We propose a hardwood species identification method based on wood hyperspectral microscopic images. A SOC710VP hyperspectral stereomicroscope was used to acquire microscopic images of a hardwood cross section. In these microscopic images, each part’s spectral features are discussed. We found that the spectral divisibility of wood vessels’ peripheral and central regions in the hyperspectral microscopic images can be used for hardwood species recognition. Mathematical morphological operation and K-L divergence were used to extract spectral features at the wood vessels’ peripheral regions and central regions, respectively. By comparing wood vessels’ spectral similarity across wood species samples, we found that wood vessels’ peripheral spectral divisibility is larger than its central. Finally, the spectral information from randomly selected regions of interest (i.e., ROI) and that of wood vessels’ peripheral and central regions have value as a classification basis. In our hardwood species classification experiments, three dimensionality reduction algorithms, principal component analysis (PCA), kernel principal component analysis (KPCA), and multidimensional scaling (MDS), and the three classifiers, BP neural network, support vector machine (SVM), and Mahalanobis distance (MD), are combined to perform hardwood species classification work. Experimental results indicate that the best recognition effect can be achieved at the peripheral region of wood vessels using PCA or MDS with the MD algorithm.
Wood is an important renewable natural resource. Different tree species
With the rapid development of machine vision, it has become possible to identify the processed hardwood species by using intelligent methods. Pan and Kudo [
In recent years, spectral analysis schemes have been proposed for wood species classification [
In this paper, the microscopic hyperspectral images of hardwood cross sections were recorded using a SOC710VP hyperspectral stereomicroscope to investigate the possibility of hardwood species identifications. The spectral features of different wood tissues such as wood vessel, wood ray, and background are discussed. It was found that spectral information from the peripheral and central regions of wood vessels can be used for hardwood species identification.
Six hardwood species were studied (Table
Species studied.
Number | English name | Scientific name |
---|---|---|
1 | Red oak |
|
2 | Ash |
|
3 | Merbau |
|
4 | Golden silk teak |
|
5 | Rose wood |
|
6 | Limba |
|
The intraclass differences of wood samples are caused by the tree’s age, cutting position, and place of growth. For each wood species, our wood blocks are taken from different individual trees at random position in the cross section of the timber to include these intraclass differences. In this way, the reliability of our proposed scheme can be ensured.
The spectral acquisition instrument used for this research was a SOC710VP hyperspectral stereomicroscope, with 128 spectral bands and a spectral range of 372–1038 nm. To improve spectral acquisition, all wood sample surfaces were flat, without burrs, and the spectral acquisition environment was consistent and stable. The dimension of captured microscopic hyperspectral images was
To control costs, we chose the visible/near infrared hyperspectral stereomicroscope as the experimental instrument. Spectral data in the visible region may vary because wood sample color may change with environmental variation. Therefore, we maintained uniform environmental conditions during the spectral experiments and wood sample preservation. In addition, 780–900 nm is the near infrared band, which best characterizes the different wood properties. A stereomicroscope was used to capture wood microscopic images conveniently for each wood sample. The wood slice is not used as its production is complex and time consuming.
The microscopic hyperspectral images of cross sections mainly include wood vessels, wood rays, and axial parenchyma, as shown in Figure
Microscopic hyperspectral images of cross sections of different hardwood species. (a)
Hyperspectral images are 3D cubes which have both spatial information and spectral information, and the spectral curves of different wood tissues for the same hardwood species are different. For example, Figure
Spectral curves for different tissues of
Mathematical morphology is widely used in image processing, which can be used to extract useful image components to enhance ROI and suppress background regions. We denote the original image, background image, and ROI image as
In order to eliminate noise, a corrosion operation is required for the obtained
This cube can be regarded as
The wood vessels’ peripheral region extraction with mathematical morphology: (a) the hyperspectral 56th band image; (b) the binary image cube of
For the binary image
The abovementioned image and spectral processing operations were performed to a ROI in a hyperspectral image of one sample. In this work, the size of ROI is set as
Spectral difference in wood vessels’ peripheral region for six hardwood species.
