You are on page 1of 8

IAWA Journal, Vol.

30 (1), 2009: 8794

ArtificiAl neurAl networks in wood identificAtion: the cAse of two Juniperus species from the cAnAry islAnds
luis Garca esteban1*, francisco Garca fernndez1, paloma de palacios de palacios1, ruth moreno romero1 and nieves navarro cano 2
SummAry

Neural networks are complex mathematical structures inspired on biological neural networks, capable of learning from examples (training group) and extrapolating knowledge to an unknown sample (testing group). The similarity of wood structure in many species, particularly in the case of conifers, means that they cannot be differentiated using traditional methods. The use of neural networks can be an effective tool for identifying similar species with a high percentage of accuracy. This predictive method was used to differentiate Juniperus cedrus and J. phoenicea var. canariensis, both from the Canary Islands. The anatomical features of their wood are so similar that it is not possible to differentiate them using traditional methods. An artificial neural network was used to determine if this method could differentiate the two species with a high degree of probability through the biometry of their anatomy. To achieve the differentiation, a feedforward multilayer percepton network was designed, which attained 98.6% success in the training group and 92.0% success in the testing or unknown group. The proposed neural network is satisfactory for the desired purpose and enables J. cedrus and J. phoenicea var. canariensis to be differentiated with a 92% probability. Key words: Juniperus cedrus, J. phoenicea var. canariensis, artificial neural network, ANN, biometry.
INTroduCTIoN

Juniperus cedrus Webb & Berthel. and J. phoenicea L. var. canariensis Guyot (Cupressaceae) are part of the native flora of the Canary Islands. Most likely the limited number of bird species which contribute to their dissemination, in conjunction with habitat and environment changes, were influential in the two species becoming isolated in the Islands (Harry et al. 1995).
1) universidad Politcnica de madrid, Escuela Tcnica Superior de Ingenieros de montes, departamento de Ingeniera Forestal, Ctedra de Tecnologa de la madera, Ciudad universitaria s/n, 28040 madrid, Spain. 2) universidad Politcnica de madrid, Escuela universitaria de Arquitectura Tcnica, departamento de Construcciones Arquitectnicas y su Control, Ctedra de Construccin III, Ciudad universitaria s/n, 28040 madrid, Spain. *) Corresponding author [luis.garcia@upm.es].

88

IAWA Journal, Vol. 30 (1), 2009

Juniperus cedrus is a shrub or small tree with a height of up to 15 metres, found only in the islands of Tenerife, La Palma, Gran Canaria and La Gomera, at altitudes above 1400 metres. Juniperus phoenicea var. canariensis is a shrub or small tree with a height of up to 8 metres, found in the islands of Tenerife, La Palma, Gran Canaria, El Hierro and La Gomera, generally at altitudes below 1000 metres (Bramwell & Bramwell 2001). Very few studies exist on the wood anatomy of Juniperus from the Canary Islands. Peraza and Lpez (1967) conducted the most extensive study, although only on a single sample of each species. As in the case of other conifer groups, the anatomical similarity of the wood of these two species makes it practically impossible to differentiate them using traditional methods of microscopic identification of conifer wood. This means that it is impossible to identify the few wood samples found in archaeological sites or therefore to establish any hypotheses on the ancestral natural area of these species in the Canary Islands archipelago. In view of this, methods other than the usual microscopic identification of conifer wood must be used. Predictive models based on artificial neural networks (ANN), inspired on biological neural networks, have been used in similar cases in other areas of knowledge. In the field of wood they have been applied in predicting thermal conductivity using the chemical composition (Avramidis & Iliadis 2005a), in obtaining hygroscopic equilibrium points (Avramidis & Iliadis 2005b), in classifying wood defects (drake & Packianather 1998; ramrez & Chacn 2005) and even in obtaining the internal bond strength values using manufacturing parameters (Cook & Chiu 1997). However, there is no record of ANNs being used in wood identification. ANNs have been used with satisfactory results as a classification method in other areas, including behavioural sciences (reby et al. 1997), wood classification using ultrasound (Jordan et al. 1998), conifer identification using spectral data (Gong et al. 1997), identification of agricultural plant varieties (Marini et al. 2004), classification of plants and seeds (moshou et al. 2001; Kavdir 2004) and plant identification using morphological features (Clark 2003). ANNs are capable of learning from a series of examples without needing to know beforehand the relations which may exist between the variables involved in the process, by adjusting the weight of the relations between the variables in order to then predict a coherent result when new data unknown to the network is entered. ANNs can be regarded as a complex group of interconnected non-linear functions (transfer functions or neurons) capable of self-adjusting using known input and output data. They can be regarded as multiple regression models in which the algorithm allowing a solution to be reached is unknown, or where the extreme complexity of the algorithm makes it impossible to use (Prez & martn 2003). ANNs consist of three layers: an input layer, a hidden layer and an output layer. The input layer receives the initial values of the variables; the output layer shows the results of the network for the input values; and the hidden layer performs the operations designed to obtain an output. The input layer must have as many neurons as the number of input variables and the output layer must have as many neurons as the number of

