Journal of Spectroscopy

Journal of Spectroscopy / 2021 / Article

Research Article | Open Access

Volume 2021 |Article ID 6088435 | https://doi.org/10.1155/2021/6088435

Peng Zhao, Zhen-Yu Li, Cheng-Kun Wang, "Wood Species Recognition Based on Visible and Near-Infrared Spectral Analysis Using Fuzzy Reasoning and Decision-Level Fusion", Journal of Spectroscopy, vol. 2021, Article ID 6088435, 16 pages, 2021. https://doi.org/10.1155/2021/6088435

Wood Species Recognition Based on Visible and Near-Infrared Spectral Analysis Using Fuzzy Reasoning and Decision-Level Fusion

Academic Editor: Daniel Cozzolino
Received04 Jun 2021
Revised09 Jul 2021
Accepted12 Jul 2021
Published23 Jul 2021

Abstract

A novel wood species spectral classification scheme is proposed based on a fuzzy rule classifier. The visible/near-infrared (VIS/NIR) spectral reflectance curve of a wood sample’s cross section was captured using a USB 2000-VIS-NIR spectrometer and a FLAME-NIR spectrometer. First, the wood spectral curve—with spectral bands of 376.64–779.84 nm and 950–1650 nm—was processed using the principal component analysis (PCA) dimension reduction algorithm. The wood spectral data were divided into two datasets, namely, training and testing sets. The training set was used to generate the membership functions and the initial fuzzy rule set, with the fuzzy rule being adjusted to supplement and refine the classification rules to form a perfect fuzzy rule set. Second, a fuzzy classifier was applied to the VIS and NIR bands. An improved decision-level fusion scheme based on the Dempster–Shafer (D-S) evidential theory was proposed to further improve the accuracy of wood species recognition. The test results using the testing set indicated that the overall recognition accuracy (ORA) of our scheme reached 94.76% for 50 wood species, which is superior to that of conventional classification algorithms and recent state-of-the-art wood species classification schemes. This method can rapidly achieve good recognition results, especially using small datasets, owing to its low computational time and space complexity.

1. Introduction

Wood species classification has been investigated for several years, as different wood species have diverse physical and chemical properties. Many such schemes have been proposed for automatic processing using sensors and computers, with spectra-based [13], chemometric-based [4], and image-based schemes [59] being the most commonly investigated. Spectral analysis technology is also widely used in the prediction of wood properties [1012] such as wood density [13], wood moisture content [14, 15], and others [16, 17]. The NIR spectral band is often used in these applications. A wood spectral analysis scheme usually pertains to one-dimensional (1D) spectral reflectance curves, which exhibits low computational complexity and, thus, high processing speeds. A wood image processing scheme usually considers the wood’s texture features for species classification, as its texture is more stable than its color [1820].

Fuzzy logic can provide nonseparated output classes and functions well with just a few learning datasets. In the past, fuzzy-rule-based schemes were successfully used in control systems and pattern recognition applications [21, 22]. In the wood industry, fuzzy rule classification has been successfully applied to wood color classification and wood defect recognition [2325]. For example, considering the uncertainty of wood color and processing speed, Bombardier et al. proposed a fuzzy rule classifier to fulfill wood color recognition [24, 25].

Fuzzy rule classification is not frequently used in wood species recognition. However, Ibrahim et al. proposed a fuzzy preclassifier to classify 48 tropical wood species into four broad categories, followed by the use of a support vector machine (SVM) classifier in each broad category to further determine the wood species of a sample [19]. One advantage of their approach was that when a new wood species was added to the system, only the SVM classifier of one broad category required retraining instead of the entire system. Similarly, Yusof et al. also proposed a fuzzy-logic-based preclassifier to classify tropical wood species into four broad categories, followed by performing specific wood species recognition using definite classifiers to increase recognition accuracy and decrease processing time requirements [7].

In this study, we propose a fuzzy rule classification scheme for the recognition of 50 wood species using 1D spectral reflectance data. The spectral range covers both the visible (VIS) band (i.e., 376.64–779.84 nm) and the near-infrared (NIR) band (i.e., 950–1650 nm). For the wood spectral curves, the VIS band may vary as the color of a wood sample may change owing to environmental variations, while the NIR band remains relatively stable. Consequently, these two bands were processed using different fuzzy classifiers, and the two classification results were further processed using an improved decision-level fusion based on the D-S evidential theory.

Our scheme has the two following advantages. First, it can achieve good recognition accuracy rapidly in a small spectral dataset. There are only 50 samples for each wood species, and each sample consists of a VIS spectral curve and an NIR curve. Second, the experimental equipment cost is low. The used Ocean Optics USB2000-VIS-NIR spectrometer costs approximately $3,000, while the Ocean Optics FLAME-NIR spectrometer costs $8,000. Several comparisons with other classical and state-of-the-art classifiers in wood species recognition also verified the merits of our scheme.

