CLASSIFICATION OF THE INITIAL DEVELOPMENT OF EUCALIPTUS USING DATA MINING TECHNIQUES

CLASSIFICAÇÃO DO DESENVOLVIMENTO INICIAL DO EUCALIPTO UTILIZANDO TÉCNICAS DE MINERAÇÃO DE DADOS

ABSTRACT

Eucalyptus plantation has expanded considerably in Brazil, especially in regions where soils have low fertility, such as in Brazilian Cerrados. To achieve greater productivity, it is essential to know the needs of the soil and the right moment to correct it. Mathematical and computational models have been used as a promising alternative to help in this decision-making process. The aim of this study was to model the influence of climate and physico-chemical attributes in the development of Eucalyptus urograndis in Entisol quartzipsamment soil using the decision tree induction technique. To do so, we used 30 attributes, 29 of them are predictive and one is the target-attribute or response variable regarding the height of the eucalyptus. We defined four approaches to select these features: no selection, Correlation-based Feature Selection (CFS), Chi-square test (χ²) and Wrapper. To classify the data, we used the decision tree induction technique available in the Weka software 3.6. This data mining technique allowed us to create a classification model for the initial development of eucalyptus. From this model, one can predict new cases in different production classes, in which the individual wood volume (IWV) and the diameter at breast height (DBH) are crucial features to predict the growth of Eucalyptus urograndis, in addition to the presence of chemical soil components such as: magnesium (Mg⁺²), phosphorus (P), aluminum (Al⁺³), potassium (K⁺), potential acidity (H + Al), hydrogen potential (pH), and physical attributes such as soil resistance to penetration and related to climate, such as minimum temperature.

Keywords:
Eucalyptus urograndis; Individual wood volume; Feature selection; Entisol quartzipsamment soil; Decision tree

RESUMO

O cultivo de eucalipto tem se expandido consideravelmente no Brasil, sobretudo em regiões em que os solos apresentam baixa fertilidade, como nos Cerrados brasileiros. Para alcançar maiores produtividades, é fundamental saber a necessidade e o momento adequado para correção do solo. Para auxiliar esse processo de tomada de decisão, modelos matemáticos e computacionais têm sido utilizados e são uma alternativa promissora. O objetivo deste trabalho foi modelar a influência dos atributos físico-químicos do solo e climáticos no desenvolvimento do Eucalyptus urograndis em Neossolo Quartzarênico, por meio da técnica de indução da árvore de decisão. Para isso foram utilizados 30 atributos, sendo 29 preditivos e um atributo-meta ou variável resposta, a saber, a altura do eucalipto. Foram avaliadas quatro abordagens para seleção de atributos: sem seleção, seleção de atributos baseado em correlação (CFS), método do Qui-quadrado (χ²) e Wrapper. Para classificar os dados foi utilizada a técnica de indução de árvore de decisão por meio do software Weka 3.6. As técnicas de mineração de dados através da indução de árvore de decisão permitem o desenvolvimento de um modelo de classificação do desenvolvimento inicial de eucalipto eficiente para previsão de novos casos em diferentes classes de produção, onde o volume individual de madeira (VOL) e o diâmetro altura do peito (DAP) são atributos determinantes para previsão do crescimento do Eucalyptus urograndis, além de atributos químicos do solo como: Magnésio (Mg⁺²), Fósforo (P), Alumínio(Al⁺³), potássio (K⁺), acidez potencial (H⁺Al) e potencial hidrogeniônico (pH), atributo físico como a resistência do solo à penetração e relacionado ao clima como a temperatura mínima.

Palavras chave:
Eucalyptus urograndis; Volume individual de madeira Seleção de atributos; Neossolo Quartzarênico; Árvore de decisão

INTRODUCTION

The plantation of Eucalyptus spp in Brazil has increased in recent years because of the following reasons: quick growth, diversification in the use of wood and the ease of adapting to different soil and climate conditions. This fact makes commercial plantations of eucalyptus in Brazil quite variable and productive, which contributes to the recognition of the country as the one with the best technologies in eucalyptus plantation currently, reaching around 60 m³.ha^-1 average productivity in rotations of seven years (SFB, 2016).

Increases in the production of wood led the Brazilian sector of pulp and paper to stand out in the world’s forest area as the fourth biggest productor of pulp, the ninth biggest productor of paper, in addition to being the 13th biggest in the consumption of paper per capita, totaling 220 companies of pulp and paper with forest activities in 540 cities located in 18 states (TOLEDO et al., 2015).