Esteban et al. Artificial neural networks in wood identification

89

outputs produced by the network. However, there is no rule to enable it to be decided beforehand how many neurons should make up the hidden layer or whether the hidden layer needs to consist of more than one sublayer (Isasi & Galvn 2004), and therefore the only way to ascertain this is by a process of trial and error. Neural networks are normally represented as [n1 n 2,1 n 2,2 n 2,m n 3], where n1 is the number of neurons in the input layer, n 2,i is the number in the hidden layers and n 3 the number in the output layer (Fig. 1). The aim of this study was to use ANNs for the first time as a predictive method for differentiating species of conifer wood through the biometry of their anatomy, using wood taken from basal discs of the two species.
Hidden layers

Input layer Variable 1

W11 W21 W22

W11 B1 W21 W12 W22 B2 W31 W11 W21 W32 W42

Output layer Output

Variable 2

W23

W14

W13

W24

B3

W41

B4

Bias

Neuron

Figure 1. General structure of an ANN [2 4 2 1].

mATErIAL ANd mETHodS

All the samples were collected in the natural forests of the two species: Juniperus cedrus in the island of La Palma and J. phoenicea var. canariensis in the islands of La Palma, La Gomera and El Hierro. In each area five trees were felled, all of which were adult specimens over 70 years old, representative of the forest. The publication by Ceballos and ortuo (1951) was used to locate the natural forests. The microscopic preparations were prepared in accordance with the usual methods. Table 1 shows the features measured and the number of measurements taken using the WinCell image analysis programme. The biometry was conducted on three preparations from each tree, in all cases on mature wood from a basal disc between rings 70 and 100. The tracheid length was measured using Ladells indirect method (Ladell 1959). The height and ray frequency were measured on 5 rays contained in one square millimetre in 5 different areas on the tangential section.

90

IAWA Journal, Vol. 30 (1), 2009

Table 1. Biometric pattern.


Tracheids diameter Length diameter tracheid pits rays Height in number of cells Height Number of rays / mm 2 Cross-fields Largest diameter cross-field pits Smallest diameter cross-field pits Number of pits per cross-field 3 slides 20 trees 25 3 slides 20 trees 25 3 slides 20 trees 50 3 slides 20 trees 25 3 slides 20 trees 25 3 slides 20 trees 5 3 slides 20 trees 50 3 slides 20 trees 50 3 slides 20 trees 25 1500 1500 3000 1500 1500 300 3000 3000 1500

Table 2. Biometry of Juniperus cedrus and J. phoenicea var. canariensis.


