An Interpretable Approach to Wood Species Identification Based on Anatomical Features in Microscopic Images

Abstract

Wood recognition plays a vital role in the trade and conservation of rare wood species. However, the computer vision-based methods classify the wood species by the features that are not used within the framework of wood anatomy, leading to results that are not interpretable. This study proposes a novel wood recognition method that detects anatomical structures such as vessels, wood rays, and parenchyma in wood microscopic images. These structures are quantified and mapped to the International Association of Wood Anatomists (IAWA) features, which are then used for species classification. Experimental results on 32 wood species demonstrate the effectiveness of the approach, achieving an accuracy of 94.1%, precision of 92.6%, recall of 93.3%, and an F1-score of 92.7%. In addition to its recognition performance, the method may offer interpretable IAWA-based classification criteria in wood science. These findings suggest that the method could serve as an anatomically interpretable framework for wood species identification, contributing to the regulation of the rare timber trade and supporting the conservation of endangered tree species.

1. Introduction

Illegal logging is one of the most economically significant natural resource crimes worldwide, with approximately 15% to 30% of global timber being sourced through illicit channels each year [1]. The illegal wood trade has caused declines in tree populations, pushed some species to the brink of endangerment, and severely disrupted the balance of forest ecosystems [2,3]. To combat illegal logging, the international community has established a series of protective measures, including the Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES) [4], the European Union Timber Regulation (EUTR) [5] enacted in 2013, and the United States Lacey Act. The effective implementation of these measures relies heavily on reliable wood identification technologies.

Traditional manual wood identification is a complex process. At the macroscopic level, wood species can be identified by observing physical characteristics of the wood surface, such as color, texture, and luster. At the microscopic level, experts typically examine anatomical features in cross, radial, and tangential sections—such as vessels, wood rays, and parenchyma—to perform wood species identification. To standardize and guide this process, the International Association of Wood Anatomists (IAWA) has published systematic lists of microscopic wood identification features [6,7,8], which is well-established and widely recognized as an important reference for wood identification. However, identification based on anatomical observation remains a challenging task. Experts require specialized knowledge in wood anatomy, and quantitatively analyzing some anatomical features is time-consuming and inefficient, making this approach unsuitable for large-scale inspections that involve the analysis of numerous wood samples [9]. In recent years, computer vision-based classification methods have rapidly gained popularity across various fields. To address the low efficiency of traditional manual wood identification, researchers have explored computer vision approaches to automate the recognition process [10,11,12], an overview of representative studies on wood classification is provided in Table 1.

Machine learning and deep learning are two commonly used approaches in computer vision. Among them, machine learning-based wood species recognition methods typically treat wood image identification as a texture recognition task, classifying species by extracting texture features from images. For example, Martins [13] employed local binary pattern (LBP) texture features to classify 112 wood species, achieving an accuracy of 86%. Filho applied Gabor features to classify 41 Brazilian tree species, reaching an accuracy of 97.77% [14]. Rosa da Silva utilized local phase quantization (LPQ) texture features to classify 77 wood species, achieving an accuracy of 88% [15]. Several other studies have also demonstrated that machine learning methods can improve the efficiency of wood identification [16,17,18]. However, most machine learning approaches rely on manually selected feature types. Specifically, Martins, Filho, and Rosa da Silva extracted LBP, Gabor, and LPQ features respectively, which perform well on specific datasets but lack generalizability.

In recent years, deep learning has made significant advances in image recognition and has gradually become the main approach for wood species identification. Among deep learning architectures, convolutional neural networks (CNNs) are the most mature and widely used, owing to their superior capability in image feature extraction. CNNs have been extensively applied to wood image recognition tasks [19,20,21]. Unlike traditional machine learning methods that rely on shallow features such as texture or shape, CNNs are capable of learning abstract and hierarchical features from images, thereby more effectively capturing the subtle anatomical differences among species and significantly improving classification accuracy and generalization performance [22,23].

The above literature review indicates that computer vision-based methods for wood species identification have significantly improved efficiency while maintaining high accuracy. However, according to the IAWA feature lists, key classification features in wood anatomy are primarily related to anatomical structures such as vessels, rays, and parenchyma observed in wood section images [24]. Whether based on texture features in machine learning or high-dimensional abstract features in deep learning, the features utilized in those methods do not align with those employed by wood anatomy experts. As a result, the recognition outcomes often lack interpretability and credibility from a wood science perspective [25], which limits their applicability in professional wood identification contexts [26].

This study proposes an interpretable wood species recognition approach that detects anatomical structures in wood microscopic images, then measures these structures, and maps the measurements to IAWA features. Due to variations in shooting angles, cutting positions, and magnification levels, a single image typically contains limited IAWA features, making it insufficient for classification [27]. Therefore, our method classifies wood species using IAWA features from multiple anatomical images. Finally, a Bayesian fusion network is utilized for wood species classification. Different from traditional methods, our approach not only identifies wood species but also provides interpretable evidence in wood science. The contributions of this study include:

1.

An interpretable wood species classification method is proposed, which leverages IAWA features as the basis for species identification.

2.

Automatic quantification of IAWA features helps reduce the manual effort traditionally required by wood anatomists.

3.