However, eucalyptus trees have been planted in regions of Cerrado, whose soils are predominantly highly weathered and acid, with low nutrient availability, low aggregation, exposed to prolonged dry seasons, which can lead to decreased productivity of wood (BARBOSA et al., 2012). To avoid decreases in productivity, it is necessary to use monitoring and evaluation methods to determine which are the variables that most affect the development and productivity of Eucalyptus.

The use of computational systems in the decision making of commercial plantations of eucalyptus is an interesting strategy. The reason is that those systems can make the sector increase even more, since it allows the evaluation of large amounts of information on soils and plants, which can lead to new findings and most appropriate strategies to increase productivity and environmental protection.

Among the computational techniques used to interfere in the development of eucalyptus, due to heterogeneity found in the soil, data mining techniques aim to find patterns in large amount of data to infer a dependent variable from a set of features associated with this variable (CRIVELENTI et al., 2009).

Among the data mining techniques, decision tree induction is a method very simple and efficient, because it allows the classification of data sets consisted of numerical and categorical variables. In addition, the knowledge found in a decision tree is represented by rules and the algorithm achieves satisfactory results when compared with other more sophisticated approaches available in the literature (HAN et al., 2011). That is why decision tree induction is seen as a promising approach in data analysis (SOUZA et al., 2010).

A decision tree is a flowchart-like structure that shows the various outcomes from a series of decisions. It is formed of nodes connected by branches and leaves. Nodes represent the variables in the data file analyzed, while the branches between the nodes represent logical tests conducted for the separation of data. The first node is called root node and it is the main one in the decision tree. Nodes located below the main node are connected by branches. Leaves are the regions associated with a label or value of the terminal node (WITTEN et al., 2011). Decision tree operation occurs with the division of a set into subsets of data in a recursive way. The separation of data occurs until each subset is homogeneous, i.e, with cases of one single class (WITTEN et al., 2011).

The main advantages of using a decision tree include the support in a decision-making process, which considers the most relevant attributes, and the facility of interpretation and understanding of the results, because the classification is obtained explicitly, simplifying its interpretation and allowing users to know which attributes influence the development of eucalyptus. In addition, the results are usually supplied quickly due to the computational efficiency provided by this technique (DAI et al., 2016; HAN et al., 2011).

In this context, this work aimed at modeling the influence of physico-chemical attributes of the soil and climate ones in the development of Eucalyptus urograndis in Entisol quartzipsamment soil by using the decision tree technique.

MATERIAL AND METHODS

Area of study

The study was carried out in the 2014/2015 agricultural year, in the experimental area of Farm Bom Retiro, owned by Eldorado Brazil Celulose, located in Três Lagoas, state of Mato Grosso do Sul, Brazil, at latitude 20^o27’S and longitude 52^o29’W, with an annual average rainfall of 1300 mm and average temperature of 23.7°C. The climate type is A_w, according to the Köppen classification, being characterized as tropical humid with a rainy weather during summer and dry one during winter. The experimental area is under cultivation of Eucalyptus urograndis. The soil was classified as Entisol quatzipsamment soil (EMBRAPA, 2013).

Collection and determination of dendrometric and soil attributes

In the experimental area, a sample mesh was installed containing 300 random points, representative of an total area approximately 12 ha. In order to determine the dendrometric attributes, 300 trees of E. urograndis were individually sampled and soil attributes were determined in the vicinity of each tree. Thus, we had a database with 300 observations for each plant and soil attribute.

The dendrometric attributes evaluated were: individual height of eucalyptus trees (IHE), diameter at breast height (DBH), and determined the volume (IWV). The cubing occurred with the tree standing. The collection of data on tree heights was performed with a clinometer and DBH was collected at 1.300 m hight from the ground with the aid of a digital caliper.

To determine the individual volume of each tree (IWV), the cubing occurred with the tree standing using we used the Huber´s formula because it assumes that the average area of a sectioned tree is at its midpoint; but this is not always that the case, it has an intermediate precision (CAMPOS; LEITE, 2002). In order to have this correction in the individual wood volume, it used the form factor equal to 0.4 (OLIVEIRA et al., 2009). The Huber’s formula, adapted and described by Péllico Netto (2004), is obtained by the product of the sectional area (taken by half of the section) and sectional length determined by equation 1. Which: IWV is the individual wood volume (m³); DHB is diameter at breast height (m); and IHE is the height of the tree (m).