2. Materials and Methods

2.1. Wood Samples and Spectral Data Acquisition

Our wood dataset consists of 50 wood species, as shown in Table 1 and Figure 1. These 50 wood species consist of softwood and hardwood species, which are representative tree species in the wood market. Moreover, this dataset also consists of some similar wood species (i.e., visually similar wood species and intragenus different species). Each wood sample was processed into a 2 × 2 × 3 cm block. We used the Ocean Optics USB2000-VIS-NIR microfiber spectrometer and an Ocean Optics FLAME-NIR microfiber spectrometer, visible and near-infrared (VIS/NIR) spectra of the wood samples being obtained using these devices. Fifty mean spectral curves were detected by using the spectrometer for each wood species—that is, there are 2500 VIS spectral curves and 2500 NIR spectral curves for 50 wood species. Our experimental setup consisted of a halogen lamp, a universal serial bus, a computer, a customized holder, a fiber spectrometer, and a section of optical fiber, as shown in Figure 2. Spectral acquisition was performed in a controlled environment at a temperature interval of . The VIS spectral band covered a range of 376.64–779.84 nm, with a spectral resolution of approximately 0.3 nm. The NIR spectral band covered a range of 950–1650 nm, with a spectral resolution of approximately 5.4 nm. The spectrometer parameter settings, spectral data acquisition, and data storage were conducted using the Spectrasuite software. The integration time was 300 ms, and the mean number of measurements at each point was 900. Before spectral acquisition, the reference and dark spectra were measured and stored, the reference spectra were acquired using a standard white panel of known 100% reflectance, and the dark spectra were acquired by blocking the radians. For each wood sample, four diffuse reflectance spectra were measured randomly using (1)—where represented the received intensity, dark intensity, and white panel intensity, respectively—at different locations across the wood sample’s cross section. In fact, wood cross section has plentiful wood structures such as pores and parenchyma, making its spectra more suitable for species recognition, which is superior to radial direction and tangential direction in wood species recognition [26] and wood density prediction [13]. Subsequently, the mean spectrum of each wood sample was calculated and stored for subsequent spectral analysis and species classification. Figures 3 and 4 show the acquired mean spectra of the 50 wood species with standard normal variables (SNVs).


Class labelLatin name

1Pterocarpus soyauxii
2Ulmus glabra
3Calophyllum brasiliense
4Intsia bijuga
5Tectona grandis
6Pouteria speciosa
7Quercus mongolica
8Magnolia fordiana
9Guibourtia ehie
10Terninalia catappa
11Cinnamomum camphora
12Guibourtia demeusei
13Swintonia florbunda
14Fraxinus mandshurica
15Pometia pinnata
16Dipterocarpus alatus
17Vernicia fordii
18Sophora japonica
19Robinia pseudoacacia
20Quercus acutissima
21Populus alba
22Fraxinus chinensis
23Pinus sylvestris
24Populus tomentosa
25Betula alnoides
26Betula platyphylla
27Platanus orientalis
28Picea asperata
29Larix gmelinii
30Salix matsudana
31Amygdalus davidiana
32Juglans mandshurica
33Pinus koraiensis
34Populus cathayana
35Tona ciliate
36Prunus avium
37Pinus massoniana
38Aucoumea klaineana
39Shorea contorta
40Entandrophragma candollei
41Millettia laurentii
42Swietenia mahagoni
43Rhodamnia dumetorum
44Cyclobalanopsis glauca
45Shorea laevis
46Chamaecyparis nootkatensis
47Juglans nigra
48Tilia mandshurica
49Pseudotsuga menziesii
50Pinups radiata

The spectral range of the USB2000-VIS-NIR spectrometer was 340–1027 nm. Both ends of this spectral range can be noisy, so the useful spectral range is 376.64–779.84 nm. The spectral data in the VIS band may be variable as the color of a wood sample may change with changing environmental conditions. To avoid this, we maintained a stable environment during spectral acquisition and wood sample preservation. The NIR band of 950–1650 nm is relatively stable and can be used to characterize different wood properties to some extent. Therefore, these two bands were processed using different fuzzy classifiers—that is, with different membership functions and fuzzy rules—the two classification results of which were further processed using an improved decision-level fusion based on the D-S evidential theory.

2.2. Spectral Dimension Reduction

For each spectral curve, the VIS/NIR band includes thousands of spectral data, so that the spectral dimension is very large. Consequently, each spectral curve contains redundant information, which reduces the wood species classification accuracy and speed. We used principal component analysis (PCA) and the T-distributed stochastic neighbor embedding (T-SNE) algorithms to compare and reduce the spectral dimension results. The PCA algorithm is a transformation algorithm in multivariate statistics, which is based on the variance maximization of a mapped low-dimension vector. The procedure is as follows: First, the initial dataset matrix is defined as —where m and n are the dimension and number of feature vectors, respectively. The mean vector is computed as . Second, we compute to form . Third, the k largest eigenvalues of are selected, their k corresponding eigenvectors forming a matrix . Finally, the new dataset matrix after dimension reduction is computed as .

Additionally, the T-SNE algorithm was also used for spectral dimension reduction. This algorithm performs nonlinear dimension reduction, an improvement on the original SNE algorithm [27]. The algorithm is based on the invariant data distribution probability of the mapped low-dimensional vector. Here, the symbol “T” in T-SNE represents the t-distribution, the freedom of the t-distribution being 1 in this study. The graphs of the three-dimensional (3D) spectral vectors for the VIS and NIR bands after spectral dimension reduction using the PCA and T-SNE algorithms are shown in Figure 5. The contribution rate (CR) and cumulative contribution rate (CCR) of the first ten principal components are shown in Figure 6 and Table 2 for the VIS and NIR bands. In Figure 5, the wood sample’s 3D spectral vectors were acquired using the PCA and T-SNE algorithms and are shown with 50 different colors for the 50 wood species. The best case is shown in Figure 5(b), where the wood samples were the most dispersed.


PC01PC02PC03PC04PC05PC06PC07PC08PC09PC10

VIS CR0.88040.07260.03010.00700.00610.00180.000749820.000559080.000478820.00027888
VIS CCR0.88040.95290.98300.99000.99610.99790.99860.99920.99970.9999
NIR CR0.82150.14680.01550.00970.00290.00210.000823300.000230930.000173110.00010637
NIR CCR0.82150.96830.98380.99350.99650.99870.99940.99970.99980.9999

2.3. Fuzzy Rule Classifier

A fuzzy rule classifier is a fuzzy reasoning system based on fuzzy linguistic rules, its core component being a fuzzy rule set. When an input sample is transformed into the rule set of one wood species, this sample is then classified as a wood species. The overall fuzzy reasoning system is shown in Figure 7 and consists of two parts: the training and testing processes. The steps in Figure 7 are presented as follows.Step 1. The membership functions are set to fuzzify the training sample’s feature and generate the training sample’s fuzzy rulesStep 2. Fuzzy rules are adjusted by using a corrected confidence coefficientStep 3. The fuzzy rule set is generatedStep 4. The testing sample’s feature is fuzzified, and the generated rule is compared with the fuzzy rule set for wood species classification

The training process consisted of three steps. First, membership functions were used to fulfill the fuzzification of the training dataset. Here, the membership functions could be a triangular function, trapezoid function, or Gaussian function. Second, fuzzy rules were generated based on the wood species of the wood samples in the training dataset. A fuzzy rule was defined as “IF … AND …, THEN …” using the conjunction or disjunction operation. Third, the fuzzy rules required modification using a confidence coefficient to form a perfect fuzzy rule set.