The study demonstrates the importance of integrating multiple anatomical views, showing that feature fusion from different perspectives can improve classification performance.

Table 1. Summary of prior studies on wood classification, detailing the image types used, feature extraction methods, classification algorithms, dataset scales (in terms of number of images and species), and achieved recognition rates.

References	Image Type	FXM	Classifiers	Images/Species	Rec. Rate (%)
Martins et al. [13]	Micro	LBP	SVM	2240/112	98.6
Filho et al. [14]	Micro	Gabor	SVM	2942/41	97.77
Rosa da Silva et al. [15]	Micro	LPQ+LBP	KNN	1221/77	88
Tou et al. [21]	Macro	CNN features	ResNet-50	10,787/25	91.8
Yang et al. [28]	Micro	LBP+GLCM	BPNN+SVM	3000/19	97.54
Zhao et al. [29]	Macro	GLCM	BPNN	1000/5	90
Zamri et al. [30]	Macro	GLCM	BPNN	5200/52	97.01
Yuliastuti et al. [31]	Macro	Gabor	MLP	400/10	95
Nasirzadeh et al. [32]	Macro	LBP	KNN	3700/37	96.60
Yusof et al. [33]	Macro	GLCM	MLBP	3000/30	95.44

Abbreviations: FXM—feature extraction methods; LBP—Local Binary Pattern; Gabor—Gabor Filter; LPQ—Local Phase Quantization; CNN—Convolutional Neural Network; GLCM—Gray-Level Co-occurrence Matrix; SVM—Support Vector Machine; BPNN—Backpropagation Neural Network; MLP—Multilayer Perceptron; KNN—k-Nearest Neighbors; MLBP—Multi-layer neural network based on Back Propagation.

Table 1. Summary of prior studies on wood classification, detailing the image types used, feature extraction methods, classification algorithms, dataset scales (in terms of number of images and species), and achieved recognition rates.

References	Image Type	FXM	Classifiers	Images/Species	Rec. Rate (%)
Martins et al. [13]	Micro	LBP	SVM	2240/112	98.6
Filho et al. [14]	Micro	Gabor	SVM	2942/41	97.77
Rosa da Silva et al. [15]	Micro	LPQ+LBP	KNN	1221/77	88
Tou et al. [21]	Macro	CNN features	ResNet-50	10,787/25	91.8
Yang et al. [28]	Micro	LBP+GLCM	BPNN+SVM	3000/19	97.54
Zhao et al. [29]	Macro	GLCM	BPNN	1000/5	90
Zamri et al. [30]	Macro	GLCM	BPNN	5200/52	97.01
Yuliastuti et al. [31]	Macro	Gabor	MLP	400/10	95
Nasirzadeh et al. [32]	Macro	LBP	KNN	3700/37	96.60
Yusof et al. [33]	Macro	GLCM	MLBP	3000/30	95.44

Abbreviations: FXM—feature extraction methods; LBP—Local Binary Pattern; Gabor—Gabor Filter; LPQ—Local Phase Quantization; CNN—Convolutional Neural Network; GLCM—Gray-Level Co-occurrence Matrix; SVM—Support Vector Machine; BPNN—Backpropagation Neural Network; MLP—Multilayer Perceptron; KNN—k-Nearest Neighbors; MLBP—Multi-layer neural network based on Back Propagation.

2. Materials

The wood samples used in the study are acquired from the Inside Wood website [34] and the Wood Specimen Museum at Southwest Forestry University (SWFU). The data from Inside Wood website is used in accordance with Inside Wood’s copyright policy. These samples come from 32 wood species(see

Table 2. Wood species information in the dataset. The information includes Family, Species, WOCI (IAWA code 192: Wood of Commercial Importance) status, CITES listing, and Geographic Origin.

ID	Family	Species	WOCI	CITES Listed	Geographic Origin
1	MELIACEAE	Swietenia macrophylla	✓	✓	(183, 184)
2	ANACARDIACEAE	Pistacia chinensis			(164, 167, 171, 173)
3	MELIACEAE	Entandrophragma angolense	✓		(178, 179)
4	EBENACEAE	Cyrilla racemiflora			(182)
5	EBENACEAE	Diospyros crassiflora	✓	✓	(178, 179)
6	EBENACEAE	Diospyros virginiana	✓		(184)
7	FAGACEAE	Quercus alba	✓		(184)
8	JUGLANDACEAE	Juglans nigra	✓		(182)
9	LEGUMINOSAE CAESALPINIOIDEAE	Gleditsia triacanthos	✓		(182)
10	LEGUMINOSAE CAESALPINIOIDEAE	Albizia ferruginea	✓		(178, 179)
11	LEGUMINOSAE CAESALPINIOIDEAE	Piptadeniastrum africanum	✓		(178, 179)
12	LEGUMINOSAE DETARIOIDEAE	Intsia bijuga	✓		(168, 169)
13	LEGUMINOSAE PAPILIONOIDEAE	Pericopsis elata		✓	(178, 179)
14	LEGUMINOSAE PAPILIONOIDEAE	Pterocarpus soyauxii	✓	✓	(178, 179)
15	LEGUMINOSAE PAPILIONOIDEAE	Robinia pseudoacacia	✓		(182)
16	MALVACEAE HELICTEROIDEAE	Triplochiton scleroxylon	✓		(178, 179)
17	MELIACEAE	Guarea cedrata	✓		(178, 179)
18	MELIACEAE	Guarea laurentii	✓		(178, 179)
19	MELIACEAE	Lovoa trichilioides	✓		(178, 179)
20	MELIACEAE	Melia azedarach	✓		(164, 166)
21	MORACEAE	Milicia excelsa	✓		(178, 179)
22	RUBIACEAE	Nauclea diderrichii	✓		(178, 179)
23	SAPINDACEAE	Acer rubrum			(182)
24	FABACEAE	Dalbergia cochinchinensis	✓	✓	(171, 172)
25	FABACEAE	Dalbergia stevensonii		✓	(184)
26	FABACEAE	Dalbergia cultrata	✓	✓	(171, 172)
27	FABACEAE	Pterocarpus erinaceus	✓	✓	(178, 179)
28	MELIACEAE	Entandrophragma cylindricum	✓		(178, 179)
29	FABACEAE	Pterocarpus macrocarpus	✓		(171, 172)
30	FABACEAE	Dalbergia oliveri		✓	(171, 172)
31	FABACEAE	Pterocarpus angolensis	✓	✓	(178, 179)
32	FABACEAE	Dalbergia frutescens	✓	✓	(183)

