Optimizing recognition models for wood species identification using multi-spectral techniques

2.1 Experimental material

In this study, four species of Pinaceae coniferous wood and one species of broadleaf wood were collected from the Hadamen Township woodland in Hunchun City, Jilin Province, China. The wood samples were harvested from the trunk at a height of approximately 1.3 m, and precisely cut into discs using a precision cutter. A total of 40 samples were obtained from each of the five species, resulting in a comprehensive set of 200 wood chip samples, each measuring 50 mm × 50 mm × 5 mm in thickness. To minimize the impact of moisture on the THz experiment, all wood specimens were naturally dried in the same atmospheric environment until they reached a dry state (25.66 + 1.01 °C).

Figure 1 depicts a representative image of the wood samples used in the experiments. To facilitate data analysis, the collected wood feature data is partitioned into training and test sets using the Kennard-Stone algorithm, which ensures that the most representative samples are selected for model training while minimizing overlap between sets. The dataset was divided with a 7:3 ratio for training and test purposes, respectively. Table 1 displays the breakdown of the dataset results.

Figure 1:

Wood sample of (a) Tilia tuan, (b) Pinus sylvestris, (c) Pinus tabuliformis, (d) Pinus koraiensis, and (e) Larix gmelinii.

2.2 Experimental instruments

In this study, three spectroscopic instruments were employed to obtain spectral data from wood samples.

Firstly, an MPA II NIR spectrometer (Bruker, Germany) was used to acquire NIR spectra of wood samples, as illustrated in Figure 2. The instrument operates over a wavelength range of 800–2,800 nm and is capable of performing 32 scans per measurement. All measurements were conducted in diffuse reflectance mode, with each locus subjected to 32 automated scans, taking approximately 10 s per scan. Three parallel experiments were carried out for each specimen, and the spectra obtained were averaged to mitigate random errors.

Figure 2:

NIR instrument: (a) NIR system, and (b) the schematic diagram of NIR system.

In addition, a SOC710VP hyperspectral imaging system was employed for hyperspectral data acquisition. The system consists of an imaging spectrometer, CCD imaging lens, halogen lights, and an upper computer for data processing (Figure 3). This setup enables fixed-point, long-term continuous measurement and offers functionalities such as visual focusing and rapid imaging, which are highly suitable for microscopic and biochemical investigation.

Figure 3:

Hyperspectral instrument: (a) hyperspectral imaging system, and (b) the schematic diagram of hyperspectral imaging system.

Furthermore, THz-TDS was conducted using the TERA K15 equipment (Menlo Systems, Germany). This system was selected due to its superior performance in capturing THz spectral information for thin samples, such as the wood slices used in this study, which had dimensions of 50 mm × 50 mm × 5 mm. The transmission THz-TDS system, illustrated in Figure 4, provides high-resolution spectra by transmitting THz radiation through the samples.

Figure 4:

THz-TDS instrument: (a) THz-TDS system, and (b) the schematic diagram of THz-TDS system.

The instrument employed in this study offers a spectral resolution of 0.011 THz, a dynamic range of 70 dB, and a scanning range of 300 ps. Additional specifications include a pulse duration of less than 90 fs and an average output power exceeding 60 mW. The system operates within a spectral range of 0–3.5 THz. Before collecting sample data, reference spectra were acquired using dry air as the medium. The time-domain signal of each wood sample was subsequently transformed into the frequency domain through a Fast Fourier Transform (FFT) for analysis. Data were collected at three distinct locations – the top, middle, and bottom of each sample. The final spectrum for each specimen was determined by averaging these data points. THz absorption coefficients were calculated by employing the optical parameter model proposed by Dorney et al. (2001), based on the characteristics of the THz frequency domain spectrum.

2.3 Spectral pre-processing

Spectral data are often affected by various noise sources, including baseline fluctuations, electrical noise, and environmental disturbances, which can compromise the reliability of subsequent analysis. To mitigate these issues and enhance the identification and quantification of spectral features, data preprocessing techniques such as smoothing, baseline correction, and noise filtering are commonly employed. In this study, the Standard Normal Variate (SNV) transformation was applied to minimize the effects of scattering and spectral noise, thus improving the quality and interpretability of the spectra.

The SNV transformation is a commonly used method for spectral correction, particularly effective in mitigating scattering effects and baseline deviations. In this study, the SNV transform was applied to correct the spectral data. The process involves calculating the mean value of each individual spectrum, subtracting it from the spectral data, and then dividing the result by the standard deviation of the spectrum. This normalization adjusts the spectrum by minimizing variations due to scattering and other interferences.