In the testing process, the testing dataset was added to the trained fuzzy rule classifier, and the same membership functions were used to fulfill its fuzzification. In this way, a corresponding fuzzy rule of a testing sample could be obtained, and this rule was compared with the generated perfect fuzzy rule set to obtain the classified wood species. This species was then compared with the correctly labeled species of the testing sample to evaluate the performance of the fuzzy rule classifier. If the corresponding fuzzy rule of a testing sample was not found in the perfect fuzzy rule set, the sample was classified as an unknown species.

2.3.1. Fuzzification of the Feature Vector

Here, the feature vector refers to the spectral vector after dimension reduction. This fuzzification step transforms spectral feature vectors into linguistic variables. A linguistic variable is defined using a triple value , where is a variable (e.g., PC1) defined on a set of reference , is the universe of discourse (field of variation of ), and is the vocabulary chosen to describe the values of in a symbolic way. The set contains the normalized fuzzy subsets of , and each fuzzy subset is defined by the membership degree .

For example, the wood spectral feature vector is obtained using the PCA algorithm. The fourth principal component (PC4) of the first ten wood species is shown in Table 3 for the NIR band, where the maximum, minimum, mean, and median values of PC4 are shown. Thus, we can design the membership functions to divide the NIR PC4 (i.e., the 4th dimension) into twenty-six intervals based on these values, as shown in Figure 8.


Wood speciesMinimum valueMean valueMedian valueMaximum value

Pterocarpus soyauxii−0.292834701−0.16041152−0.1699747120.007724295
Ulmus glabra0.0253157880.1957505660.2108186090.363952969
Calophyllum brasiliense−0.192778352−0.117200877−0.117229875−0.034786326
Intsia bijuga−0.389917083−0.139350711−0.1377726120.114471138
Tectona grandis−0.0575578260.1481586360.1533149030.280022623
Pouteria speciosa−0.0950309850.0538110880.0500285470.191835329
Quercus mongolica−0.0629643380.0561465120.0465165520.338022639
Magnolia fordiana−0.154656742−0.037434328−0.0481849820.126992175
Guibourtia ehie−0.338868726−0.178740936−0.2012918660.085960311
Terninalia catappa−0.1871783530.0607163560.0537035390.267102296

As the different spectral PCs or dimensions may be divided into different intervals, the overall number of fuzzy rules is the product of the interval number of different PCs or dimensions:where N is the dimension of the spectral feature vectors and is the number of intervals for variable (i.e., the element of the spectral feature vector). In fact, can be determined by both N and . Therefore, a large N or produces a large resulting in high computational complexity. In other words, N and should be designed to be as small as possible in view of the recognition accuracy of our fuzzy rule classifier. If they cannot be simultaneously small, the dimension N should be as small as possible.

For instance, if the PCA algorithm is used for wood spectral dimension reduction, the CCR should be above 99% in view of the PC number. As can be seen from Table 2, the CCR of the first four PCs reaches 99% and 99.35% for the VIS and NIR bands, respectively, but remains approximately stable when , so we should set .

After several tests with the PCA algorithm, N and were as follows: for the VIS band, , . For the NIR band, , . The triangular, trapezoid, and triangular-trapezoid functions were used and compared 50 times—in terms of the ORA and mean training time requirement (i.e., the wood sample number of the training set was 2000) for the fuzzy rule set—as described in Tables 4 and 5 for the VIS and NIR bands, respectively. The maximum and practical in terms of N and are listed in Table 6. We observed that the training time requirement for the fuzzy rule set was practically invariant with regard to the fuzzy classifiers based on different membership functions. However, this time requirement was only relevant to the N dimension of the spectral feature vectors. In fact, both the time requirement and the increased rapidly when the dimension N increased. Moreover, the overall recognition accuracy in terms of the triangular-function-based fuzzy classifier increased when the dimension N increased, but the trapezoid-function- or triangular-trapezoid-function-based classifier remained practically invariant when the dimensions were and for the VIS and NIR bands, respectively. Finally, using our fuzzy classifier, the optimal trapezoid function parameters were dimension , , and for the VIS band and for the NIR band.


Dimension NTriangular functionTrapezoid functionTriangular-trapezoid
ORA (%)TR (s)ORA (%)TR (s)ORA (%)TR (s)

375.651051.2776.851154.1277.801219.26
489.0524393.2489.6025128.7590.2024983.20
589.60312764.5290.05299023.2090.25327279.92
690.20556719.4790.40469466.5390.65483718.36


Dimension NTriangular functionTrapezoid functionTriangular-trapezoid
ORA (%)TR (s)ORA (%)TR (s)ORA (%)TR (s)

377.651822.2382.801733.4681.151807.12
490.2545125.0292.9044783.2493.2548353.50
591.55657271.1593.10631285.4692.85718375.61
692.051038488.1893.05940615.3493.351223481.25


Dimension N
345

(VIS) maximum fuzzy rule number4532595182514277375
(VIS) practical fuzzy rule number98713621475
(NIR) maximum fuzzy rule number52640136864027372800
(NIR) practical fuzzy rule number113516741689

2.3.2. Fuzzy Rule Generation

After the fuzzification of the spectral feature vectors, the fuzzy rules in the form of “IF … AND …, THEN …” required definition. For example, if three linguistic variables are considered for input and the output is wood species Y, the fuzzy rule is as follows:

This expression indicates that when is classified as , is classified as …, and the output class is . Therefore, we need to calculate the membership degrees for input . In this expression, “AND” is a Cartesian product between the linguistic variables and the operation is performed with a T-Norm, as follows:

This equation is then used to calculate the product of the membership degrees for the spectral linguistic variables :

For one wood sample of wood class/species l, there are often some products of the membership degrees . For example, if are not equal to 0, we obtain products of the membership degrees for one wood sample (i.e., ). We should then choose the largest product to form the corresponding rule. Thus, we obtain a fuzzy rule for wood class/species l as follows:

2.3.3. Fuzzy Rule Adjustment

When the fuzzy rule is generated, there are usually several rules with the same judgment condition (i.e., the same “IF … AND …,”) but different classification results (i.e., different “THEN …”). Consequently, fuzzy rule adjustment is usually required to add a confidence coefficient (CC) to each fuzzy rule.