Geographic origin(IAWA Code): 164 Europe and temperate Asia; 166 Mediterranean including Northern Africa and Middle East; 167 Temperate Asia (China), Japan, Russia; 168 Central South Asia; 169 India, Pakistan, Sri Lanka; 171 Southeast Asia and Pacific; 172 Indochina (Thailand, Laos, Vietnam, Cambodia); 173 Indomalesia (Indonesia, Philippines, Malaysia, Brunei, Papua New Guinea, Solomon Islands); 178 Tropical mainland Africa and adjacent islands; 179 Tropical Africa; 181 Southern Africa (south of the Tropic of Capricorn); 182 North America, north of Mexico; 183 Neotropics and temperate Brazil; 184 Mexico and Central America; 186 Tropical South America. Only populations of Diospyros crassiflora from Madagascar are listed under CITES II, while those from continental Africa are not listed.

). Most of the species are of commercial importance in the global timber trade, and several are listed under the CITES appendices. In addition, the geographic origins of the majority of these species span three major regions—Africa, Asia, and the Americas.

The images are acquired from the cross section and the tangential section. The cross section image is acquired perpendicular to the growth rings, which contains vessels and parenchyma. The tangential section image is acquired parallel to the growth rings, which contains wood rays. The positions of cross section and tangential section in a wood log are as shown in Figure 1.

The workflow of wood microscopic image acquisition is shown in Figure 2. The microscopy imaging is conducted under consistent illumination conditions using brightfield reflected light with an LED light source to ensure clear anatomical structures. Prior to imaging, standard wood anatomical sample preparation procedures are followed, including specimen softening, slicing, staining, and Mounting. A mild safranin stain is applied to enhance the contrast of the anatomical structures. Overlapping acquisition regions are avoided during imaging to ensure non-redundant coverage. The microscope used for image acquisition is a Leica DM2000 LED, manufactured by Leica Microsystems in Wetzlar, Germany. equipped with objective lenses of 5×, 10×, and 20× magnifications. Depending on the size and clarity of the anatomical structures, images were captured at these three levels, corresponding to scale bars of 500 μm, 200 μm, and 100 μm, respectively.

Figure 3 presents representative images from the dataset, where subfigures (a)–(e) correspond to cross section images, and (f)–(i) correspond to tangential section images. The image has a resolution of 1205 × 1536 pixels and includes scale information, which provides a reference for converting pixel-based measurements into real-world units and is crucial for accurately measuring the size and quantity of anatomical structures.

A total of 3926 images are collected in this study, comprising 2104 cross section images and 1822 tangential section images, yielding a tangential-to-cross section ratio of approximately 1:1.15. For each of the 32 wood species, around 122 images are acquired to ensure class balance. All anatomical features in the images are annotated with polygonal masks using LabelMe software. The data are split into training and testing sets at a ratio of 8:2 according to section type within each wood species.

3. Methods

The proposed wood recognition method consists of three steps, as shown in Figure 4. First, anatomical feature detection models are employed to detect the contours of anatomical structures, such as vessels, wood rays, and parenchyma. These detection models were trained using the YOLOv8-seg [35] architecture. Second, a series of measurement methods are employed to quantify density, diameter, arrangement, etc., and map them to the IAWA feature codes according to the IAWA principle. Finally, a Bayesian network is adopted to classify wood species using the fused IAWA feature vector from multiple anatomical images.

The method outputs three parts, the first two parts are classification evidence, and the last is the classification result. The first output consists of annotated images, in which vessels, wood rays, and parenchyma are labeled with distinct colors. These color-labeled images serve as visual evidence to support the classification results and facilitate the interpretation of the detected IAWA features. The second output is a list of IAWA codes representing the anatomical features extracted from the images, providing interpretable classification evidence in wood science. The final output is the predicted wood species, derived from the detected IAWA features.