$$IWV=\left[DB{H}_2\stackrel{.}{}\left(3.14/4\right)\stackrel{.}{}IHE\right]\stackrel{.}{}0.4$$ (1)

Soil attributes were collected around each evaluated tree, totaling 300 sampling points. The following soil attributes were assessment: mechanical resistance to penetration (RP), gravimetric moisture (GM), volumetric moisture (VM), bulk density (BD), particle density (PD), total porosity (TP), sand, silt, clay, phosphorus (P), organic matter (OM), hydrogen potential (pH), potassium (K⁺), calcium (Ca²⁺), magnesium (Mg²⁺), potential acidity (H⁺+Al³⁺), aluminium (Al³⁺), sum of bases (SB), cationic exchange capacity (CEC), saturation by bases (SB), calcium with cationic exchange capacity (Ca/CEC), magnesium with cationic exchange capacity (Mg/CEC), and aluminum saturation (m) collected in layers of 0-0.20 m and 0.20-0.40 m. These were determined according to the methodology proposed by Embrapa (2011), Stolf (1991), and Raij et al. (2001). All analyses were performed in the laboratory of physics and fertility of the Faculty of Engineering of Ilha Solteira, Sao Paulo.

Data mining

The original data set was composed by 30 attributes (29 predictive attributes and one target-attribute or response variable) (Table 1), which were added to the set of data, totaling 300 observations for each attribute. The target attribute refers to the height of Eucalyptus urograndis and is the target of the classification.

TABLE 1
Description of predictive attributes (physical, chemical, climatic, and dendrometric) used in the database to predict the height of Eucalyptus urograndis using decision trees

Attributes		Description	Unity
Physical	PR	Penetration resistance	MPa
	GM	Gravimetric moisture	kg.kg-1
	VM	Volumetric moisture	m3.m-3
	BD	Bulk density	Mg.dm-3
	PD	Particle density	m3.m-3
	TP	Total porosity	m3.m-3
	San	Sand	g.kg-1
	Sil	Silt	g.kg-1
	Cla	Clay	g.kg-1
Chemicals	P	Exchangeable phosphorus	mg.dm-3
	OM	Organic matter	mg dm-3
	pH	pH in calcium chloride	-
	K+	Exchangeable potassium	mmolc.dm-3
	Ca2+	Exchangeable calcium	mmolc.dm-3
	Mg2+	Exchangeable magnesium	mmolc.dm-3
	H+ + Al3+	Potential acidity	mmolc.dm-3
	Al3+	Exchangeable aluminum	mmolc.dm-3
	SB	Sum of bases	mmolc.dm-3
	CEC	Cation exchange capacity	mmolc.dm-3
	V	Base saturation	%
	Ca/CEC	Relation Ca/CEC	mmolc.dm-3
	Mg/CEC	Relation Mg/CEC	mmolc.dm-3
	m	Aluminium saturation	%
Dendrometric	DBH	Diameter at breast height	cm
	IWV	Individual wood volume	m3
>Climatic	Prec	Precipitation	mm.dia-1
	T	Minimum/maximum/average temperature	°C

Different heights of E. urograndis were submitted to a discretization procedure in categories, i.e, continuous data were transformed into discrete data (intervals), since the categorization of the response variable in intervals makes the information simpler, facilitating the interpretation and the decision-making when analyzing a decision tree. To do so, height values of E. urograndis were organized in ascendent order and divided equally into three classes (low, medium and high) (Table 2).

TABLE 2
Distribution of Eucalyptus urograndis heights according to low, medium and high classes and their respective limits, aimed at inducing decision trees to predict different heights of Eucalyptus.

Classes	Interval	Frequency
Low	[1.4 - 3.8]	99
Medium	[3.9 - 9.9]	101
High	[10.1 - 12.6]	100

Selection of attributes

To select attributes, the most common methods are those based on the theory of information, such as information gain, which represents the expected reduction in entropy caused by partitioning the examples according to an attribute (HAN et al., 2011).

Entropy measures the amount of information brought by an attribute, thus characterizing the uncertainty or randomness of a set of examples (Shannon, 1949). Given a set of examples concerning the target attribute (S) of interest, and a categorization (C) of the same attribute in n classes S1, S2, ..., Sn, the entropy H (S) is defined by the equation 2, as follows, where pi is the ratio of favorable cases in the class Si.

$$H\left(S\right)={\displaystyle \sum }_{i=1}^n-\left(pi\right).\log \left(pi\right)$$ (2)

Because of the large number of attributes generated in data preprocessing, a selection procedure was applied to attributes with the purpose of removing those highly correlated, considering they can bring redundancy to the generated model. To do so, we evaluated four approaches to select the attributes:

1-Without attribute selection, in which all attributes were used. We characterized this approach by the absence of selection;

2-Attribute selection based on correlation (CFS) searches for a set of correlated attributes to prevent the use of the same information already used previously;

3-The Chi-square test (χ²) is based on the concept of statistical independence. To accomplish that, attributes are evaluated individually by using the χ² measure regarding the class of interest held by equation 3.