A feature selection technique is essential for optimizing the selection of spectral bands, as raw spectra contain redundant information across numerous spectral characteristics. In this study, the NIR spectra, due to their high dimensionality and numerous bands, were processed using a combination of UVE and CARS algorithms. These methods have proven effective in selecting the most informative bands, reducing data redundancy, and improving model performance. The number of bands in HSI information data is also large, and the redundant features are obvious, using CARS can effectively improve the ability of high-dimensional data processing, redundant feature removal and classification performance enhancement. For THz spectra, where the number of bands is typically smaller, the SPA was employed for feature selection. This decision is based on previous findings (Wang et al. 2019), which achieved recognition accuracy of 96 % in wood identification after applying SPA to screen THz spectral data. Therefore, SPA is particularly well-suited for the THz spectrum in this study, facilitating efficient feature extraction and enhancing the model’s predictive accuracy.

The CARS algorithm is a feature selection technique that integrates partial least squares (PLS) techniques with Monte Carlo sampling, effectively optimizing the identification of relevant spectral features (Li et al. 2009). In CARS, adaptive reweighted sampling (ARS) is employed to iteratively select a subset of points based on the regression coefficients derived from the PLS model. Specifically, the absolute values of these coefficients guide the selection process, where points with higher weights are retained, and those with lower weights are progressively eliminated. This adaptive process continues over multiple iterations, each constructing a new PLS model on the remaining subset of points. At each iteration, the wavelength with the smallest cross-validated root-mean-square error (RMSE) is identified and chosen as the feature wavelength.

The UVE algorithm is a variable selection method based on PLS analysis. It uses the intrinsic correlations within the sample data to remove the uninformative variables based on the magnitude of their contribution value, resulting in a reduction in the degree of data redundancy. UVE works by randomly partitioning the original training dataset into sub-datasets through cross-validation. Uninformative features were systematically removed by assessing the contribution value of each variable, leaving only the most significant variables for further analysis.

Additionally, the SPA is employed as a forward variable selection technique. SPA operates by projecting each wavelength onto other wavelengths and comparing the magnitude of the resulting projection vectors. Through this iterative process, SPA identifies the final set of feature wavelengths based on their relevance to the model, ensuring that the selected variables possess the least covariance and redundancy. This approach effectively enhances the precision of spectral analysis by isolating the most informative wavelengths, improving the model’s interpretability and predictive power.

2.4 Recognition model

The optimal recognition models for the NIR spectrum, HSI spectrum, and THz spectrum were identified through a comparative analysis of the accuracies of five different wood species in both training and test sets, utilizing five commonly employed models: THz: SPA + RF, HSI: CARS + LSTM, and NIR: UVE + CARS + ELM.

ELM is characterized as a single hidden layer feedforward neural network characterized that employs randomly chosen input layer parameters. This model minimizes the loss function to determine the connection weights between the hidden and output layers, based on Moore–Penrose theory. Owing to its reduced number of training parameters, ELM facilitates rapid learning, exhibits a lower tendency for overfitting, and demonstrates enhanced generalization capabilities.

LSTM networks represent a specialized variant of Recurrent Neural Network (RNN) that incorporate gating units to mitigate the issues of gradient vanishing and explosion, which are commonly encountered in traditional RNNs. Consequently, LSTM are well-suited for effectively processing significant events that occur over extended time intervals and temporal delays in time series data.

RF is an ensemble classifier that utilizes multiple decision trees, with the mode of the classifications produced by each tree determining the final output categories. This algorithm was developed by Leo Breiman and Adele Cutler, who also trademarked the term “Random Forests.” In addition to facilitating dimensionality reduction, Random Forests are applicable to both classification and regression problems. They outperform individual decision trees in prediction and classification tasks and exhibit strong resilience to noise and outliers.

3.1 Analysis of spectra

The analysis was conducted over a wavelength range from 800 to 2,787 nm, with wavelength plotted on the horizontal axis and reflectance on the vertical axis. Figure 5 presents a comparison of the NIR spectra before and after SNV processing. The spectral curves of the five wood species exhibit significant overlap in the 800–1,400 nm region, as demonstrated by the original spectra in Figure 5, which complicates the identification. In contrast, the SNV-treated spectra reveal a more distinct separation among the five curves in the regions of 1,500–1,600 nm, 2,300–2,400 nm, and 2,500–2,700 nm, compared to the other spectral bands.

Figure 5:

NIR absorption coefficient of wood samples: (a) the original spectra, (b) the spectra pre-processed by SNV.

The comparison of HSI between the original and SNV-treated spectra is illustrated in Figure 6, where reflectance is plotted on the vertical axis and wavelength ranges from 370 to 1,035 nm on the horizontal axis. The spectral curves of the five wood species show considerable similarity in the 400–1,000 nm range, exhibiting an increasing trend in the 400–850 nm band. Beyond 850 nm, the spectral curves gradually flatten or even decline, with no noticeable absorption peaks observed. This suggests that the material composition of the five wood species is approximately similar. Notably, the spectral curves of Larix gmelinii are more distinctive compare to those of the other four species.