3.1. Anatomy Structure Detection

Anatomy structure detector (ASD) is a fine-tuned YOLOv8-seg model, which was trained on the wood section images dataset. The backbone of the ASD is a feature pyramid, which is used to extract the anatomical structure feature at different scales. The neck of ASD is a fusion module, which combines the features of each scale to enhance the structural feature. The heads of the ASD are used to predict the masks of anatomical structures, such as vessels, rays, and parenchyma. The top header is used to detect small targets, and the bottom is used to detect big targets.

The input of ASD is anatomy images, including cross section images and tangential section images, and the output of ASD is the label, position, and boundary of the anatomy structure. Assuming the image list is $Img=(i_1,i_2,\dots ,i_3)$, the output label is L, the output mask is M, and the output boundary is B, then the processing of the ASD can be represented as Equation (1).

$$L,B,M=\mathit{ASD}\left(\mathit{Img}\right)$$

(1)

where B is the bounding box point list, M is a binary image matrix, and L is the type of anatomy structure.

3.2. Auto Measurement

IAWA features are important in the wood identification task. In our experiments, features related to the material and suitable for automatic computation were selected. The selected features are shown in

Table 3. Selected IAWA feature codes and the descriptions.

IAWA Code	Description
Vessel arrangement and groupings
6	Vessels in longitudinal bands
9	Vessels exclusively solitary (90% or more)
10	Vessels in radial multiples of 4 or more common
11	Vessel clusters common
Tangential diameter of vessel lumina
40	<=50 μm
41	50–100 μm
42	100–200 μm
43	>=200 μm
Vessels per square millimetre
46	<=5 vessels per square millimetre
47	5–20 vessels per square millimetre
48	20–40 vessels per square millimetre
49	40–100 vessels per square millimetre
50	>=100 vessels per square millimetre
Paratracheal axial parenchyma
79	Axial parenchyma vasicentric
80	Axial parenchyma aliform
82	Axial parenchyma winged-aliform
83	Axial parenchyma confluent
85	Axial parenchyma bands more than three cells wide
Ray width
96	Rays exclusively uniseriate
97	Ray width 1 to 3 cells
98	Larger rays commonly 4- to 10 seriate
99	Larger rays commonly > 10-seriate
Rays per millimetre
114	<=4/mm
115	4–12/mm
116	>=12/mm

, in which, the first column is the IAWA codes and the second column is the description of the IAWA codes. The measurement module calculates the features of each anatomical structure and maps the IAWA codes. The selected IAWA features contain 25 features related to vessel arrangement, vessel diameter, vessel density, parenchyma type, ray width, and ray density.

To determine vessel arrangement and grouping features (IAWA codes 6, 9, 10, and 11), a clustering-based approach is employed. As illustrated in Algorithm 1, the center points of the detected vessels are clustered using the DBSCAN algorithm [36]. For each cluster, a linear regression is performed on the center points to estimate the main orientation. The slope of the fitted line is used to identify the vessel arrangement: near-horizontal slopes indicate tangential arrangement (feature 6), whereas near-vertical slopes indicate radial arrangement (feature 10). Feature 11 is determined by evaluating whether the vessels in a cluster exhibit both tangential and radial dispersion, quantified by standard deviations in the horizontal and vertical directions. To assess feature 9, the method checks for overlapping vessels within each cluster. If over 90% of the vessels are not involved in any overlaps, the feature is considered present.

Algorithm 1 Determination of IAWA Features 6, 9, 10, and 11

Require:   Vessel center points $P=\{p_1,p_2,\dots ,p_n\}$ 1: Apply DBSCAN clustering on P, obtain clusters $\{C_1,C_2,\dots ,C_k\}$ 2: for each cluster $C_i$ do 3:    if $|C_i|\ge 4$ then 4:     Fit a line $L_i$ to the center points in $C_i$ 5:     Compute slope ${\theta }_i$ of $L_i$ 6:     if $|{\theta }_i|<{\theta }_t$ then 7:      Mark IAWA feature 6 $|{\theta }_i|>{\theta }_r$ 8:     else if $|{\theta }_i|>{\theta }_r$ then 9:      Mark IAWA feature 10 10:   end if 11:  end if 12:  if $|C_i|\ge 3$ then 13:   Compute $x_{std},y_{std}$ of $C_i$ 14:   if ${\tau }_1$ < $x_{std}$/$y_{std}$ < ${\tau }_2$ then 15:    Mark IAWA feature 11 16:   end if 17:  end if 18: end for 19: Count solitary vessels 20: if solitary vessels/total vessels $>=90\%$ then 21:  Mark IAWA feature 9 22: end if

IAWA features (40–43) relate to the vessel tangential diameter, which can be calculated by the vessel boundary box. Define the pixel width of the vessel d, the pixel width is converted to the real width by $d\ast s$. The features (46–50) relate to the vessel density, which is calculated by Equation (2).

$$\begin{array}{cc}\hfill D& ={\displaystyle \frac{N}{A}}\hfill \\ \hfill A& =w\ast h\ast s\hfill \end{array}$$

(2)

where N is the number of detected vessels, w is the image width, h is the image height, s is the scale of pixels to millimeters, and A represents the area of the image in millimeters squared.