Feature [mean value Sd and range (minmax)] Tracheids
diameter (m) Length (mm) diameter tracheid pits (m)

J. cedrus

J. phoenicea var. canariensis La Palma La Gomera El Hierro


26.7 3.6 (18.239.3) 2.1 0.8 (1.05.7) 25.8 3.5 (16.538.3) 2.4 0.5 (1.4 4.9) 30.4 5.4 (17.4 48.0) 3.3 0.9 (1.6 7.0)

27.2 4.7 (17.239.3) 2.7 0.9 (1.39.0)

15.4 1.6 (11.319.8) 2 (117)

12.5 1.4 (8.617.9) 2 (119)

13.5 1.5 (9.918.8) 2 (114)

14.6 2.4 (9.922.6) 2 (118)

rays

Height (m)

most common height in number of cells and range

Number of rays/mm 2

70.3 38.0 (15.4314.7) 65.1 7.5 (4982)

73.0 45.9 (13.4 352.9) 60.0 6.6 (4883) 5.1 0.8 (2.78.0)

55.8 27.4 (14.4 275.1) 65.2 10.8 (44 90) 5.1 0.7 (3.17.3)

58.3 33.1 (12.5277.1) 63.8 6.3 (4780) 4.8 0.7 (2.86.9) 1.8 0.3 (1.02.7) 2.2 0.9 (15) 2

Cross-fields

Largest diameter cross-field pits (m)

Smallest diameter cross-field pits (m) Number of pits per cross-field

6.4 0.8 (4.210.3) 3.2 0.5 (2.05.2) 1.7 0.8 (14) 1

2.6 0.5 (1.05.0) 1.7 0.8 (1 4) 1

2.2 0.4 (1.3 4.0) 1.9 0.9 (1 4) 2

most frequent number of pits per cross-field

Esteban et al. Artificial neural networks in wood identification

91

As the input vector is very large, for the design of the ANN an initial analysis was made of the principal components to determine if there was any type of correlation between the input variables (demuth et al. 2002). This analysis is conducted in three steps: orthogonalisation of the input vector, ordering so that the principal components come first, and removal of the correlated components. The input data were then pre-processed in order to make the network more efficient (Reby et al. 1997; drake & Packianather 1998; demuth et al. 2002; Peng et al. 2007). X X min X = X max X min X: Value after normalisation of vector X X min and X max: minimum and maximum values of vector X The characterisation of the ANN is based on the definition of the type of network and transfer functions. A feedforward multilayer percepton was used for the network, trained by means of the backpropagation algorithm, one of the most commonly used algorithms in the references consulted (reby et al. 1997; drake & Packianather 1998; myhara & Sablani 2001; Nordmark 2002; Panchariya et al. 2002; Clark 2003; Hernndez-Prez et al. 2004; Avramidis & Iliadis 2005a, 2005b; diamantopoulou 2005; Peng et al. 2007). Three different types of transfer functions were used in order to progressively adapt the output values to the final result (0 for J. cedrus and 1 for J. phoenicea var. canariensis): the linear function (purelin), the log sigmoid (logsig) and the hyperbolic tangent sigmoid (tansig), a variant of the hyperbolic tangent (myhara & Sablani 2001; Clark 2003; Hernndez-Prez et al. 2004; diamantopoulou 2005), which achieves the output much faster and is mathematically equivalent, improving the functioning of the network (demuth et al. 2002). The network training method was carried out by means of supervised learning (Hagan et al. 1995; Haykin 1998; Prez & martn 2003; Isasi & Galvn 2004). For this purpose, the results were divided into two groups selected at random and without repetition: the training group (67% of the total) and the testing group (33% of the total), with the percentages being very similar to those used in the study by Cook et al. (2000). during the training or learning phase the network is shown values of the input variables with the result of the output so that it can adapt to the problem to be solved. during the testing phase the network is shown values unknown to it and the output result is compared with the real result. For the training phase the Bayesian regularization backpropagation was used. This minimises the errors as well as their influence, thereby improving the network generalisation (demuth et al. 2002). To assess the results, both in the training and the testing phase, the formula for the degree of confidence or overall average accuracy was used (Gong et al. 1997; Clark 2003). n corr OAA(%) = 100 n tot