$\text{χ}2={{\displaystyle \sum }}^{\text{}}{\left(Observed-Expected\right)}^2\cdot Expecte{d}^{-1}$ (3)

The observed frequencies are directly obtained in the sampling data, and the expected frequencies are calculated based on the observed ones.

The Wrapper approach is applied together with a basic learning algorithm. This method generates a subset of attributes, which is tested by the learning algorithm of interest. Such a process is repeated for each subset of attributes until the stop criterion is satisfied (WITTEN et al., 2011).

Induction and validation of the classification model (decision tree)

Induced models with a variation in the number of instances (or observations), per leaf, were assessed using the cross-validation method using 10 folds. The selection of the best model was made based on the following measures: (i) accuracy; (ii) number of leaves (number of rules), generally associated with the ease of interpretation of the model; (iii) Kappa coefficient, which is a measure of agreement between predicted and observed classes of the classifier.

As a result of the induction of the decision tree model, a data analyst obtains the confusion matrix (Table 3), widely used in statistical analysis of agreement (HAN et al., 2011).

From the confusion matrix, it is possible to get the measures of performance evaluation. Accuracy is the percentage of examples that were correctly classified by the classifier and can be expressed as in Equation 4.

TABLE 3
Example of a confusion matrix of 2 × 2 size.

Predicted Observed	Class A	Class B	Total
Class A	TP	FN	P
Class B	FP	TN	N
Total	P’	N’	P + N

$Accuracy=\left(TP+TN\right)\cdot {(P+N)}^{-1}$ (4)

The Kappa coefficient is used to describe the measure of agreement between the predicted and observed classes. Such a coefficient ranges from 0 to 1, representing poor and excellent ranking results, respectively. It can be defined by Equation 5, as follows (WITTEN et al., 2011), where Pr (a) is a relative agreement observed for a given class in the confusion matrix; Pr (e) is the probability of the expected agreement in this same class.

$K=\mathrm{Pr}\left(a\right)-\text{Pr}\left(e\right)\cdot {(1-\text{Pr}\left(e\right))}^{-1}$ (5)

The Kappa coefficient is calculated considering all the classes available in a data set. A possible interpretation of models performance from the Kappa statistic method was introduced by Landis and Koch (1977) and can be seen in Table 4.

TABLE 4
Performance of classification models using the Kappa statistics method.

Kappa statistics	Quality
< 0,00	Very bad
0.00-0.20	Bad
0.21-0.40	Average
0.41-0.60	Good
0.61-0.80	Very good
0.81-1.00	Excellent

For data classification, we used the method of binary decision tree available in the Weka software 3.6 (WITTEN et al., 2011). The induction algorithm used was the J48, widely known as C4.5 and developed by Quinlan (1993). With the aim of minimizing a possible overfitting effect, tree pruning techniques were used to reduce the number of internal nodes, generating smaller and less complex trees and, therefore, easier to be understood.

RESULTS AND DISCUSSION

Results of the use of different methods for attribute selection can be found in Table 5. In this table, we present some metrics such as accuracy, Kappa coefficient, classes precision, and the number of rules generated in each tree for each method of attribute selection.

TABLE 5
Accuracy,Kappa coefficient, and the number of rules for different methods of attribute selection for the induction of decision trees to predict the height of Eucalyptus urograndis.

Met. Selection	Accuracy	Kappa	Precision by class			Number Rules
			Low	Medium	High
Without selection	84.00	0.78	0.89	0.75	0.87	18
CFS	84.67	0.77	0.90	0.75	0.91	14
χ2	83.66	0.75	0.88	0.74	0.90	21
Wrapper	84.67	0.77	0.87	0.77	0.90	12

In general, the performance achieved by different methods of attribute selection was very similar. IIn particular, the accuracy ranged from 83.66% to 84.67%, while Kappa ranged from 0.75 to 0.77. So the classification of all the methods is considered very good, as can be seen in Table 5. In addition, it is possible to notice a wide variation in the number of selected attributes and the number of generated rules, ranging from 12 to 21.