The IAWA features 79, 83, and 85 are related to the shape of the parenchyma, which are derived from the ASD detection results. For parenchyma’s predicted label is ‘wing’, features 80 and 82 are classified based on the ratio of the bounding box height to its width.

The IAWA features 96–99 are related to wood rays and determined by the average of the ray width. According to the IAWA description, the width of a wood ray is defined as the number of cells at its widest part. As shown in Figure 5a, firstly, the rays detected by the model are converted into binary images. Next, a cell detection algorithm [37] is used to detect the cells. The middle part between $1/3$ and $2/3$ is used to compute the ray width. Equally dividing the middle part into multiple horizontal segments, and counting the ray cells per segment. The IAWA features 96–99 are calculated by the average width of segments.

The IAWA features 114–116 are related to the number of rays. The image is divided into n equal parts using horizontal lines, and the number of wood rays intersected by each horizontal line is counted. The principle is shown in Figure 5b, in which the anatomy image is divided by 5 lines. The average number of divided parts is used for calculating the rays per image width in pixels. Finally, using a scale to convert the rays per pixel to rays per millimeter.

3.3. Classification

Wood species classification is based on IAWA code ($IAC$) from both cross section and tangential section images. As illustrated in Figure 6, define the IAWA features of each image as $A_n$. A union operation is used to integrate the IAWA features, which is shown in Equation (3). The presence and absence of the IAWA features are calculated using Equation (4). The result 1 means the feature is present, which is detected from the images. The result 0 means the feature is absent, which is not detected in images.

$$D=A_1\cup A_2\cup \dots \cup A_n$$

(3)

$$f_a\left(x\right)=\left\{\begin{array}{cc}1\hfill & \mathrm{if} x\in D\hfill \\ 0\hfill & \mathrm{if} x\notin D\hfill \end{array}\right.$$

(4)

The $\mathit{IAC}$ is calculated using Equation (5).

$$\mathit{IAC}=\left\{x\right|f_a\left(x\right)\},\hspace{1em}x\in \{6,9,\dots ,116\}$$

(5)

A Naive Bayes-based classifier was employed to classify the wood species. Define the x denotes the IAWA feature code matrix $\mathit{IAC}$, and Y denotes the wood species. The probability of x belonging to Y is calculated using Formula (6). Where the variable c represents the index of the wood species.

$$\underset{c}{\mathit{argmax}} P(Y=c\mid x)=\underset{c}{\mathit{argmax}} {\displaystyle \frac{P(Y=c){\displaystyle \prod _i}(x_i\mid Y=c)}{{\displaystyle \sum _c}P(Y=c){\displaystyle \prod _i}(x_i\mid Y=c)}}$$

(6)

3.4. Evaluation

The performance of classification is evaluated by metrics accuracy, precision, recall, and F1-Score. Accuracy represents the overall correctness of a classification model, calculated as the ratio of correctly predicted samples (both true positives and true negatives) to the total number of samples. Precision is the proportion of true positives among all predicted positives. Recall is the proportion of true positives among the positives, indicating the model’s ability to detect all positive instances. F1-score is a balance of precision and recall. These metrics are computed using Equations (7)–(10).

$$\mathrm{Accuracy}={\displaystyle \frac{\mathit{TP}+\mathit{TF}}{\mathit{TP}+\mathit{FP}+\mathit{TN}+\mathit{FN}}}$$

(7)

$$\mathrm{Precision}={\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}}$$

(8)

$$\mathrm{Recall}={\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}}$$

(9)

$$\text{F1-Score}=2\times {\displaystyle \frac{\mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(10)

where True Positives (TP) denotes the number of correctly predicted positives, False Positives (FP) represents the number of incorrectly predicted positives, True Negatives (TN) indicates the number of correctly predicted negatives, and False Negatives (FN) corresponds to the number of incorrectly predicted negatives.

The intersection over union (IoU) measures the overlap between the predicted and ground truth masks. It is defined as the ratio of their intersection to their union, as shown in Formula (11). The average precision (APx) is used to evaluate the detection performance of the anatomy structure under IoU threshold x, which is the mean of precisions calculated from images. The metric mAP is the mean of APx under different IoU thresholds, and is calculated using the Formula (12).

$$\mathrm{IoU}={\displaystyle \frac{A_{\mathit{pred}}\cap A_{\mathit{gt}}}{A_{\mathit{pred}}\cup A_{\mathit{gt}}}}$$

(11)

where $A_{\mathit{pred}}$ is the predicted area and $A_{\mathit{gt}}$ is the ground truth area.

$$\mathrm{mAP}={\displaystyle \frac{1}{m}}\sum A{P}_i, i\in (0.5,0.55,0.6,\dots ,0.95), m=\mathit{len}\left(P\right)$$

(12)

where $P_i$ denotes the precision calculated under the IoU threshold i.

4. Results

4.1. Classification Performance

The wood species classification performance of the proposed method is evaluated in the testing dataset, which consists of 32 species, and the results are shown in

Table 4. Wood species classification results. The table presents the precision, recall, and F1-score for each species, along with the overall accuracy and weighted average metrics across all species. The weighted average accounts for the number of samples per species, giving more influence to species with more samples.

