NC State
BioResources
Kim, J. H., Purusatama, B. D., Savero, A. M., Prasetia, D., Jang, J. H., Park, S. Y., Lee, S. H., and Kim, N. H. (2024). “Convolutional neural network performance and the factors affecting performance for classification of seven Quercus species using sclereid characteristics in the bark,” BioResources 19(1), 510-524.

Abstract

Based on the sclereids in the bark of oak species, a convolutional neural network (CNN) was employed to validate species classification performance and its influencing factors. Three optimizers including stochastic gradient descent (SGD), adaptive moment estimation (Adam), root mean square propagation (RMSProp), and dataset augmentation were adopted. The accuracy and loss stabilized at approximately 15 to 20 and 70 to 80 epochs for the augmented and non-augmented condition, respectively. In the last five epochs, the RMSProp-augmented condition achieved the highest accuracy of 89.8%, whereas the Adam-augmented condition achieved the lowest accuracy of 73.8%. Regarding the loss, SGD-non-augmented condition was the lowest at 0.498, whereas Adam-augmented condition was the highest at 2.740. The highest accuracy was influenced by RMSProp at 0.194. Dataset augmentation had a significant influence on accuracy at 0.456. Homogeneous subsets among the validation conditions indicated that the accuracy and loss were classified into the same subset using an augmented dataset during the training, regardless of the optimizer. Only Adam and RMSProp with non-augmented datasets were categorized into the same subset during the test. Hence, species classification using CNN and sclereid characteristics in the bark was feasible, and RMSProp with augmented datasets showed optimal performance for species classification.


Download PDF

Full Article

Convolutional Neural Network Performance and the Factors Affecting Performance for Classification of Seven Quercus Species using Sclereid Characteristics in the Bark

Jong Ho Kim,a Byantara Darsan Purusatama,b Alvin Muhammad Savero,a Denni Prasetia,a Jae Hyuk Jang,c Se Yeong Park,a Seung Hwan Lee,a and Nam Hun Kim a,*

Based on the sclereids in the bark of oak species, a convolutional neural network (CNN) was employed to validate species classification performance and its influencing factors. Three optimizers including stochastic gradient descent (SGD), adaptive moment estimation (Adam), root mean square propagation (RMSProp), and dataset augmentation were adopted. The accuracy and loss stabilized at approximately 15 to 20 and 70 to 80 epochs for the augmented and non-augmented condition, respectively. In the last five epochs, the RMSProp-augmented condition achieved the highest accuracy of 89.8%, whereas the Adam-augmented condition achieved the lowest accuracy of 73.8%. Regarding the loss, SGD-non-augmented condition was the lowest at 0.498, whereas Adam-augmented condition was the highest at 2.740. The highest accuracy was influenced by RMSProp at 0.194. Dataset augmentation had a significant influence on accuracy at 0.456. Homogeneous subsets among the validation conditions indicated that the accuracy and loss were classified into the same subset using an augmented dataset during the training, regardless of the optimizer. Only Adam and RMSProp with non-augmented datasets were categorized into the same subset during the test. Hence, species classification using CNN and sclereid characteristics in the bark was feasible, and RMSProp with augmented datasets showed optimal performance for species classification.

DOI: 10.15376/biores.19.1.510-524

Keywords: Bark; Convolutional neural networks (CNNs); Oak; Sclereids; Species classification

Contact information: a: Department of Forest Biomaterials Engineering, College of Forest and Environmental Sciences, Kangwon National University, Chuncheon 24341, Republic of Korea; b: Institute of Forest Science, Kangwon National University, Chuncheon 24341, Republic of Korea; c: FC Korea Land Co., Ltd., Seoul 07271, Republic of Korea; * Corresponding author: kimnh@kangwon.ac.kr

INTRODUCTION

Globally, there is excess demand for wood (UNECE 2007), resulting in a gradual appreciation of the value of wood as a resource in various sectors. Species identification is increasingly recognized as a significant process for enhancing value of wood resources and is required in numerous fields such as optimizing the utilization of conventional wood resources, protecting endangered species, and facilitating practical customs clearance operations.