Among the methods for attribute evaluation, that one without attribute selection showed better efficiency of the model, with and accuracy of 84% and Kappa of 0.78, indicating a very good classification. This method used all attributes in the set of data and generated a decision tree with 18 rules. In addition, it is possible to notice that this method showed high accuracy for all classes evaluated: 89% for the low class, 75% for the medium class, and 87% for the high class, indicating the model efficiency in classifying all the classes.

Despite presenting a good accuracy (84.67%), the CFS classifier generated a decision tree with only 14 rules and used only four attributes, namely: Diameter at breast height (DBH), individual wood volume (IWV) and minimum temperature (T min). In the decision trees interpretation, there is a preference for smaller trees, because they are more comprehensive, present less redundancy of attributes and are susceptible to have greater accuracy because of optimization by pruning; however, in this case, the generated tree does not meet the identification of the standards previously mentioned in the database, since it is quite general (only adding little information), since it uses only four attributes out of 29 predictive ones.

From the 29 predictive attributes in the dataset, the Wrapper method selected just five of them: diameter at breast height (DBH), individual wood volume (IWV), organic matter (OM), Potential acidity (H+Al) and sum of bases (SB) at a depth from 0.20 to 0.40 m, and potential acidity at a depth from 0.00 to 0.10 m with an accuracy of 84.67% and Kappa of 0.77, indicating that the use of these five attributes are enough to achieve a very good classification. However, this selection method also generated a decision tree with a number of rules and attributes excessively low.

The method based on the Chi-square test excluded 16 attributes of the data set, thus only remaining the following ones: BD, PD, SAND, Clay, Silt, pH, Ca, H=Al, SB, CEC, V and Mg/CEC. However, this method presented a lower accuracy among all other evaluated methods for attribute selection (83.66%). In parallel, this method also presented one of the smallest Kappa coefficient (0.75), totalling 21 generated rules.

In Table 6, we present data from the confusion matrix generated through the execution of the J48 algorithm for decision tree induction without selection of attributes, which gives details on both good and bad matches of the model. The results revealed that the model achieved a high accuracy, as can be seen in the main diagonal in which 252 cases were correctly classified and distributed in low, medium and high classes, confirming an accuracy of 84% and Kappa coefficient of 0.78 (Table 6). The values available out of the main diagonal represent the model error, totalling 48 cases, corresponding to an error rate of 16%.

TABLE 6
Confusion matrix obtained by the execution of the J48 algorithm for the decision tree induction, without attribute selection.

Predicted Observed	Low	Medium	High
Low	86	13	0
Medium	11	79	11
High	0	13	87

The decision tree generated by the execution of the J48 algorithm is available in Figure 1, without attribute selection.This decision tree is composed of 18 rules, five of them belonging to the low class, nine belonging to the medium class and four belonging to the high class, corresponding to 27%, 50% and 22% of the total of generated rules, respectively (Table 7).

TABLE 7
Number of rules generated by using the J48 algorithm, without attribute selection to predict the growth of Eucalyptus urograndis.

Classes	Number of rules
Low	5
Medium	9
High	4
Total	18

Analyzing the decision tree illustrated in Figure 1, one can infer that the individual wood volume (IWV) attribute showed the highest importance because of its higher information gain. So it is placed in the root of the tree. According to Quinlan (1993), the root node is the attribute with lower entropy and, consequently, it has the highest gain of information, optimizing the generation process of decision trees.

FIGURE 1
Resultant decision tree from the execution of the J48 algorithm without attribute selection to predict the growth of Eucalyptus urograndis.

The organization of a decision tree follows a hierarchy, so that the highest hierarchy attributes are also the highest in information gain in the decision-making process. Hence, variables with greater proximity to the root node of the tree have higher relevance in eucalyptus growth classification.