K-L divergence is widely used in information theory as a signal similarity measuring tool [
Therefore, the difference between two band images of a hyperspectral image cube can be measured using the K-L divergence equation (
Suppose a hyperspectral image cube
In this way, we can calculate the K-L divergence of two band images to obtain a symmetric matrix
Spatial clustering results with spectral information and
As shown in Figure
The clustering process in Figure
To solve these problems, we propose an automatic identification scheme for the wood vessels’ central region. For a ROI
Here
We calculated
Spectral difference in wood vessels’ central region for six hardwood species.
We may use spectral properties of a specific wood region or tissue to guide wood species identification using both interclass and intraclass differences. It is obvious that large interclass differences and small intraclass differences will produce more reliable results. These two differences can be represented by the spectral curve’s interclass and intraclass distances. Therefore, here, we use the divergence matrix to indicate the spectral curve’s interclass and intraclass differences. We define the intraclass divergence matrix for the
Here
Finally, the pattern divisibility rule is defined as follows:
In terms of a specific wood region or tissue, we considered the wood vessels’ peripheral region, central region, and randomly selected ROI. Spectral extractions in wood vessels’ peripheral and central regions have been discussed in Sections
The smoothing procedure for the spectral preprocessing and the SNV calibration procedure are usually required before spectral extraction and pattern recognition as they cancel the noise disturbance and waveform offset. This preprocessing process is illustrated in Figure
Spectral preprocessing procedure: (a) original spectral curve; (b) smoothed spectral curve; (c) spectral curve with SNV processing.
Table
Item | Vessels’ central region | Vessels’ peripheral region | Random ROI |
---|---|---|---|
|
1.6506 | 9.0970 | 6.6387 |
Finally, the time required was compared for spectral extractions in the wood vessels’ peripheral region, central region, and random ROI. The computer configuration used was as follows: CPU Intel Core I7-6700, internal memory 8G, hard disk 1T, display card AMD Radeon R7200 Series. In these three spectral extractions, one hyperspectral image was cut into 30 ROI of
Time requirement comparisons for 3 spectral extractions.
Method | Process | Computing time (s) |
---|---|---|
Vessels’ peripheral region |
|
0.6095 |
|
0.0056 | |
|
0.0008 | |
|
0.6199 | |
|
||
Vessels’ central region |
|
0.7714 |
|
0.2252 | |
Central region selection | 0.0087 | |
|
1.0181 | |
|
||
Random ROI |
|
0.0874 |
Before sending the extracted spectral vectors into classifiers, spectral dimensionality reduction was usually required. In this work, three algorithms, PCA, KPCA, and MDS, were used to reduce dimensionality. In spectral classification, three classifiers, BP neural networks, SVM, and MD, were used for identification comparisons.
Firstly, we used the MD classifier to calculate the MD between a test sample and the training sample sets of the 6 hardwood species. The usual nearest neighbor classification was used to judge the test sample’s wood species. The classification results are illustrated in Figure
Spectral classification comparisons with MD classifier.
The classification accuracy was the calculated CCR, which was the overall classification ratio (%) for the 540 test set samples. This classification accuracy is illustrated in Figures
Identification accuracy of BP network for 25 training times (i.e., 25 networks).
Spectral classification comparisons with BP neural network classifier.
Spectral classification comparisons with SVM classifier.
Best classification results with different classifiers and spectral dimensionality reductions.
Dimensionality reductions | PCA | MDS | KPCA | ||||
---|---|---|---|---|---|---|---|
Classifiers | Spectral region | CCR (%) | Spectral dimension | CCR (%) | Spectral dimension | CCR (%) | Spectral dimension |
BP | Central vessel region | 81.78 | 13 | 80.42 | 13 | 79.4 | 13 |
Random ROI | 97.80 | 9 | 97.57 | 9 | 94.25 | 9 | |
Peripheral vessel region | 98.67 | 7 | 98.14 | 8 | 97.65 | 9 | |
|
|||||||
MD | Central vessel region | 81.29 | 14 | 78.14 | 11 | 66.48 | 12 |
Random ROI | 88.88 | 13 | 91.11 | 11 | 94.44 | 15 | |
Peripheral vessel region | 100 | 8 | 100 | 9 | 98.70 | 7 | |
|
|||||||
SVM | Central vessel region | 83.70 | 15 | 83.33 | 14 | 83.33 | 13 |
Random ROI | 97.78 | 7 | 97.78 | 8 | 94.62 | 12 | |
Peripheral vessel region | 99.07 | 8 | 99.07 | 7 | 98.88 | 7 |
Secondly, we used the BP neural network classifier. Since the threshold and weight for the network initialization were selected randomly, the classification accuracy of every trained network may fluctuate slightly. Figure
Thirdly, the SVM classifier was used. The predictive power of SVM depends on the kernel function used. The radial basis kernel function was usually used. When determining the penalty factor
The best identifications with different classifiers and spectral dimensionality reductions are given in Table
Figures
Concerning the spectral dimensionality reductions, good results were obtained using PCA and KPCA with a low spectral dimension. However, if MDS was used, the CCR stabilized only when the spectral dimensionality was larger than 7. In general, PCA was the best choice in terms of identification accuracy and spectral dimension.