Recently, there has been a surge in research aimed at automating the classification of wood species and enhancing their precision of automated classification. These efforts have been driven by the objectives of streamlining the process and reducing subjectivity. Ilic (1993) proposed the potential for computer-based automated species identification, while Wheeler and Baas (1998) emphasized the importance of computer-aid specification identification. Since the 2000s, studies applying computer vision to identify wood species have been conducted in earnest. Tou et al. (2007) proposed a system to identify wood species in real time using macroscopic images of wood cross sections. Bremananth et al. (2009) attempted to computerize a wood species recognition system using computer vision technology. In the 2010s, research on artificial intelligence-based wood species identification increased explosively, owing to the rapid development of computing performance and machine learning technology. As a representative example, Hermanson and Wiedenhoeft (2011) introduced an overview and advantages of species identification using machine vision in their review paper. Since the first meeting of the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) in 2010, many high-performance convolutional neural network (CNN) models, such as VGGNet (Simonyan and Zisserman 2015), GoogLeNet (Szegedy et al. 2015), and ResNet (He et al. 2016) have been developed. These inventions have led to rapid developments in image classification using CNN.

Recently, CNNs have been employed to classify wood species using datasets containing wood images. Kwon et al. (2017) reported the possibility and performance of softwood species classification using CNN-based models, such as LeNet and miniVGGNet. Kwon et al. (2019) improved the performance of automated species classification by using a CNN with ensemble methods. Yang et al. (2019) also employed a CNN with ensemble methods for classifying Korean softwood species using datasets from near-infrared spectra (NIR) analysis results and macroscopic images of radial sections. Hwang et al. (2020) attempted to visualize anatomical features using machine learning technology for quantitative analysis. Zhao et al. (2021) demonstrated the efficacy of a CNN for accurately classifying various wood species based on microscopic image datasets. Huang et al. (2021) demonstrated the efficacy of a transfer-learning model with enhanced pooling layers for enhanced wood identification using cross-sectional wood images. Hwang and Sugiyama (2021) conducted a methodological examination of computer-vision-based species identification. Their research anticipated the potential of computer vision to enhance the accessibility of wood identification to the public and make significant contributions to the field of wood science. Cao et al. (2022) utilized an artificial neural network trained using species-specific thermal conductivity trends to classify different species.

Most studies on wood species identification have focused on the xylem anatomical characteristics (Kim et al. 2021; Savero et al. 2022; Savero et al. 2023). However, bark also displays distinct characteristics in each species, rendering it a helpful tool for species identification (IAWA Committee 2016). Species identification using bark has a notable benefit because bark can be obtained from standing trees without timber harvesting. A few studies (Fiel and Sablatnig 2010; Bertrand et al. 2017; Carpentier et al. 2018; Kim et al. 2022) have attempted to utilize bark for automated species identification using artificial intelligence. They focused only on the outer appearance of the bark from standing trees for dataset preparation. However, the anatomical features of the bark could be more systematic than the surface features, showing a higher efficiency in feature selection and extraction, which affects species classification performance using computer vision. Sclereids are sclerenchyma cells that are variable in form and size, but typically not much elongated, with thick, often polylamellate, lignified secondary walls with many pits. They develop mainly in the nonconducting phloem, cortex, and periderm by modification of parenchyma cells. However, there are some examples of earlier development directly from cambial derivatives. Sclereids are extremely variable in size and shape, and this diversity has been classified into different sclereid types for the plant body, including brachysclereids, columnar sclereids, osteosclereids, astrosclereids, and filiform sclereids. Thus, sclereids can be used as keys for species differentiation (IAWA committee 2016).

Therefore, in the present study, the performance and performance-influencing factors of CNNs using sclereid characteristics in the bark to identify seven oak species were investigated. Three optimizers, stochastic gradient descent (SGD), adaptive moment estimation (Adam), and root mean square propagation (RMSProp), were used to analyze the feasibility and efficacy of the classification performance of the wood species.