Species	Precision	Recall	F1-Score
Swietenia macrophylla	0.960	1.000	0.980
Pistacia chinensis	0.935	1.000	0.967
Entandrophragma angolense	1.000	0.962	0.980
Cyrilla racemiflora	0.895	1.000	0.944
Diospyros crassiflora	0.980	0.970	0.980
Diospyros virginiana	0.975	0.985	0.980
Quercus alba	0.944	1.000	0.971
Juglans nigra	0.980	0.990	0.985
Gleditsia triacanthos	0.909	0.833	0.870
Albizia ferruginea	0.692	0.818	0.750
Piptadeniastrum africanum	0.957	0.917	0.936
Intsia bijuga	0.867	0.929	0.897
Pericopsis elata	0.688	0.647	0.667
Pterocarpus soyauxii	1.000	0.950	0.974
Robinia pseudoacacia	0.885	1.000	0.939
Triplochiton scleroxylon	0.905	1.000	0.950
Guarea cedrata	0.941	0.889	0.914
Guarea laurentii	0.970	0.980	0.975
Lovoa trichilioides	0.895	0.773	0.829
Melia azedarach	0.737	0.778	0.757
Milicia excelsa	0.824	1.000	0.903
Nauclea diderrichii	0.938	0.833	0.882
Acer rubrum	1.000	0.900	0.947
Dalbergia cochinchinensis	1.000	0.950	0.974
Dalbergia stevensonii	0.933	1.000	0.966
Dalbergia cultrata	0.960	1.000	0.980
Pterocarpus erinaceus	1.000	0.857	0.923
Entandrophragma cylindricum	0.965	0.975	0.970
Pterocarpus macrocarpus	1.000	0.913	0.955
Dalbergia oliveri	0.970	0.985	0.977
Pterocarpus angolensis	0.975	0.980	0.978
Dalbergia frutescens	0.960	0.970	0.965
Accuracy			0.941
Weighted Average	0.926	0.933	0.927

. Among all species, Rose wood, such as Dalbergia oliveri, Diospyros crassiflora, and Pterocarpus macrocarpus achieved perfect or near-perfect precision values (≥0.97), while maintaining high recall rates (>0.91), indicating it has good performance in these species. Although some species, such as Albizia ferruginea and Pericopsis elata have low F1-score, the method achieved a high F1-score on average. The average accuracy of all species reached 0.941, precision reached 0.926, the average recall reached 0.933, and the average F1-score reached 0.927. These indicate that our proposed method has high robustness in the testing dataset.

4.2. Interpretable Wood Recognition

This method not only predicts the wood species but also outputs image evidence and IAWA features for classification. The full outputs are shown in Figure 7, where three classification examples are presented. Among them, the first two species are correctly predicted, while the last one represents a misclassification case. In Output1, the anatomical structures are color masked with structure names, and different colors represent different structures. These colored masks serve as evidence of the classification in the context of wood anatomy. The cross section images provide vessel and parenchyma structures, and tangential section images provide wood ray features. The images illustrate the detection results for vessels, wood rays and parenchyma. Output2 represents the IAWA feature codes derived from statistical analysis of detected anatomical structures. It provides visible indicators to explain the IAWA features. Output3 is the classification result, which contains the wood species.

4.3. Anatomy Feature Detection

The model was trained on an NVIDIA RTX 4090 GPU, with hyperparameters summarized in

Table 5. Hyperparameter settings used for YOLO-seg model training.

Hyperparameter	Value	Description
epochs	250	Number of training epochs
batch	16	Batch size per iteration
imgsz	640	Input image size
learning_rate	0.01	Initial learning rate
optimizer	SGD	Type of optimizer
weight_decay	0.0005	Weight decay for regularization
momentum	0.937	Momentum factor for SGD

. The training process, including the trends of loss and F1-score over epochs, is shown in Figure 8. Figure 9 presents the detected anatomies from the deep learning model. The first row illustrates the detection results for vessels. The vessels are colored with different colors and used to calculate the IAWA features. The second row exhibits the wood rays, which are covered by color polygons. The third row represents the parenchyma structure detection and classification. The parenchyma contains three types, including winged parenchyma (IAWA 80, 82), parenchyma confluent (IAWA 83), and banded parenchyma (IAWA 85). These images demonstrate that the proposed method highly detects the vessels, pharenchyma, and wood rays. The boundaries of the anatomical structure are accurately recognized using polygons, which provide precise data for calculating the IAWA features.

The number and density of vessels and rays are key parameters used in calculating IAWA features. To evaluate the accuracy of the proposed automated measurement method, 50 images were manually annotated to obtain reference counts of vessels and rays. The comparison between the results from the automated module and manual counting is presented in Figure 10. It can be clearly observed that the vessel and ray counts generated by the automated system are highly consistent with those obtained through manual annotation. The maximum and average measurement errors for vessel count are 7 and 1.16, respectively, while those for ray count are 5 and 1.6. Despite these levels of error, the measured IAWA features remain within the same categorical range, indicating that the measurement results still maintain a high degree of accuracy.