First, for one wood sample of wood class/species l, the sum of the products of the membership degrees is calculated as follows:

Then, the mean value is calculated using

Thus, we can compute the CC of the fuzzy rule produced as follows:

Moreover, when the same rules are produced by other samples, the CC of the fuzzy rule should be increased as follows:

If the same rule is not produced by other samples, the CC of this fuzzy rule remains invariant. Some final rules with their CCs are listed in Table 7. In practical classification work, when there are rules that have the same judgment condition but different classification results, the final classification is determined by the fuzzy rule with the largest CC.


ABCDCCWood species

224860.9939Pinups radiata
5209150.9876Tilia mandshurica
282920150.9072Juglans nigra
1721260.7993Entandrophragma candollei
3231160.6161Chamaecyparis nootkatensis

2.4. Decision-Level Fusion

The D-S evidential theory [28] is an effective decision-level fusion scheme, but the classifier’s reliability is not considered fully when each classifier’s result is converted into a classification probability distribution before the decision-level fusion. To solve this problem, we proposed an improved D-S decision-level fusion method that determines each classifier’s weight in view of the differences in the classification accuracy ratio and correlation coefficient.

As the NIR band is more stable than the VIS band, these two bands should be processed using different fuzzy parameters (e.g., different membership functions). Consequently, spectral dimension reduction and fuzzy rule classification were applied to the two bands, respectively, so that the two classification results could be fused with our improved D-S decision-level fusion, as described below. The overall improved D-S decision-level fusion model is shown in Figure 9. The steps in Figure 9 are outlined as follows:Step 1. The basic classification probability distribution of two classifiers is generated using classification resultsStep 2. Classifier weights are determined using classification the accuracy ratio (CAR) and classification correlation coefficient (CCC)Step 3. The adjusted classification probability distribution is generated using classifier weightsStep 4. The final classification results are calculated using D-S decision-level fusion

2.4.1. Conversion from the Classification Result to Classification Probability Distribution

Based on (4), we calculate the products of membership degrees for the dimension of the VIS spectral feature vectors or the dimension of the NIR spectral vectors for one wood sample. Every product and its corresponding rule form one row of the matrix :where n is the maximum rule number such that for and for . refers to rule l with product .

Then, the rows with are deleted in this matrix. The rule in every row in the matrix is checked to determine whether it belongs to the fuzzy rule set—Table 7—in Section 2.3.3. The rules in the fuzzy rule set remain, while those that are not in the set are deleted in order to form the matrix :

The added final column refers to the corresponding wood species. Some rules in correspond to the same wood species (i.e., for ). For an that corresponds to several rules, the maximum rule (i.e., with the largest ) remains, while the others are deleted, so that the matrix is formed. Therefore, for one wood sample, the classification probability distribution can be computed as follows:

Here, is the number of . Note that, for wood species that do not exist in , their corresponding classification probabilities are set as 0. For example, the classification probability distribution in the VIS and NIR bands for a wood sample is shown in Table 8.


Class labelVIS FRC (%)NIR FRC (%)

10.000.00
20.000.00
30.000.00
40.000.00
50.000.00
627.222.75
70.740.00
80.000.00
90.000.00
100.000.00
110.000.00
1221.4593.17
130.000.00
141.680.00
150.000.00
160.000.00
1714.370.00
180.000.00
190.290.00
200.000.00
210.000.00
220.000.00
230.000.07
2413.150.17
250.000.00
260.000.00
270.000.00
284.700.00
290.000.62
300.000.00
310.000.00
320.000.00
330.000.00
340.000.00
350.000.00
360.000.00
370.002.12
380.000.00
390.000.00
400.000.00
413.650.00
424.320.00
430.000.00
440.000.35
450.000.00
467.920.00
470.000.00
480.000.75
490.510.00
500.000.00

2.4.2. Weight Determination of VIS and NIR Fuzzy Rule Classifier

In this section, we propose a novel classifier’s weight determination based on the difference between the fuzzy classification accuracy ratio (CAR) and the correlation coefficient for the VIS and NIR bands using an analytical hierarchy process. In Section 2.3.1, we concluded that the fuzzy classification accuracy was 90.20% after the PCA algorithm processing for , while it was 93.25% for in the training set.

Regarding the calculation of the classification correlation coefficient (CCC), the training set is used to calculate the classification probability distribution for each wood sample using (13). Then, the following equation is used to calculate the CCC for the VIS and NIR classifiers:

Here, refers to the covariance between the corrected and actual classification probability distributions. For instance, as described in Table 8, the actual classification probability distribution for one wood sample for the NIR and VIS classifiers was and , respectively, while the corrected one was . The represent the standard deviations for the corrected and actual classification probability distribution vectors in the training set, respectively. The computed fuzzy CCC was 0.8073 after PCA algorithm processing for , whereas it was 0.8862 for .