92

IAWA Journal, Vol. 30 (1), 2009

OAA: overall average accuracy ncorr : Number of correctly classified samples ntot : Total number of test samples The ANN Neural Network Toolbox ver. 4.0.2 application, part of the mATLAB Ver. 6.5.0. release 13 programme, was used to develop the ANN.
rESuLTS ANd dISCuSSIoN

Table 2 shows the biometry of the two species studied. Analysis of the principal components concluded that the only variables with a statistical influence on the process were the tracheid diameter, tracheid length and the smallest and largest diameters of the cross-field pits. The remaining variables involved were combinations of these four, resulting in a network input layer of four neurons. The process for obtaining the internal structure of a network is complex, with no defined methodology, which means that it is normally obtained by trial and error (Isasi & Galvn 2004). In this study, after a variety of combinations, it was found that the most effective network had a hidden layer consisting of three sublayers: the first consisting of eight neurons with a linear transfer function; the second consisting of eight neurons with a variant of the hyperbolic tangent transfer function; and the final sublayer consisting of five neurons with a log sigmoid transfer function. The output layer consists of a single neuron with a log sigmoid transfer function (Fig. 2). The output range of the log sigmoid transfer function of the output layer is a real number between zero and one. For values lower than 0.5 the output result was J. cedrus and for values higher than 0.5 it was J. phoenicea var. canariensis. For simplicity, the output result was rounded to the next integer.
Hidden sub-layer 1 Hidden sub-layer 2

Hidden sub-layer 3
Tracheid diameter Tracheid length Smallest diameter cross-field pits Largest diameter cross-field pits

Input layer

Output layer Round 1 Neuron logsig 5 Neurons logsig 8 Neurons purelin 8 Neurons tansig

Figure 2. General structure of the ANN developed [4 8 8 5 1].

Esteban et al. Artificial neural networks in wood identification

93

The values of the interlayer connections and the network biases were initialised randomly, and the network learning rate was 0.5. The training phase consisted of 80,000 cycles, after which the degree of confidence began to decrease in what is known as network overtraining (Hagan et al. 1995; Haykin 1998; Isasi & Galvn 2004). Table 3 shows the results of the network in both the training and the testing phase.
Table 3. results of the ANN.
Phase Learning Testing oAA (%) 98.6 92.0

The oAA obtained in the training phase (98.6%) and the testing phase (92.0%) is within the margins obtained by other authors in the references consulted (Gong 1997; Clark 2003; Kavdir 2004; marini et al. 2004), which means that the structure of the ANN obtained can be accepted. The parameters chosen to differentiate the two species are therefore regarded as appropriate for this purpose and the ANN obtained allows the two species to be differentiated with a probability of 92%. However, as in the case of qualitative wood identification, for the ANN to be reliable it is necessary to compare wood from similar plant parts; that is, mature wood with mature wood, juvenile with juvenile, and branchwood with branchwood. This study was conducted with wood from basal stem discs, and further studies are needed to confirm whether or not this method is also reliable in the case of wood from other parts of the tree.
ACKNoWLEdGEmENTS The authors are grateful to the forest services of the islands of La Gomera, La Palma and El Hierro for their assistance in collecting samples. This study is part of the AGL2004-02528 Project of the Spanish National Plan for Scientific Research, Development and Technological Innovation, funded by the Spanish ministry of Education and Science and the European regional development Fund (ErdF).

rEFErENCES
Avramidis, S. & L. Iliadis. 2005a. Predicting wood thermal conductivity using artificial neural networks. Wood Fiber Sci. 37: 682690. Avramidis, S. & L. Iliadis. 2005b. Wood-water sorption isotherm prediction with artificial neural networks: A preliminary study. Holzforschung 59: 336341. Bramwell, d. & Z. Bramwell. 2001. Flores silvestres de las Islas Canarias. rueda, madrid. Ceballos, L. & F. Ortuo. 1951. Estudio sobre la vegetacin y flora forestal de las Canarias Occidentales. ministerio de Agricultura, madrid. Clark, J.Y. 2003. Artificial neural networks for species identification by taxonomists. Biosystems 72: 131147. Cook, d. F. & C. C. Chiu. 1997. Predicting the internal bond strength of particleboard, utilizing a radial basis function neural network. Eng. Appl. Artif. Intell. 10: 171177.