5. Discussion

5.1. Advantage of Multiple Outputs

Compared to traditional deep learning-based methods of wood species recognition, the proposed method enhances the interpretability of the recognition process in a certain field. Unlike deep learning models that only output a classification label, the proposed method outputs not only the wood species but also visually interpretable evidence that includes segmented anatomical structures and IAWA features. These evidences improve the credibility, transparency, and scientific relevance of the recognition results in wood science.

By combining computer vision techniques with established principles from wood science, the proposed method may help narrow the gap between computer science and botanical taxonomy. This interdisciplinary approach could contribute to the development of explainable AI systems tailored for analyzing wood anatomical structures and supporting automated wood identification. In addition, it may facilitate collaboration between data scientists and wood anatomists, offering the potential for more reproducible and interpretable recognition tools in wood science applications.

5.2. Performance of Different Models

Anatomical structure detection is crucial in this method, as it forms the foundation for both IAWA feature calculation and wood species classification. In the experiments, the detection performance of anatomy structures in YOLOv8-seg, CondInst [38], PointRend [39], and Mask R-CNN [40] models is evaluated to find the best structure detection model. The comparison is shown in Figure 11, in which, the first column displays the original image, and the other columns are detection results from image detection models. The first row is vessel images, the middle row is ray images, and the last row is parenchyma images. It is evident that the YOLOv8-seg model successfully detected nearly all the targets, whereas the other models missed many in the first row. In the detection results on other images, YOLOv8-seg consistently outperformed or matched other models. Overall, YOLOv8-seg demonstrates a stronger capability in detecting all target objects compared to other models, which is particularly important for the accurate calculation of IAWA features.

The anatomical feature detection capabilities of different models are detailed in

Table 6. Detection performance comparison of different models for anatomical structures. Metrics reported include Average Precision (AP) at IoU thresholds 0.50 (AP50) and 0.75 (AP75), and the mean Average Precision (mAP) across different IoU thresholds evaluated for vessel, ray, and parenchyma.

Model	Vessel			Ray			Parenchyma
	AP50	AP75	mAP	AP50	AP75	mAP	AP50	AP75	mAP
Mask-RCNN	0.770	0.747	0.646	0.948	0.811	0.674	0.760	0.582	0.494
PointRend	0.741	0.705	0.596	0.930	0.806	0.674	0.531	0.501	0.457
CondInst	0.764	0.660	0.544	0.903	0.613	0.551	0.533	0.496	0.471
YOLOv8-seg	0.912	0.864	0.715	0.935	0.834	0.716	0.730	0.623	0.572

$mAP$: average of $APs$, mean of ($AP50,AP55,\dots ,AP95$). The Bolded values represent the best scores.

. Notably, YOLOv8-seg achieves the highest vessel detection performance with an $AP50$ score of 0.912, surpassing other models by approximately 0.2 points. Meanwhile, for the detection of wood rays and parenchyma, YOLOv8-seg reached the highest mAP of 0.716 and 0.572. Compared to other models, YOLOv8-seg demonstrates superior detection performance overall, which can be attributed to its efficient feature extraction and fusion mechanisms that enhance its ability to detect densely packed small targets.

5.3. Impact of Multiple Images

The IAWA features are regarded as the international standard for wood species identification. Due to a single microscopic image usually including partially anatomical characteristics, reliable wood identification typically requires multiple microscopic images to capture enough IAWA features. For finding the best combination of images, we evaluated the impact of cross section and tangential section images on classification performance metrics precision and recall, as illustrated in Figure 12. The combination of cross section and tangential section images yields significantly higher classification metrics compared to using either type of image alone. Specifically, the combined input (CS+TS) achieves both precision and recall about 80%. Furthermore, among the single-section inputs, cross-sectional images consistently outperform tangential section images in classification performance. Notably, as Figure 13 shows, the accuracy first peaks when the data includes 6 cross section images and 4 tangential section images. Considering efficiency, this combination is recommended, achieving an accuracy of 94%.

The above results can be attributed to the complementary nature of anatomical features captured by different section types. The combination of cross section and tangential section images provides a more comprehensive representation of wood anatomy. By integrating the two views, the model benefits from diverse features, leading to enhanced classification performance. When comparing single-view performance, cross section images outperform tangential section images. This outcome can be attributed to the more IAWA features present in the cross section. According to the IAWA feature table used in this study, a total of 18 features, accounting for 72% are related to the cross section. The limited features in the tangential section may restrict the model’s learning capacity and negatively affect classification performance. Overall, these findings highlight the importance of view integration in achieving reliable wood species classification.

5.4. Classification Methods

To evaluate the effectiveness of different machine learning algorithms for wood species classification based on IAWA features, we compared five commonly used classification methods: Support Vector Machine (SVM) [41], Naive Bayes, K-Nearest Neighbors (KNN) [42], decision tree [43], and Random Forest [44]. Their performance is summarized in

Table 7. Recognition accuracy of five classification models evaluated on the train set and test set.

Model	Accuracy of Test Set	Accuracy of Train Set
SVM	0.939	0.946
Naive Bayes	0.941	0.945
Decision Tree	0.900	0.973
Random Forest	0.937	0.942
KNN	0.921	0.940

The Bolded values represent the best scores.