It is possible to see in Figure 1 that the second most important attribute was the diameter at breast height (DBH). Lima et al. (2010), studying the physico-chemical attributes of a oxysol from Cerrado and its association with the dendrometric characteristics of Eucalyptus camaldulensis, found a high positive correlation between individual wood volume (IWV) and the diameter at breast height (DBH). Carvalho et al. (2012) also styding Eucalyptus camaldulensis found a significant correlation between individual wood volume (IWV) and the diameter at breast height (DBH) (r=0.95**), and between individual wood volume (IWV) and the height of eucalyptus (ALT) (r=0.85**). Rosa Filho et al. (2011), evaluating the spatial variability of dendrometric properties of eucalyptus and physical attributes of a Red Oxysol with Eucalyptus urophylla, verified a correlation between the height of eucalyptus and the diameter at breast height (DBH) (r = 0.677**). In addition, these results are in agreement with Pinkard and Neilsen (2003), who found greater volume per tree due to the higher DBH, in a settlement with 500 trees of Eucalyptus nitens per hectare, compared with densities up to 1,667 trees per hectare. By analyzing the decision tree depicted in Figure 1, we noted two main branches in the tree from the root node, one on the left side and the other on the right. Right branches are associated to individual wood volume (IWV) higher than 0.001796m³, in which there is the predominance of medium and high classes. On the other hand, in the left branch there is the predominance of medium and low classes related to the individual wood volume (IWV) equal tor lower than 0.001796 m³. Thus, we observed sampling data of higher individual wood volume (> 0.001796 m³) in the highest levels of production of Eucalyptus urograndis. The increase in the individual wood volume is conditionated to mineral fertilization of the soil, which is a commonly used tool to modify the quality of forest area and, thus, increase the growth rate of trees. According to Barbosa et al. (2014), the fertilization techniques used for the initial development of eucalyptus trees causes an increase in the volume of wood, since this is highly sensitive to environmental changes.

Moreover, it can be observed that for the determination of the high class, the most decisive attributes were those associated with soil chemistry, such as: Magnesium (Mg⁺²), phosphorus (P) and aluminum (Al⁺³) in the layer from 0.20 to 0.40 m and potassium (K⁺) in the layer from 0.00 to 0.10 m, which have been present throughout the right branch. The relationship of these chemical attributes was due to the fact that they are indispensable elements for the development of eucalyptus. Magnesium plays an essential role in chlorophyll constitution, which is very important in photosynthesis. Phosphorus is the most limiting nutrient in the initial growth of eucalyptus in Brazil because of its participation in metabolic processes of roots development and in reproductive initiations (RAIJ, 2011). Potassium is one of the nutrients more demanded by the eucalyptus because it is directly associated with the water balance of trees; potassium deficiency is one of the most frequent in forest plantations of eucalyptus, which directly influences foreign exchange activities and wood formation (FREITAS et al., 2015). Aluminum toxicity is one of the factors that most restrict development and forestry production in acid soils, especially in pH less than 5.0, because it interferes in the fixation of phosphorus, decreases root respiration and interferes in absorption, transportation and in the use of nutrients and water by plants (COSTA et al., 2015).

The results also showed that the only physical attribute capable of influencing the high production of eucalyptus was soil resistance to penetration in the layer of 0.20 to 0.40 m. Lima et al. (2010) observed simple linear correlations between individual wood volume of Eucalyptus camaldulensis and soil attributes when paired with significant mechanical resistance to penetration and pH. On the other hand, Rosa Filho et al. (2010) observed simple linear correlations between individual wood volume of Eucalyptus camaldulensis and soil attributes, which showed to be significant when both were paired with resistance to penetration only in depth from 0.2 to 0.3m.

Apart from that, we also observed that of all climatic attributes evaluated only the minimum temperature was present in the decision tree, and only temperatures above 18.25 provided greater growth of eucalyptus. Studies associated with a decrease in the temperature below the optimum temperature tend to reduce plant growth by limiting the production of wood and damaging the formation of homogeneous plantations (FLORIANI et al., 2013; GAVITO et al. 2001; PENG; DANG 2003). In particular, young eucalyptus plants are more sensitive to low temperatures than adult plants, which can lead them to death because they are not completely established in the environment where they were planted (MORAES et al., 2015).

Attributes associated with the soil acidity, such as potential acidity (H + Al) in a layer from 0.20 to 0.40 m, pH in layers from 0.00 to 0.20 m and 0.20 to 0.40 m were decisive in the generation of rules related to the low class, and these attributes are present in 60% of rules of the low class of eucalyptus growth. Lima et al. (2010) observed that the pH in surface layer of the soil was the attribute that showed the best interaction with the productivity of individual wood volume (IWV) of Eucalyptus camaldulensis to conditions in Selvíria (Mato Grosso do Sul, Brazil) and therefore it can be thought of as the best indicator of the quality of the soil studied, when intended for the productivity of wood.

CONCLUSION

The use of data mining techniques, in particular the data classification algorithm based on decision tree allowed for the development of a model for the initial development of eucalyptus to predict new cases in different production classes. The decision tree showed to be a representative, comprehensive and understandable model that allows for the identification of decision limits of the data evaluated, contributing to the understanding of the attributes. In addition, this model helps domain experts interpret the development of Eucalyptus.