Figure 6:

HSI spectral information reflectance spectra of wood samples: (a) the original spectra, (b) the spectra pre-processed by SNV.

Figure 7 compares the THz absorption coefficient spectra of the raw and SNV-treated samples, with frequency plotted on the horizontal axis and absorption coefficient on the vertical axis. The THz absorption coefficient spectra of the five wood species do not exhibit significant overlap in the 0.2–1 THz range. Notably, the absorption coefficient spectrum of L. gmelinii displays a distinct position for its absorption peak relative to the other four species. Additionally, the five species demonstrate markedly in the intensity of their absorption coefficients, ranked as follows: Tilia tuan Szyszyl, L. gmelinii, Pinus koraiensis, Pinus tabuliformis, and Pinus sylvestris.

Figure 7:

THz absorption coefficient of wood samples: (a) the original spectra, (b) the spectra pre-processed by SNV.

3.2 Feature selection

To achieve optimal recognition results, this study employs multiple algorithms for the selection of distinct features based on three spectral types: THz, HSI, and NIR.

The CARS algorithm was employed for secondary dimensionality reduction, successfully reducing the dimensionality of feature variable to 51, which accounts for 4.42 % of the total feature bands. This final accuracy surpasses that obtained by independently applying the CARS and UVE algorithms. Initially, the NIR spectrum was processed using the UVE algorithm. However, due to the high dimensionality of the NIR bands, 577 features remained after selection, complicating subsequent data processing.

Owing to the benefits of the CARS algorithm in screening features from HSI data (Masia et al. 2015), this method was similarly for feature selection within the HSI spectrum. Following the screening process, 39 feature bands – representing 30.47 % of the total – were identified, with the majority concentrated between 800 and 1,000 nm in wavelength.

For THz frequency bands, the SPA method was utilized due to its exceptional performance in THz frequency selection (Wang et al. 2019). After applying the SPA algorithm, 15 feature bands – accounting for 9.20 % of the total – were identified. These bands were primarily distributed in the 0.2–0.6 THz range.

3.3 Analysis of model results

To develop the optimal recognition model for each of the three feature-extracted spectra, the most suitable methods were selected and integrated with the corresponding feature extraction algorithms.

3.3.1 Recognition models for NIR spectra

Five widely recognized models – ELM, LSTM, RF, CNN, and SVM – were applied to the NIR spectral data obtained through the UVE + CARS method, as presented in the previous section. This analysis aimed to determine the model that provides the best recognition performance. Among the five wood species, T. tuan Szyszyl and L. gmelinii are generally more easily identifiable, as shown in Table 2. The average recognition accuracy of the SVM model across the five models was only 71.4 %, which can be attributed to significant spectral overlap among the wood types. In contrast, the ELM model demonstrates superior recognition performance, achieving an accuracy of 100.0 % in the training set and 96.7 % in the test set. CNN, LSTM, and ELM all exhibited high accuracy, exceeding 90 %. Figure 8 illustrates the recognition performance of the five wood species following the application of the ELM model, which resulted in a recognition accuracy of 96.7 %. According to the results, only two of the 60 samples in the test set exhibited misclassification.

Table 2:

Comparison of five recognition models for five wood species in NIR spectroscopy.

Wood species	Recognition models	Training set (%)	Testing set (%)
Tilia tuan Szyszyl	ELM	100.0	100.0
	SVM	83.9	78.2
	LSTM	100.0	99.0
	RF	95.7	88.2
	CNN	100.0	100.0
Pinus sylvestris	ELM	100.0	98.4
	SVM	80.8	73.9
	LSTM	100.0	100.0
	RF	93.9	81.4
	CNN	100.0	100.0
Pinus tabuliformis	ELM	100.0	100.0
	SVM	75.5	76.0
	LSTM	100.0	91.9
	RF	98.1	75.8
	CNN	100.0	92.9
Pinus koraiensis	ELM	100.0	97.1
	SVM	68.7	59.8
	LSTM	100.0	94.4
	RF	95.5	82.6
	CNN	100.0	91.3
Larix gmelinii	ELM	100.0	94.5
	SVM	85.2	69.0
	LSTM	100.0	96.8
	RF	89.6	80.2
	CNN	100.0	92.5

Figure 8:

Recognition results of ELM for five wood species in the NIR spectrum.