The difference in the CAR was , while that of CCC was . The membership functions are shown in Figure 10. The corresponding credibility for these interval differences is shown in Table 9. We observed that the was in the 1st interval, which indicated that the VIS fuzzy and NIR fuzzy classifiers had the same credibility in view of the CAR, while was in the 2nd interval, which indicated that the NIR fuzzy classifier was more credible than the VIS classifier in view of the CCC. To determine the classifier’s weight using (17), we first need to compute the classifier’s scores using (15) and then compute the mean score using (16):


RR and CC differenceCredibilityDescription

1st interval1The same trusted
2nd interval2Between 1 and 3
3rd interval3A little trusted
4th interval4Between 3 and 5
5th interval5Obviously trusted
6th interval6Between 5 and 7
7th interval7Strongly trusted
8th interval8Between 7 and 9
9th interval9Extremely trusted

2.4.3. Improved D-S Decision-Level Fusion Based on the Classifier’s Weight

In this section, we modified the classification probability distribution, as shown in Table 8, in the VIS and NIR bands using the weight of the classifier. First, the maximum probability was modified using (18), and then, we computed the sum of this modified maximum probability with the other 19 probabilities using (19). Finally, we obtained the modified classification probability distribution using (20). As an example, the classification probability distribution in Table 8 was modified in this manner, as shown in Table 10.


Class labelVIS FRC (%)NIR FRC (%)

10.000.00
20.000.00
30.000.00
40.000.00
50.000.00
619.952.10
70.810.00
80.000.00
90.000.00
100.000.00
110.000.00
1223.5994.78
130.000.00
141.850.00
150.000.00
160.000.00
1715.810.00
180.000.00
190.320.00
200.000.00
210.000.00
220.000.00
230.000.05
2414.460.13
250.000.00
260.000.00
270.000.00
285.170.00
290.000.47
300.000.00
310.000.00
320.000.00
330.000.00
340.000.00
350.000.00
360.000.00
370.001.62
380.000.00
390.000.00
400.000.00
414.010.00
424.750.00
430.000.00
440.000.28
450.000.00
468.710.00
470.000.00
480.000.57
490.570.00
500.000.00

Finally, we used the modified VIS and NIR classification probability distributions to compute the normalized constant and composed mass function as follows (i.e., ):

Here, is the ith wood species with . Based on these equations, we can compute 50 mass functions . The first two largest are defined as , respectively. refers to the mass function for an uncertain set, and in our wood species, the recognition work . Wood sample x is then classified as wood species with the following rules: ). Otherwise, the judgment is the “rejected classification”:

3. Results and Comparisons

In our experiments, there were 2,500 spectral curves for 50 wood species. These samples were divided into a training set and a testing set using k-fold cross validation [29] with , so that the training set consisted of 2000 samples and the testing set consisted of 500 samples. Consequently, there were 40 training samples and 10 testing samples for each wood species. The PC1 and PC2 of the training and testing datasets were distributed as shown in Figure 11, where the distributions of both the training and testing sets were dispersed and uniform, indicating that the dataset was suitable for pattern classification.

3.1. Comparisons with Conventional Algorithms

The fuzzy rule classifier was applied to the wood species spectral classification experiments and compared with other important classification schemes, such as the Bayes classifier (BC) [30], random forest (RF) [31], BP network (BPN) [32], SVM [33], and LeNet-5 [34]. The overall recognition accuracy (ORA) and processing speed for a testing sample from our testing dataset (i.e., time requirement, TR) are listed in Table 11. The computer configurations used are presented in Table 12. The ORA was calculated as follows (i.e., is the sample number correctly classified, and is the total sample number of testing sets):


ModelDatasetORA (%)TR (ms)

BCVIS (PCA7dim)76.200.050256‬
NIR (PCA7dim)77.800.037920
VIS (TSNE7dim)55.000.047946
NIR (TSNE7dim)61.600.052136‬

RFVIS (PCA7dim)83.802.437714
NIR (PCA7dim)84.002.326404
VIS (TSNE7dim)76.402.654658
NIR (TSNE7dim)80.402.310316

BPNVIS (PCA7dim)63.840.135030
NIR (PCA7dim)61.480.117026
VIS (TSNE7dim)54.040.123924
NIR (TSNE7dim)60.800.140712

LIBSVMVIS (PCA7dim)83.60 (linear)0.092936
54.80 (RBF)0.123525
NIR (PCA7dim)84.40 (linear)0.081890
80.20 (RBF)0.131751
VIS (TSNE7dim)72.40 (linear)0.086184
23.80 (RBF)0.253162
NIR (TSNE7dim)77.60 (linear)0.091526
23.00 (RBF)0.211325

LeNet-5VIS92.400.763721

FRCVIS (PCA4dim)90.200.137274
NIR (PCA4dim)92.920.141672
VIS (TSNE4dim)71.880.159326
NIR (TSNE4dim)81.120.173528

D-S FRCVIS + NIR (PCA4dim)93.843.026516
FW-D-S FRCVIS + NIR (PCA4dim)94.763.071578

VIS: visible band; NIR: near-infrared band; FRC: fuzzy reasoning classifier; PCA: principal component analysis; BC: Bayes classifier; RF: random forest; BPN: BP network; CNN: convolutional neural network; LIBSVM: support vector machine; T-SNE: T-distributed stochastic neighbor embedding; D-S: Dempster–Shafer.

ConfigurationModelParameter

CPUIntel I5-84006 core 6 thread, CPU clock speed 2.8 GHz
RAMSamsung8 GB, DDR4 2800 MHz
Hard diskSamsung256 GB, SSD read: 470 MB/s, write: 473 MB/s

In the RF classification, the time complexity was —where m is the feature dimension, d is the tree depth, and the RF number is 500. In the BP classification, the time complexity was — where refers to the node number of the ith layer and d refers to the overall layer number, including the input and output layers. The structure of the BP network was 7 (input layer)-31 (hidden layer)-20 (output layer), and the transfer functions for the input and hidden layers were logsig and purelin, respectively. The BP network’s iteration times were 1000 with a target error and learning rate of 0.01. In the LeNet-5 classification, spectral dimension reduction was not used, whereas the spectral band 451–984 nm was used to form a matrix. The training dataset was used to train a CNN with a structure comprising six convolution-2 pooling-12 convolution-2 pooling. The size of the convolution cores in the convolutional layers was 5, with a learning rate of 1 and training times of 1000. The time complexity was —where is the size of the output feature image in the ith layer, is the size of the convolution cores in the ith convolutional layer, is the core number of the ith convolutional layer, and d is the overall convolutional layer number. In the SVM classification, a linear kernel function for the PCA algorithm and a radial basis function (RBF) for the T-SNE were used with a penalty factor of 1.2.