The results showed the classification models generated in the initial development of eucalyptus were very efficient to predict new cases in different classes of production, notably when the individual wood volume (IWV) and the diameter at breast height (DBH) are crucial attributes to predict the growth of Eucalyptus urograndis.

ACKNOWLEDGEMENTS

The authors would like to thank FAPESP (process number: 2013/25329-5) and the encouraging fundraising FUNDUNESP (process number: 2333/002/14-PROPe/CDC) for the finantial support, and Eldorado Brasil Celulose from Três Lagoas, Mato Grosso do Sul, Brazil state, to allow us to use the study area.

REFERENCES

• BARBOSA, B.M.; COLODETTE, J.L.; CABRAL, C.P.T.; GOMES, F.J.B.; SILVA, V.L. Efeito da fertilização na qualidade da madeira de Eucalyptus spp. Scientia Forestalis, v. 42, n. 101, p. 29-39, 2014.
• BARBOSA, C.E.M.; FERRARI, S.; CARVALHO, M.P.; PICOLII, P.R.F.; CAVALLINI, M.C.; BENETT, C.G.S.; SANTOS, D.M.A. Inter-relação da produtividade de madeira do pinus com atributos físico-químicos de um latossolo do cerrado brasileiro. Revista Árvore, v. 36, n. 1, p. 25-35, fev., 2012.
• CAMPOS, J.C.C.; LEITE, H.G. Mensuração Florestal: Perguntas e Respostas. Viçosa: Ed. UFV, 2002. 407 p.
• CARVALHO, M.P.; MENDONÇA, V.Z.; PEREIRA, F.C.B.L.; ARF, M.V.; KAPPES, C.; DALCHIAVON, F.C. Produtividade de madeira do eucalipto correlacionada com atributos do solo visando ao mapeamento de zonas específicas de manejo. Ciência rural, v. 42, n. 10 p. 1797-1903, 2012.
• COSTA, C.D.O.; ALVES, M.C.; SOUSA, A.P.; SILVA, H.R. Propriedades químicas dos solos de uma sub-bacia hidrográfica sob processo de degradação ambiental. Revista de Ciências Ambientais, v. 9, n. 2, p. 37-50, 2015.
• CRIVELENTI, R. C.; COELHO, R.M.; ADAMI, S.F.; OLIVEIRA, S.R.M. Mineração de dados para inferência de relações solo-paisagem em mapeamentos digitais de solo. Pesquisa Agropecuária Brasileira, v. 44, n. 12, p. 1707-1715, 2009.
• DAI, Q. Y.; ZHANG, C. P.; WU, H. Research of Decision Tree Classification Algorithm in Data Mining. International Journal of Database Theory and Application, v. 9, n. 5, p. 1-8, 2016.
• EMBRAPA. EMPRESA BRASILEIRA DE PESQUISA AGROPECUÁRIA. Centro Nacional de Pesquisa de Solos, 2011. Manual de métodos de análise de solos. 2. Ed. Rio de Janeiro; Embrapa, 2011, 230p.
• EMBRAPA. EMPRESA BRASILEIRA DE PESQUISA AGROPECUÁRIA. Sistema Brasileiro de Classificação de solos. 3. Ed. Rio de Janeiro: Embrapa, 2013, 306p.
• FLORIANI, M.M.P.; STEFFENS, C.A.; CHAVES, D.M.; AMARANTE, C.V.T.; PIKART, T.G.; RIBEIRO, M.S. Relação entre concentrações foliares de carboidratos solúveis totais e tolerância ao frio em diferentes espécies de Eucalyptus spp. Ciência Florestal, v. 23, n. 1, p. 165-174, 2013.
• FREITAS, P.C.; SETTE JUNIOR, C.R.; CASTRO, V.R.; GILLES CHAIX, G.; LACLAU, J.P.; TOMAZELLO FILHO, M. Efeito da disponibilidade hídrica e da aplicação de potássio e sódio nas características anatômicas do lenho juvenil de Eucalyptus grandis. Revista Árvore , v. 39, n. 2, p. 405-416, 2015.
• GAVITO, M.E., CURTIS, P.S., MIKKELSEN, T.N., JAKOBSEN, E.I. Interactive effects of soil temperature, atmospheric carbon dioxid and soil N on root development, biomass and nutrient uptake of winter wheat during vegetative growth. Journal of Experimental Botany, v. 52, p. 1913-1923, 2001.