Wood hyperspectral image acquisition may be influenced by external environmental factors such as temperature, humidity, and illumination. In this section, a simulation experiment on noisy wood hyperspectral image classification was performed. Since the training set of wood images can be acquired in advance in a controlled environment, we only added white noises to the test set. The signal-to-noise ratio (SNR) was used to measure the test set’s image quality.
The robustness of our three spectral extraction schemes (i.e. vessels’ central, vessels’ peripheral, random ROI) was compared using an error rate defined in terms of SNR.
Different SNR with error rates: (a) SVM; (b) BP; (c) MD.
In the wood vessels’ peripheral region extraction, the two key parameters are the size
Figure
Wood vessels’ peripheral region extraction with different
Figure
Wood vessels’ peripheral region extraction with different
In summary, we conclude that small
Best values for parameters
Tree |
|
|
---|---|---|
Red oak | 45 | 26 |
Ash wood | 30 | 35 |
Merbau | 43 | 30 |
Golden silk teak | 45 | 30 |
Rose wood | 50 | 45 |
Limba | 20 | 30 |
Vessel peripheral region extraction with best
To test the incremental identification performance of our scheme, two new wood species were added (i.e., Manchurian Ash (
Spectral classification comparisons for 8 wood species.
Finally, the proposed scheme is compared with the two algorithms described in references [
The retained images after waveband selection were defined as
These new wavelet coefficients were then used for inverse wavelet transform to obtain the fused digital image
Schematic of image fusion with wavelet transform. (a) Band
The second aforementioned method consists in applying principal component analysis (PCA) to reduce the dimension of hyperspectral images. The first four principal components are presented in Figure
Results of dimensionality reduction from PCA ((a) PC1, (b) PC2, (c) PC3, and (d) PC4).
The method described in reference [
According to the descriptions in reference [
Result of compression of pixel values ((a) original 8-bit image; (b) compressed 4-bit image).
Then, the feature matrix
To reduce the dimension of the feature vector, the square matrix
Finally, the square matrix
IBGLAM feature vector.
Processing time when applying the methods described in references [
Spectral dimension reduction | Texture feature computation | ||
---|---|---|---|
K-L divergence | PCA | IBGLAM | GLCM |
42.9767 (s) | 0.6414 (s) | 0.1698 (s) | 0.0329 (s) |
Figure
Comparison of proposed method vs. other methods: (a) vessel peripheral spectral extraction method; (b) IBGLAM and GLCM methods.
It can be seen from Figure
We have proposed a novel hardwood species identification scheme using hardwood microscopic hyperspectral images. The spectral feature of wood vessels is especially considered to investigate the possibility of tree identification to the species level. Specifically, the spectral features of vessel peripheral regions are acquired using morphological operations and those of central regions by use of K-L divergence and an automatic identification scheme. The spectral divisibility rules in Section
Our work attempted to identify hardwood trees to species level using spectral analysis. Considering the instrument’s cost, we chose the visible/near infrared SOC710VP hyperspectral imager with 128 spectral bands and a spectral range of 372–1038 nm. Spectral data in the visible region may vary due to changes in wood color produced by environmental variation. Therefore, we kept the environment constant during spectral work and wood sample preservation. In addition, 780–900 nm is the near infrared band, which best characterizes different wood properties.
The method described in reference [
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 wood data.
The authors declare that there are no conflicts of interest.
This research was supported by the National Natural Science Foundation of China (grant no. 31670717), the Fundamental Research Funds for the Central University (grant no. 2572017EB09), and the Heilongjiang Province Natural Science Foundation (grant no. C2016011).