In our fuzzy reasoning classifier (FRC) classification, we observed that the PCA algorithm performed much better than the T-SNE algorithm in terms of spectral dimension reduction. The best ORA of 96.3636% was achieved with our FRC, and the D-S decision-level fusion scheme was improved with the use of four PCs for the VIS band and six PCs for the NIR band, with a TR of 2.535915 ms for one testing sample.

In other words, the RF algorithm could achieve good and efficient classification results using a large training dataset. However, its classification performance was poor with our small dataset—our training dataset only consisted of 2000 wood samples—and with low-dimensional feature vectors. Similarly, the BP neural network and LeNet-5 also required a large training dataset to complete network training. Regarding spectral dimension reduction, the classification results using our FRC and T-SNE were relatively poor compared with those of the PCA algorithm as the T-SNE is a nonlinear dimension reduction algorithm.

3.2. Comparisons with State-of-the-Art Algorithms

We compared the proposed scheme with other state-of-the-art wood species recognition schemes, all of which performed wood species classification using digital images of wood sample cross sections. We set up the image collection experimental equipment, as shown in Figure 12, consisting of a charge-coupled device (CCD) camera, an optical microscope with a magnification ratio of 10–100, and a light-emitting diode (LED) light source. The wood color image selected was of size pixels and magnification ratio 50. We used this imaging equipment to capture the cross-sectional color image of each wood sample so that each species included 50 images. This color image was then converted into a grayscale image of size , although the color image was used in the CNN scheme. The computer configuration used is presented in Table 13, and the classification performance comparison is given in Table 14. We found that our FRC and improved D-S decision-level fusion scheme outperformed all other state-of-the-art classification schemes in terms of ORA.


ConfigurationModelParameter

CPUIntel I7–9700F8 core 8 thread, CPU clock speed 3.00 GHz
GPUNVIDIA GeForce8 GB, GDDR6 1650 MHz
RTX 2060 SUPER2076 stream processor
RAMCorsair Memory16 GB, DDR4 3000 MHZ
Hard diskIntel256 GB, SSD read: 1570 MB/s, write 540 MB/s


ReferenceModelORA (%)

[35]I-BGLAM72.32
[36]LBP68.68
[19, 37]SPPD7.32
SPPD + I-BGLAM32.68
Fuzzy + SPPD + BGLAM34.68
[18]GA25.68
GA + KDA29.00
[38]VGG1683.36
[39]SqueezeNet86.28
[40]ResNet1894.40
[41]GoogLeNet92.80
[42]CNN83.00
Our schemeFRC + improved D-S fusion94.76

ORA: overall recognition accuracy; TR: time requirement; I-BGLAM: improved basic gray-level aura matrix; LBP: local binary pattern; SPPD: statistical property of pore distribution; GA: genetic algorithm; KDA: kernel discriminant analysis; CNN: convolutional neural network; FRC: fuzzy reasoning classifier; D-S: Dempster–Shafer.

In these state-of-the-art wood species recognition schemes, Yusof et al. employed texture feature operators (e.g., basic gray-level aura matrix (BGLAM), improved basic gray-level aura matrix (I-BGLAM)), and pore distribution features (i.e., statistical properties of pores distribution (SPPD)) and several feature extraction algorithms, such as the genetic algorithm (GA) and kernel discriminant analysis (KDA) to classify more than 50 tropical wood species [18, 19, 35, 37]. However, their schemes were only suitable for the classification of hardwood species with pores. For softwood species without pores, the classification performance of their schemes was poor. Because our 50 wood species consist of both hardwood and softwood species, their schemes delivered poor classification results. Another texture feature operator, the local binary pattern (LBP), and a deep learning neural network were also used for classification comparisons [36, 3842].

4. Conclusions

In this study, a novel wood species spectral classification scheme was proposed based on a fuzzy linguistic rule classifier and an improved decision-level fusion technique based on the D-S evidential theory. The VIS/NIR spectral reflectance curves of a wood sample’s cross section were measured using a USB 2000-VIS-NIR spectrometer and a FLAME-NIR spectrometer. As the VIS and NIR bands for the wood sample’s cross section have different spectral stabilities, they were processed using two different FRCs, and their classification results were further fused using an improved decision-level fusion technique based on the D-S evidential theory. The test results showed that the best ORA of our scheme reached 94.76% for 50 wood species, with little computational overhead.

More generally, our scheme has the following advantages compared with existing wood species recognition schemes. First, the acquired wood spectral curves were usually noisy and variable because of the variability of the external environment—such as optical illumination intensity—and the wood sample’s internal factors—such as humidity, surface roughness, and intraspecies differences. The uncertainty of these wood spectral features could be resolved using fuzzy processing to some extent. Second, our scheme rapidly achieved good recognition results, particularly with small datasets, and the cost of the VIS/NIR spectrometer used was relatively low. The spectral dataset was a 1D signal set, which had a relatively low computational overhead compared to other 2D image datasets. Therefore, our scheme could be readily applied to wood recognition and classification in the industry in the near future.

However, our scheme can only classify the known wood species which is included in our training dataset. In practical work, some unknown wood species samples may be encountered by our wood classification system. This issue is an unknown species identification problem under the open-world assumption, which can be solved in the certainty or uncertainty framework. For example, in an uncertainty framework, Tang et al. have proposed a new method for generation of generalized basic probability assignment (GBPA) based on evidence theory [43]. In this method, the maximum, minimum, and mean values were used to produce triangular fuzzy number of each class in different attributes, and the intersection points of test samples with triangular fuzzy numbers were calculated. Based on this method, it is feasible to produce a classifier which can effectively recognize the known and unknown classes. Therefore, we plan to investigate and improve this method for wood species classification, since the wood species quantity is greatly larger than the class quantity in [43]. On the other hand, in a certainty framework, we have also used a one-class classifier Support Vector Data Description (SVDD) [44], which can preclassify the wood samples into two categories (i.e., the known and unknown wood categories).