3.3.2 Recognition models for HSI information

Five widely utilized recognition models – ELM, LSTM, RF, CNN, and SVM – were applied to the hyperspectral imaging (HSI) data acquired through the CARS method to achieve optimal recognition results. Among the five wood species, T. tuan Szyszyl and L. gmelinii are comparatively easier to identify, as indicated in Table 3. SVM performs poorly in both the hyperspectral and near-infrared spectra, probably because the hyperspectral and near-infrared spectra share some common bands. The CNN, LSTM, and RF models demonstrate the highest recognition accuracy. The training set accuracy stabilizes at 100 % after 600 iterations with an LSTM layer of six and a maximum of 1,000 iterations, while the test set accuracy stabilizes at 96.7 % after 1,000 iterations. The recognition performance of the five wood species following the LSTM model is displayed in Figure 9, with only two of the 60 samples in the test set identified as misclassified.

Table 3:

Comparison of five recognition models for five wood species with HSI information.

Wood species	Recognition models	Training set (%)	Testing set (%)
Tilia tuan Szyszyl	ELM	100.0	98.4
	SVM	87.8	84.6
	LSTM	100.0	100.0
	RF	100.0	100.0
	CNN	100.0	100.0
Pinus sylvestris	ELM	100.0	88.3
	SVM	90.6	82.1
	LSTM	100.0	96.0
	RF	100.0	90.8
	CNN	100.0	97.8
Pinus tabuliformis	ELM	100.0	84.5
	SVM	89.1	84.4
	LSTM	100.0	92.8
	RF	98.5	94.8
	CNN	100.0	92.7
Pinus koraiensis	ELM	93.9	87.1
	SVM	92.7	79.8
	LSTM	100.0	100.0
	RF	100.0	100.0
	CNN	100.0	98.9
Larix gmelinii	ELM	92.6	96.8
	SVM	83.1	91.0
	LSTM	100.0	100.0
	RF	100.0	97.1
	CNN	98.5	100.0

Figure 9:

Recognition results of LSTM for five wood species with HSI information.

3.3.3 Recognition models for THz spectrum

To achieve optimal recognition results, five widely utilized models – ELM, LSTM, RF, CNN, and SVM – were applied to the THz spectral data obtained through the SPA described in the previous section. Table 4 illustrates that the THz spectrum exhibits superior recognition accuracy compared to both the NIR and HSI spectral types. Furthermore, no discernible variation in recognition performance was observed among the five wood species, as all models achieved accuracy levels exceeding 90 %. These findings suggest that the THz spectrum is more readily recognized than both the NIR spectrum and the hyperspectral data. With 50 decision trees, the error decreases to 0.02 when the number of trees reaches 35. The RF model demonstrated the highest recognition accuracy, achieving 98.3 % in the test set and 100 % in the training set. The recognition performance of the five wood types following RF model application is depicted in Figure 10. The results indicate that, out of the 60 test set samples, only one sample exhibited misclassification.

Table 4:

Comparison of five recognition models for five species of wood in the THz spectrum.

Wood species	Recognition models	Training set (%)	Testing set (%)
Tilia tuan Szyszyl	ELM	100.0	94.5
	SVM	99.3	96.3
	LSTM	99.3	97.2
	RF	100.0	100.0
	CNN	100.0	96.0
Pinus sylvestris	ELM	98.8	92.1
	SVM	100.0	95.1
	LSTM	98.9	94.0
	RF	100.0	98.8
	CNN	100.0	100.0
Pinus tabuliformis	ELM	97.9	94.3
	SVM	98.4	94.9
	LSTM	97.0	91.8
	RF	100.0	97.8
	CNN	100.0	94.7
Pinus koraiensis	ELM	100.0	93.7
	SVM	99.1	98.8
	LSTM	99.0	95.8
	RF	100.0	100.0
	CNN	100.0	98.5
Larix gmelinii	ELM	100.0	100.0
	SVM	97.7	93.0
	LSTM	98.5	90.6
	RF	100.0	95.4
	CNN	100.0	95.0

Figure 10:

Recognition results of RF for five wood species in the THz spectrum.

Following a comparison of five models – CNN, SVM, LSTM, RF, and ELM – the findings indicated that L. gmelinii and T. tuan Szyszyl were more readily identifiable. The feature-selected ELM model for the NIR spectral data achieved 100 % identification accuracy on the training set and 96.7 % on the test set. Similarly, the LSTM model demonstrated robust performance in the analysis of hyperspectral data, maintaining a constant accuracy of 100 % in the training set and 96.7 % in the test set. In contrast, all models achieved identification rates exceeding 90 % when applied to THz spectrum data, with the RF model showing the highest accuracy at 100 % in the training set and 98.3 % in the test set. Overall, the results suggest that THz spectra perform more effectively in wood recognition compared to both NIR and HSI data.