• HAN, J.; KAMBER, M.; PEI, J., 2011. Data mining: concepts and techniques. 3^rd edition. San Francisco: Morgan Kaufmann Publishers.
• LANDIS, J.R.; KOCH, G.G. The measurement of observer agreement for categorical data. Biometrics, v. 33, n.1, p. 159-174, 1977.
• LIMA, C.G.R.; CARVALHO, M.P.; NARIMATSU, K.C.P.; SILVA, M.G.; QUEIROZ, H.A. Atributos físico-químicos de um latossolo do cerrado brasileiro e sua relação com características dendrométricas do eucalipto. Revista Brasileira de Ciência do Solo, v.34, p.163-173, 2010.
• MORAES, C.B.; CARVALHO, E.V.; ZIMBACK, L.; LUZ, O.S.L.; PIERONI, G.B.; MORI, E.S.; LEAL, T.C.A.B. Variabilidade genética em progênies de meios-irmãos de eucaliptos para tolerância ao frio. Revista Árvore , v. 39, n. 6, p. 1047-1054, 2015.
• PÉLLICO NETTO, S. Equivalência Volumétrica: Uma nova metodologia para estimativa do volume de árvores. Revista Acadêmica: Ciências Agrárias e Ambientais, v. 2, n. 1, p. 17-30, 2004.
• PENG, Y.Y.; DANG, Q. Effects of soil temperature on biomass production and allocation in seedlings of four boreal tree species. Forest Ecology and Management. v. 180, p. 1-9, 2003.
• PINKARD, E.A.; NEILSEN, W.A. Crown and stand characteristics of Eucalyptus nitens in response to initial spacing: implications for thinning. Forest Ecology and Management , v. 172, n. 2/3, p. 215-227, 2003.
• QUINLAN, J. R. C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers. 1993.
• RAIJ, B. V. Fertilidade do solo e manejo de nutrientes. Piracicaba: Ed. IPNI, 2011. 420p.
• RAIJ, B.; ANDRADE, J.C.; CANTARELLA, H.; QUAGGIO, J.A., 2001. Análise química para avaliação da fertilidade de solos tropicais. Campinas: Instituto Agronômico.
• ROSA FILHO, G.; CARVALHO, M.P.; MONTANARI, R.; SILVA, J.M.; SIQUEIRA, G.M.; ZAMBIANCO, E.C. Variabilidade espacial de propriedades dendrométricas do eucalipto e de atributos físicos de um Latossolo Vermelho. Bragantia, campinas, p. 439-446, 2011.
• SERVIÇO FLORESTAL BRASILEIRO - SFB. As florestas plantadas. Brasília: SFB, 2016. Disponível em: < http://www.florestal.gov.br/snif/recursos-florestais/estoque-das-florestas>. Acesso em: 12 mai. 2016.
  » http://www.florestal.gov.br/snif/recursos-florestais/estoque-das-florestas
• SHANNON, C.E., 1949. The mathematical theory of communication. Warren Weaver. University of Illinois Press.
• SOUZA, Z. M.; CERRI, D. G. P.; COLET, M. J.; RODRIGUES, L. H. A.; MAGALHÃES, P. S. G.; MANDONI, R. J. A. Análise dos atributos do solo e da produtividade da cultura de cana-de-açúcar com o uso da geoestatística e árvore de decisão. Ciência Rural, Santa Maria, v.40, n.4, p.840-847, abr., 2010.
• STOLF, R. Teoria e teste experimental de fórmulas de transformação dos dados de penetrômetro de impacto em resistência do solo. Revista Brasileira de Ciência do Solo , v. 15, n. 3, p. 229-235,1991.
• TOLEDO, F. H. S. F.; VENTURIN, N.; CARLOS, L.; DIAS, B. A. S.; VENTURIN, R. P.; MACEDO, R. L. G. Composto de resíduos da fabricação de papel e celulose na produção de mudas de eucalipto. Revista Brasileira de Engenharia Agrícola e Ambiental, v.19, n.7, p.711-716, 2015.
• WITTEN, I.H.; FRANK, E.; HALL, M.A. Data mining: practical machine learning tools and techniques. São Francisco, CA: The Morgan Kaufmann series in data management systems, 2011. 665p.

Publication Dates

• Publication in this collection
  Apr-Jun 2017

History

• Received
  12 Jan 2017
• Accepted
  04 May 2017