Data Availability

Access to research data is restricted. The experimental instruments and wood species samples are owned by other academic institutions; therefore, we cannot publicly release the research spectral and image wood dataset.

Conflicts of Interest

The authors declare no conflicts of interest.

Acknowledgments

This research was supported by the National Natural Science Foundation of China (Grant no. 31670717) and by the Fundamental Research Funds for the Central University (Grant no. 2572017EB09).

References

  1. B. K. Lavine, C. E. Davidson, A. J. Moores, and P. R. Griffiths, “Raman spectroscopy and genetic algorithms for the classification of wood types,” Applied Spectroscopy, vol. 55, no. 8, pp. 960–966, 2001. View at: Publisher Site | Google Scholar
  2. V. Piuri and F. Scotti, “Design of an automatic wood types classification system by using fluorescence spectra,” IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), vol. 40, no. 3, pp. 358–366, 2010. View at: Publisher Site | Google Scholar
  3. A. K. Moore and N. L. Owen, “Infrared spectroscopic studies of solid wood,” Applied Spectroscopy Reviews, vol. 36, no. 1, pp. 65–86, 2001. View at: Publisher Site | Google Scholar
  4. I. Miranda, J. Gominho, S. Ferreira-Dias, and H. Pereira, “Pattern recognition as a tool to discriminate softwood and hardwood bark fractions with different particle size,” Wood Science and Technology, vol. 48, no. 6, pp. 1197–1211, 2014. View at: Publisher Site | Google Scholar
  5. M. Khalid, L. Y. L. Eileen, R. Yusof et al., “Design of an intelligent wood species recognition system,” International Journal of Simulation: Systems, Science and Technology, vol. 9, no. 3, pp. 9–19, 2008. View at: Google Scholar
  6. U. Khairuddin, R. Yusof, M. Khalid et al., “Optimized feature selection for improved tropical wood species recognition system,” ICIC Express Letters, Part B, vol. 2, no. 2, pp. 441–446, 2011. View at: Google Scholar
  7. R. Yusof, M. Khalid, and A. S. Mohd Khairuddin, “Fuzzy logic-based pre-classifier for tropical wood species recognition system,” Machine Vision and Applications, vol. 24, no. 8, pp. 1589–1604, 2013. View at: Publisher Site | Google Scholar
  8. A. R. De Geus, A. R. Backes, A. B. Gontijo et al., “Amazon wood species classification: a comparison between deep learning and pre-designed features,” Wood Science and Technology, vol. 24, 2021. View at: Publisher Site | Google Scholar
  9. A. Fabijánska, M. Danek, and J. Barniak, “Wood species automatic identification from wood core images with a residual convolutional neural network,” Computers and Electronics in Agriculture, vol. 181, Article ID 105941, 2021. View at: Google Scholar
  10. L. R. Schimleck, C. Mora, and R. F. Daniels, “Estimation of the physical wood properties of green Pinus taeda radial samples by near infrared spectroscopy,” Canadian Journal of Forest Research, vol. 33, no. 12, pp. 2297–2305, 2003. View at: Publisher Site | Google Scholar
  11. M. C. Popescu, C. M. Popescu, G. Lisa et al., “Evaluation of morphological and chemical aspects of different wood species by spectroscopy and thermal methods,” Journal of Molecular Structure, vol. 988, no. 1–3, pp. 65–72, 2011. View at: Publisher Site | Google Scholar
  12. L. Schimleck, R. Evans, and J. Ilic, “Application of near infrared spectroscopy to the extracted wood of a diverse range of species,” IAWA Journal, vol. 24, no. 4, pp. 429–438, 2003. View at: Publisher Site | Google Scholar
  13. Z.-H. Jiang, A.-M. Huang, and B. Wang, “Near infrared spectroscopy of wood sections and rapid density prediction,” Spectroscopy and Spectral Analysis, vol. 26, no. 6, pp. 1034–1037, 2006, in Chinese. View at: Google Scholar
  14. Z. Zhou, S. Rahimi, S. Avramidis et al., “Species- and moisture-based sorting of green timber mix with near infrared spectroscopy,” Bioresources, vol. 15, no. 1, pp. 317–330, 2020. View at: Google Scholar
  15. V. T. H. Tham, T. Inagaki, and S. Tsuchikawa, “A new approach based on a combination of capacitance and near-infrared spectroscopy for estimating the moisture content of timber,” Wood Science and Technology, vol. 53, no. 3, pp. 579–599, 2019. View at: Publisher Site | Google Scholar
  16. A. M. Taylor, S. H. Baek, M. K. Jeong et al., “Wood shrinkage prediction using NIR spectroscopy,” Wood and Fiber Science, vol. 40, no. 2, pp. 301–307, 2008. View at: Google Scholar
  17. R. Stirling, T. Trung, C. Breuil et al., “Predicting wood decay and density using NIR spectroscopy,” Wood and Fiber Science, vol. 39, no. 3, pp. 414–423, 2007. View at: Google Scholar
  18. R. Yusof, M. Khalid, and A. S. M. Khairuddin, “Application of kernel-genetic algorithm as nonlinear feature selection in tropical wood species recognition system,” Computers and Electronics in Agriculture, vol. 93, no. 1, pp. 68–77, 2013. View at: Publisher Site | Google Scholar
  19. I. Ibrahim, A. S. M. Khairuddin, M. S. Abu Talip, H. Arof, and R. Yusof, “Tree species recognition system based on macroscopic image analysis,” Wood Science and Technology, vol. 51, no. 2, pp. 431–444, 2017. View at: Publisher Site | Google Scholar
  20. B. Panagiotis, D. Kosmas, B. Loannis et al., “Wood species recognition through multidimensional texture analysis,” Computers and Electronics in Agriculture, vol. 144, no. 2, pp. 241–248, 2018. View at: Publisher Site | Google Scholar
