, commonly used on an operational scale by forestry companies in Brazil, was adjusted.
Where: ln = Napierian logarithm;  i = regression parameters; dbh = diameter at 1.3 m height; Hd = dominant height of plot; and

Application of the NEAT method
The computational system used to apply the NEAT was the NeuroForest 3.3 (Binoti, 2012), available free of charge from their homepage (http:// neuroforest.ucoz.com/).The system allows the user to configure the population size and density parameters of the initial connections.
In this work population sizes of 100, 200, 300, and 400 individuals (number of nets tested in each iteration), and initial connection densities of 0.1, 0.5, and 1.0 (initial connection rate of each iteration, for example, 0.1 equals 10%) were used.Five ANN were trained for each of the combinations of the NEAT parameters, totaling 60 ANN.
The NEAT was set to stop training when the number of times or cycles (number of times all the training set data are applied to the RNA for parameter adjustment) reached 3,000 or when the mean squared error was equal to 0.0001, whichever happened first.
The processing time of the NEAT method for the combinations of the initial parameters was quantified in a computer with an AMD FX™-8350 Eight-Core, 4.00 GHz processor, 32.0 GB of RAM, and a 64-bit operating system.Estimation of height eucalyptus tres with...

Evaluation of methods
Estimates of the total height of the trees obtained with RNA and the regression model were evaluated based on the Pearson linear correlation coefficient between the observed and estimated values in the percentage error dispersion graph, and the percentage frequency histogram of the percentage errors and root mean square error (RMSE %).
The root mean square error (RMSE%) evaluates the mean square difference between the observed values and the estimated values.The lower the RMSE, the better the accuracy of the estimate being: where, (Ht) ´ is the mean of observed heights (field measurements),?Ht?_i is the observed height of the i-th tree, (Ht) ^_i is the estimated height of the i-th tree and n is the total number of data.
To evaluate the average performance of the ANN for each configuration option of the NEAT parameters, the correlation and the mean RMSE% of the five ANNs adjusted for each combination were evaluated.

RESULTS
Among the various combinations of the parameters of the NEAT method tested, the 12 ANN that presented the best results in terms of linear correlation and RMSE% are presented in Table 1, along with the corresponding statistics for the hypsometric model.ANN were found that presented higher correlation and lower values of RMSE% in comparison to the hypsometric model in both training and generalization.
The best result was obtained with ANN 49, which completed its training after 3,000 cycles, with population size equal to 400 and initial density of connections equal to 0.1 (Table 1).The percentage errors of the estimates by ANN 49, 95% in training were between ± 10%, and 98% in this same interval in generalization (Figure 3).On average, a population size value of 300 and an initial density of connections value of 0.1 obtained the highest correlation and the lowest RMSE% (Table 2).
Table 2 indicates that the processing time spent for RNA generation was not high, even with a relatively large database, maintaining the accuracy of the traditional method commonly used and reducing the number of trees collected in the field.

DISCUSSION
There was a small improvement in the analyzed statistics with increased population size, but at the cost of a significant increase in the processing time of each ANN.According to Heaton (2011), this increase in computational cost during network training with increasing population size may lead to non-convergence of the algorithm or hardware locking.
It was not possible to perceive any influence of the initial density of the connections in the quality of the estimates, as good results were obtained in all values tested.
During the process of training the ANNs for operational application, the measurement team faced the problem of defining the structure and the ideal topology for training.Achieving consistency and avoiding bias in the estimates was a concern.
Because of the dynamics of the forest company, where innumerable hypsometric estimates must be obtained with temporal and computational constraints, tests on the best structure and/or topology of the ANNs used could not be implemented as they would have resulted in a low operational measurement performance.Algorithms that automatically define the structure and topology of the ANNs are thus extremely important to facilitate their application.This applies not only in the forest measurement sector, but also in more diverse sectors of forestry engineering where the aim is to reduce operational and processing costs.
Measuring tree heights in the forest inventory is carried out indirectly.Several factors can introduce errors in the measurement process of this variable.These include wind, tree positioning, tree visibility, luminosity, and equipment accuracy (Hush et al., 2003).Although a simple change of methodology for the estimation of tree height will not provide significant gains for the forest measurement team, associated reduction in the number of trees measured in the field does provide the advantage of reducing both costs and errors associated with the difficulty of measuring this variable (Binoti, 2012   Histograma de frequência percentual dos erros percentuais no treinamento e generalização, respectivamente.

Figure 3 -
Figure 3 -Dispersal of percentage errors on the amounts of the observed heights (m), training and generalization.Percentage frequency histogram of percentage errors in training and generalization, respectively.Figura 3 -Dispersão dos erros percentuais em função dos valores das alturas observadas (m), no treinamento e generalização.Histograma de frequência percentual dos erros percentuais no treinamento e generalização, respectivamente.

Table 1 -
). Characteristics of the twelve best artificial neural networks by NEAT adjusted for each combination of parameters and their statistics for the training and generalization.Tabela 1 -Características das doze melhores redes neurais artificiais ajustadas pelo NEAT para cada combinação de parâmetros e suas respectivas estatísticas para o treinamento e generalização.