EXPERIMENTAL

Materials

The barks of six domestic oak species obtained from the research forest of Kangwon National University, and Quercus suber, donated by FC Korea Land Co., Ltd. (Seoul, Korea), were used in this study. Three stems of each of the six domestic oak species were harvested, and bark samples were collected from the breast height of the stems. Several Q. suber bark samples with dimensions of 60 × 100 cm were also used. Comprehensive information about the samples is presented in Table 1.

Table 1. Sample Information

*D.B.H.: Diameter at breast height

Methods

Sample preparation for the dataset

Wood discs (3 cm thick) were prepared using a chainsaw from the breast height of each oak trees. The barks were carefully detached from the discs. After separating the bark from the disc, a table saw (Professional Cabinet Saw model with 100-teeth saw blade, SawStop, Oregon, USA) was used to cut the top and bottom of the bark specimen to a uniform thickness of about 1 cm for easier microscopic observation. Transverse sections were sanded using a series of progressively finer sandpapers: #80, #120, #220, and #400. After the sanding treatment, the surface was cleaned using an air compressor. Twenty specimens of each species were examined using a visual microscope (MM-40, Nikon, Tokyo, Japan) equipped with a 2.5X objective lens (CF Plan 2.5X, Nikon, Tokyo, Japan). Digital micrographs of the inner phloem of the bark specimens were obtained using a microscope camera (IMTCam; IMT, British Columbia, Canada) and subsequently used as a dataset. The collection contained 1,400 (200 for each species) microscopic images.

General macroscopic features

Sclereid characteristics, which are representative macroscopic anatomical features of the seven oak species, were analyzed using micrographs obtained to prepare a dataset. Five micrographs of each species were analyzed to determine the quantitative characteristics of the sclereids. The number of sclereids within one micrograph (3.75 × 2.50 mm), the area of individual sclereids, and the minimum and maximum area of sclereids for each species were measured.

Dataset pretreatment

The training dataset was augmented to examine the influence of the dataset size on the training procedure for the neural network architecture. Several parameters were applied to quantitatively pretreat and augment the dataset images. The parameters included rescaling the data using a ratio of 1/255 for normalization, 10° rotation, shifting the data horizontally and vertically by 10%, zooming the data by 20%, and performing horizontal and vertical flipping. Figure 1 illustrates the examples of data augmentation.

Fig. 1. Example micrographs of Quercus aliena: (a) original image; (b-k) augmented images

The dataset consisted of 1,400 images, with 200 images per species. The dataset was divided into 80% for training and 20% for testing. The test dataset was created by using a file random extraction program to extract 40 out of 200 images of each species dataset. The selected images were then saved in the separated paths with the training dataset. Data augmentation was exclusively implemented on the training dataset. Table 2 shows the disparity in quantity between the pre- and post-augmentation data.

Table 2. Composition of Training and Test Dataset in Augmentation

Verification factors influencing CNN

Classification performance and the factors influencing performance were analyzed using basic CNN architecture, as depicted in Fig. 2. The architecture of the CNN was designed by referring to the general structure reported in the past (Elgendy 2021; Loy 2020). The CNN design comprised four convolutional layers, four MaxPooling layers, and two fully connected layers. Furthermore, the model was enhanced by incorporating two dropout layers and a flattened layer. Finally, the Softmax activation function was applied.

Fig. 2. Convolutional neural network architecture for species classification in the present study

The CNN architecture was trained using a dataset comprising seven oak species. Categorical cross-entropy was used as the loss function for the multiclass classification. Optimizers are tools used to optimize the resources used during weight updates in each training phase (Cho 2018). In this study, three specific optimizers (i.e., SGD, Adam, and RMSProp) were used and compared to assess the influence of different optimizers on the efficiency of a CNN. The learning rates for each optimizer were set to SGD 0.0001, Adam 0.001 (default), and RMSProp 0.00001, respectively.

Statistical analysis of factors influencing CNN

Bivariate correlation analysis (SPSS 26.0; IBM, New York, USA) was used to examine the Pearson correlation coefficients between the variables. The optimizer type and augmentation for the analysis were used as nominal variables, whereas the accuracy and loss rate were used as scale variables. Additionally, a one-way ANOVA and Duncan’s post-hoc analysis were used to examine the homogeneous subsets of the results.