  21. J. A. Roubos, M. Setnes, and J. Abonyi, “Learning fuzzy classification rules from labeled data,” Information Sciences, vol. 150, no. 1-2, pp. 77–93, 2003. View at: Publisher Site | Google Scholar
  22. T. Nakashima, G. Schaefer, Y. Yokota, and H. Ishibuchi, “A weighted fuzzy classifier and its application to image processing tasks,” Fuzzy Sets and Systems, vol. 158, no. 3, pp. 284–294, 2007. View at: Publisher Site | Google Scholar
  23. V. Bombardier, C. Mazaud, P. Lhoste, and R. Vogrig, “Contribution of fuzzy reasoning method to knowledge integration in a defect recognition system,” Computers in Industry, vol. 58, no. 4, pp. 355–366, 2007. View at: Publisher Site | Google Scholar
  24. V. Bombardier, E. Schmitt, and P. Charpentier, “A fuzzy sensor for color matching vision system,” Measurement, vol. 42, no. 2, pp. 189–201, 2009. View at: Publisher Site | Google Scholar
  25. V. Bombardier and E. Schmitt, “Fuzzy rule classifier: capability for generalization in wood color recognition,” Engineering Applications of Artificial Intelligence, vol. 23, no. 6, pp. 978–988, 2010. View at: Publisher Site | Google Scholar
  26. P. Barmpoutis, K. Dimitropoulos, I. Barboutis, N. Grammalidis, and P. Lefakis, “Wood species recognition through multidimensional texture analysis,” Computers and Electronics in Agriculture, vol. 144, pp. 241–248, 2018. View at: Publisher Site | Google Scholar
  27. L. V. D. Maaten and G. Hinton, “Visualizing data using t-SNE,” Journal of Machine Learning Research, vol. 9, pp. 2579–2605, 2008. View at: Google Scholar
  28. P. Smets, “The combination of evidence in the transferable belief model,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 12, no. 5, pp. 447–458, 1990. View at: Publisher Site | Google Scholar
  29. Y. Bengio and Y. Grandvalet, “No unbiased estimator the variance of K-fold cross-validation,” Journal of Machine Learning Research, vol. 5, pp. 1089–1105, 2004. View at: Google Scholar
  30. D. D. Lewis, “Naive (Bayes) at forty: the independence assumption in information retrieval,” Lecture Notes in Computer Science European Conference on Machine Learning, vol. 1398, pp. 4–15, 1998. View at: Google Scholar
  31. M. Belgiu and L. Drăguţ, “Random forest in remote sensing: a review of applications and future directions,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 114, no. 4, pp. 24–31, 2016. View at: Publisher Site | Google Scholar
  32. D. Zipser and R. A. Andersen, “A back-propagation programmed network that simulates response properties of a subset of posterior parietal neurons,” Nature, vol. 331, no. 6158, pp. 679–684, 1988. View at: Publisher Site | Google Scholar
  33. C.-C. Chang and C.-J. Lin, “Libsvm,” ACM Transactions on Intelligent Systems and Technology, vol. 2, no. 3, pp. 1–27, 2011. View at: Publisher Site | Google Scholar
  34. Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998. View at: Publisher Site | Google Scholar
  35. M. I. A. P. Zamri, F. Cordova, A. S. M. Khairuddin, N. Mokhtar, and R. Yusof, “Tree species classification based on image analysis using improved-basic gray level aura matrix,” Computers and Electronics in Agriculture, vol. 124, pp. 227–233, 2016. View at: Publisher Site | Google Scholar
  36. T. Ojala, M. Pietikäinen, and T. Mäenpää, “Multiresolution gray-scale and rotation invariant texture classification with local binary patterns,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 7, pp. 971–987, 2002. View at: Publisher Site | Google Scholar
  37. I. Ibrahim, A. S. M. Khairuddin, H. Arof, R. Yusof, and E. Hanafi, “Statistical feature extraction method for wood species recognition system,” European Journal of Wood and Wood Products, vol. 76, no. 1, pp. 345–356, 2018. View at: Publisher Site | Google Scholar
  38. K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” 2014, https://arxiv.org/abs/1409.1556. View at: Google Scholar
  39. F. N. Iandola, S. Han, M. W. Moskewicz et al., “SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5 MB model size,” 2016, https://arxiv.org/abs/1602.07360. View at: Google Scholar
  40. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, Las Vegas, NV, USA, June 2016. View at: Publisher Site | Google Scholar
  41. C. Szegedy, W. Wei Liu, Y. Yangqing Jia et al., “Going deeper with convolutions,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9, Boston, MA, USA, June 2015. View at: Publisher Site | Google Scholar
  42. A. S. Oktaria, E. Prakasa, and E. Suhartono, “Wood species identification using convolutional neural network (CNN) architectures on macroscopic images,” Journal of Information Technology and Computer Science, vol. 4, no. 3, pp. 274–283, 2019. View at: Publisher Site | Google Scholar
  43. Y. Tang, D. Wu, and Z. Liu, “A new approach for generation of generalized basic probability assignment in the evidence theory,” Pattern Analysis and Applications, 2021. View at: Publisher Site | Google Scholar
  44. P. Zhao, Z.-Y. Li, and Y. Li, “An incremental self-adaptive wood species classification prototype system,” Journal of Spectroscopy, vol. 2019, Article ID 9247386, 11 pages, 2019. View at: Publisher Site | Google Scholar

Copyright © 2021 Peng Zhao et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.


More related articles

 PDF Download Citation Citation
 Download other formatsMore
 Order printed copiesOrder
Views150
Downloads198
Citations

Related articles

Article of the Year Award: Outstanding research contributions of 2020, as selected by our Chief Editors. Read the winning articles.