94

IAWA Journal, Vol. 30 (1), 2009

Cook, d. F., C.T. ragsdale & r. L. major. 2000. Combining a neural network with a genetic algorithm for process parameter optimization. Eng. Appl. Artif. Intell. 13: 319396. demuth, H., m. Beale & m. Hagan. 2002. Neural Network Toolbox users guide, version 4. The mathworks Inc., mA 01760. Diamantopoulou, M. J. 2005. Artificial neural networks as an alternative tool in pine bark volume estimation. Comput. Electron. Agric. 48: 235244. drake, P.r. & m.S. Packianather. 1998. A decision tree of neural networks for classifying images of wood veneer. Int. J. Adv. manuf. Technol. 14: 280285. Gong, P., r. Pu & B. yu. 1997. Conifer species recognition: An exploratory analysis of in situ hyperspectral data. remote Sens. Environ. 62: 189200. Hagan, m.T., H. B. demuth & m. Beale. 1995. Neural Network design. PWS Pub. Co., Boston. Harry, I.S., C. martinez & T.A. Thorpe. 1995. Plantlet regeneration from mature embryos of Juniperus cedrus. Plant Cell Tissue organ Cult. 41: 7578. Haykin, S. 1998. Neural Networks: A comprehensive foundation. Ed. 2. Prentice Hall, New Jersey. Hernndez-Prez, J.A., m.A. Garca-Alvarado, G. Trystam & B. Heyd. 2004. Neural networks for the heat and mass transfer prediction during drying of cassava and mango. Innovative Food Science & Emerging Technologies 5: 5764. Isasi, P. & I.M. Galvn. 2004. Redes Neuronales Artificiales, un enfoque prctico. Pearson Educacin S.A., madrid. Jordan, R., F. Feeney, N. Nesbitt & J.A. Evertsen. 1998. Classification of wood species by neural network analysis of ultrasonic signals. ultrasonics 36: 219222. Kavdir, . 2004. Discrimination of sunflower, weed and soil by artificial neural networks. Comput. Electron. Agric. 44: 153160. Ladell, J.T. 1959. A new method of measuring tracheid length. Forestry 32: 124125. Marini, F., J. Zupan & A. Magri. 2004. On the use of counterpropagation artificial neural networks to characterize Italian rice varieties. Anal. Chim. Acta 510: 231240. moshou, d., E. Vrindt, B. de Ketelaere, J. de Baerdemaeker & H. ramon. 2001. A neural network based plant classifier. Comput. Electron. Agric. 31: 516. Myhara, R. M. & S. Sablani. 2001. Unification of fruit water sorption isotherm using artificial neural networks. dry. Technol. 19(8): 15431554. Nordmark, U. 2002. Knot identification from CT images of young Pinus sylvestris sawlogs using artificial neural networks. Scand. J. Forest Res. 17: 7278. Panchariya, P. C., d. Popovic & A. L. Sharma. 2002. desorption isotherm modeling of black tea using artificial neural networks. Dry. Technol. 20(2): 351362. Peng, G., X. Chen, W. Wu & X. Jiang. 2007. modeling of water sorption isotherm for corn starch. J. Food Eng. 80: 562567. Peraza, C. & A. Lpez. 1967. Estudio de las principales maderas de Canarias. IFIE, madrid. Prez, m. L. & Q. martn. 2003. Aplicaciones de las redes neuronales a la estadstica. Cuadernos de Estadstica. La muralla S.A., madrid. Ramrez G. & M.I. Chacn. 2005. Clasificacin de defectos en madera utilizando redes neuronales artificiales. Computacin y Sistemas. 9: 1727. Reby, D., S. Lek, I. Dimopoulos, J. Joachim, J. Lauga & S. Aulagnier. 1997. Artificial neural networks as a classification method in the behavioural sciences. Behav. Processes 40: 3543.

You might also like