RESULTS AND DISCUSSION

Figure 3 shows cross-sectional micrographs of the bark of the seven oak species in the dataset. Except for Q. suber bark, sclereids were found in the bark of all species. The sclereids’ size, shape, and frequency varied among the species. Table 3 shows a comparison of the number of sclereids per unit area and the size of individual sclereids by species. The sclereids in Q. variabilis and Q. aliena (181.9 × 103 µm2 and 175.4 × 103 µm2, respectively) were larger than those in the other species, while Q. serrata and Q. acutissima showed smaller sizes (52.7 × 103 µm2 and 65.4 × 103 µm2, respectively) compared to other species. This represented 2- to 4-fold differences in the sclereids size between species. Q. variabilis and Q. aliena showed lower sclereid frequencies than other species. As shown in Fig. 3f, sclereids in Q. acutissima, in particular, had a large variation in size. Prasetia et al. (2022) previously reported that sclereids were frequently found in the bark of Q. variabilis and were absent in Q. suber. Kim (1993) conducted a similar study on the anatomical characteristics of Q. variabilis and Q. suber and reported that sclereids were rarely present in Q. suber.

Table 3. Quantitative Characteristics of Sclereids in the Six Oak Species

Fig. 3. Cross-section micrographs of the barks from seven Quercus species: Quercus dentata (a), Quercus serrata (b), Quercus mongolica (c), Quercus variabilis (d), Quercus aliena (e), Quercus acutissima (f), and Quercus suber (g). White arrows indicate sclereids. Scale bars: 1,000 µm

Table 4. Comparison of Average Loss and Accuracy on the Last Five Steps Per Optimizer

Table 4 lists the average losses and classification accuracies during the last five stages for each validation condition. During the training phase, Adam produced the lowest loss and highest accuracy, suggesting superior performance irrespective of the implementation of data augmentation. However, Adam’s performance was relatively poor compared to the other optimizers during the test phase. SGD demonstrated a classification accuracy of approximately 83%, whereas RMSProp achieved a classification accuracy ranging from 79% to 90%. SGD is a function related to the random extraction and calculation of partial data at each learning stage and updating them more quickly and frequently. RMSProp is a function of applying an exponential moving average to prevent gradient loss problems, which can be obtained by lowering the reflection weight of the initial training and increasing the reflection weight of the recent gradient, thereby preventing the vanishing gradient problem during early training (Géron 2020). Therefore, in this study, SGD demonstrated superior performance in high-resolution microscopic image analysis required for species classification using bark, while RMSProp showed excellent performance in microscopic image analysis of the bark despite potential variations.

The classification accuracy and loss of bark for the seven oak species using a CNN are shown in Figs. 4 and 5. Under most training and test conditions, there was a notable increase in accuracy and a decrease in loss as the number of epochs progressed. Nevertheless, a noteworthy anomaly was detected in the training phase when the Adam optimizer was employed with the augmented datasets. In this particular instance, the observed loss exhibited an increasing pattern as the number of epochs increased, in contrast to the outcomes observed under other conditions. The observed pattern indicates overfitting, a phenomenon in which the model demonstrates high performance on data that closely resemble the training set, but exhibits poor performance on test or validation data (Oh 2021). This was because the noise generated during the augmentation process of the dataset negatively affected the learning results and only caused overfitting when the Adam optimizer with an augmented dataset was used for training. In this study, all conditions except the Adam-augmented dataset condition exhibited a notable trend of reduced loss and enhanced accuracy as the number of epochs increased, particularly when utilizing the augmented dataset as opposed to the non-augmented dataset. The utilization of the augmented dataset led to the stability of both the loss and classification accuracy after roughly 15 to 20 epochs, whereas the non-augmented dataset needed approximately 70 to 80 epochs for stabilization.

Fig. 4. Verification results of convolutional neural networks (CNNs) architecture for the seven oak species classification using the barks in training phase