. All five classification methods exhibited strong classification performance on the test dataset, with accuracy scores exceeding 0.90. Among them, the Naive Bayes model achieved the highest overall performance, with an accuracy of 94.1% on the test set. Notably, its accuracy on the training set was 94.5%, indicating a well-balanced performance between the training and test sets. In contrast, the Decision Tree model achieved an accuracy of 97.3% on the training set but dropped to 90.0% on the test set, suggesting a noticeable overfitting issue.

The performance of all five classifiers can be attributed to the two sections of images capturing a wide range of IAWA features, and these features exhibit high discriminability for wood species classification. Among the five classification methods, the Naive Bayes model achieved the best performance. The Naive Bayes model is based on a strong assumption that all features are conditionally independent. In this study, the anatomical features used for classification are clearly defined and relatively independent of each other. These features reduce the likelihood of strong correlations, making the Naive Bayes model especially suitable and effective for this task. Moreover, Naive Bayes exhibits inherent robustness in handling high-dimensional and sparse feature spaces, as well as missing values, since its probabilistic framework allows each feature to contribute independently to the final prediction [45]. Compared to more complex models like SVM or Random Forest, which may be more sensitive to feature interactions or require extensive hyperparameter tuning, Naive Bayes offers a simpler yet highly effective approach for classification. Its consistent performance on both training and testing datasets further suggests strong generalization capability under the given data distribution.

The importance and correlation of IAWA features with wood species were assessed using SHAP values and Pearson correlation coefficients [46] derived from a random forest model. The results were visualized in the form of a heat bubble chart shown in Figure 14. From a species-level perspective, features 42, 46, 47, 79, and 98 exhibit relatively high SHAP values and strong correlations in Swietenia macrophylla. From a feature-level perspective, features 50 and 114 generally display low SHAP values across most species, with only a few species exhibiting notable contributions from these features. Overall, among all IAWA features, those related to density and diameter—which are relatively easy to quantify—tend to have higher SHAP values and stronger correlations.

5.5. Limitations and Challenges

In our experiments, we observed that certain wood species exhibited relatively low classification accuracy. For example, as shown in the fourth classification result in Figure 7, the model misclassified Albizia ferruginea as Pericopsis elata. A comparison of the IAWA features of the two species revealed a high degree of similarity, with Albizia ferruginea possessing an additional characteristic—IAWA feature 98: “Larger rays commonly 4- to 10-seriate”—which the model failed to identify. Upon further inspection of the images, it was found that some samples suffered from low image quality, as illustrated in Figure 15. Although the model was able to detect the general outline of wood rays, after the binarization process, the boundaries of ray cells became indistinct or even disappeared, leading to a significant underestimation in cell count. In some extreme cases, such as when ray cells are overlapping or poorly contrasted, ray cells were difficult to distinguish even by human observation, making their identification and counting particularly challenging.

To address this issue, future work may consider incorporating measurable anatomical features, such as the actual width and height of wood rays, into the classification process—aligned with the IAWA standard, which often uses cell width as a discriminative feature. In addition, previous studies have shown that introducing a third viewing plane—the radial section—can reveal more diagnostic features and improve classification performance [27]. The radial section offers a new perspective of the wood sample, providing anatomical features that cannot be observed in the cross and tangential sections, enabling a more comprehensive and interpretable identification. Thus, integrating radial-section images into the current framework may help reduce uncertainty caused by the limitations of cross and tangential sections.

In this study, the model achieves an average inference time of approximately 80 ms per image, which is suitable for near real-time processing in laboratory conditions. However, this may pose challenges for deployment on resource-constrained devices. To improve efficiency and broaden applicability, future research could explore lightweight network architectures and model compression techniques to reduce computational requirements. Moreover, the current method still relies on high-quality images, particularly in terms of anatomical clarity under the microscope. The acquisition of microscopic images typically requires a certain level of technical expertise, including sample sectioning, staining, and microscopic imaging, which may limit its use in non-laboratory or field settings. As a future direction, we will explore the feasibility of using non-microscopic images, such as macroscopic section images, to extract IAWA features for species classification. This approach may help reduce the reliance on specialized equipment while maintaining the anatomical interpretability.

6. Conclusions

In this study, we proposed a novel wood recognition method that integrates a deep learning-based anatomical structure detector, quantitative feature extraction aligned with IAWA standards, and an interpretable Bayesian classifier for wood species identification. By incorporating both cross and tangential section images, the method achieved a classification performance with an accuracy of 94.1%. The results suggest that the method can achieve high classification accuracy while offering a degree of biological interpretability based on wood anatomical principles, which may help domain experts to understand and validate the model’s decisions.

Nevertheless, two main limitations of the proposed approach should be acknowledged. First, the method requires high-quality microscopic images, which rely on specialized equipment and precise sample preparation. This requirement may hinder its practical deployment in field environments or in contexts where such resources are unavailable. Second, although the model demonstrates near real-time performance under laboratory conditions, it may pose challenges for deployment on resource-constrained devices.

Future research will focus on reducing dependency on high-resolution microscopic inputs by exploring the use of macroscopic images and on optimizing the model for lightweight deployment. These directions are crucial for enhancing the practicality, scalability, and accessibility of AI-assisted wood identification systems in real-world applications.