Summary
Computer vision-based wood identification has been successfully applied to recognize tree species using digital images of wood sections or surfaces. However, this image-to-species approach can only recognize a limited number of species due to two main reasons: 1) the lack of a good reference database requiring high-quality standardized images from multiple individuals of hundreds or even thousands of traded timber species, and 2) species not included in the reference database cannot be identified without expert knowledge. Another bottleneck is that the feature extraction process used by these species recognition approaches is a black box, thereby creating a discrepancy between machine learning features and wood anatomical features. This discrepancy prevents wood anatomists from understanding how these machine-learning algorithms work. Here, we survey currently existing methods used in feature extraction, classification, and deep learning methods applied in wood identification along with their pitfalls and opportunities. As an example of how the field could move forward, we launch the idea of building an image-to-features-to-species identification approach based on microscopic wood images as well as text files comprising wood anatomical descriptions. If we can manage machine learning-based algorithms to recognize the main wood anatomical traits that experts use to identify species in a (semi-)automated way, this would boost wood identification in two ways: (1) extensive reference databases for each species would become less crucial as the databases are ordered at the trait level, (2) timber identification would become more feasible for species that have not yet been included in the reference database as long as wood anatomical descriptions are available.
Introduction
Wood identification plays a vital role in forest law enforcement, conservation, timber industry, and scientific research. Accurate identification helps the fight against illegal logging, as the combination of illegal and legal logging activities further contributes to harmful environmental and social impacts, such as deforestation, habitat destruction, loss of biodiversity, and exploitation of local communities (Dormontt et al. 2015; UNODC Committee 2016, 2020; Jahanbanifard et al. 2019; Chandra 2022; Low et al. 2022). For instance, accurate wood identification tools would enable international and local authorities to enforce legal regulations such as the US Lacey Act, the Australian Illegal Logging Prohibition Act (AILPA), the European Union (EU) initiated ‘Forest Law Enforcement, Governance and Trade’ (FLEGT), former European Timber Regulation (EUTR) and the upcoming new EU regulation on deforestation-free supply chains (EUDR), and the international Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES COP 19 2022) (European Commission 2003, 2010, 2023; Johnson & Laestadius 2011; Dormontt et al. 2015; UNODC Committee 2016, 2020). Furthermore, it would promote sustainable and environmentally responsible practices across the timber supply chain (He et al. 2020; Olschofsky & Köhl 2020; Liu et al. 2022; Ravindran et al. 2022a,b). On the other hand, it is important to realise that not all logging operations, even if legal, adhere to sustainable practices where timber is harvested in such a way that the integrity and health of the forest ecosystem over the long term are maintained (Putz et al. 2022).
Unfortunately, a widely applicable and reliable method to identify woods remains elusive. A variety of wood identification methods are available, such as DNA-based methods (Lowe et al. 2010; Nithaniyal et al. 2014; Jiao et al. 2015, 2018, 2022; He et al. 2019; Yin et al. 2020), chemical profiling based on mass spectrometry (Musah et al. 2015; Evans et al. 2017; Zhang et al. 2019; Price et al. 2021, 2022; Deklerck 2023), near-infrared spectroscopy (Braga et al. 2011; Bergo et al. 2016; Pan et al. 2022), detector dogs (Braun 2013), and wood anatomy (Gasson et al. 2010; Wheeler 2011; Helmling et al. 2018; Brandes et al. 2020; Lens et al. 2020; Liu et al. 2022). According to the best practice guide for forensic timber identification, issued by the United Nations Office on Drugs and Crime (UNODC Committee 2016), wood anatomy remains the most commonly used approach to identify wood samples as it is quick, cost-effective, and relies on a well-established, standardized set of wood anatomical traits that have been particularly selected for the purpose of wood identification by experienced wood anatomists, for both angiosperms (IAWA Committee 1989) and gymnosperms (IAWA Committee 2004). However, in most cases, wood anatomy mostly only allows for identification up to the genus level as closely related species typically have nearly identical microscopic patterns in their wood anatomy (Wheeler & Baas 1998; Gasson 2011). That being said, all wood identification methods suffer from accuracy issues at the species level (Price et al. 2022).
The rapid advancement of Artificial Intelligence (AI) has aroused a surge of interest in automated Computer Vision-based Wood Identification (CVWID). This is especially true for Machine Learning (ML) which relates to a subdiscipline of AI focusing on computer algorithms with the capability to learn from data, and Deep Learning (DL) which is an ML-subcategory taking advantage of deep neural networks. These pioneering approaches enable computers to learn from expertise and improve without explicit programming, making it ideal for classifying timber species based on digital images of microscopic or macroscopic wood anatomy. The potential benefits of these methods extend to aiding the work of trained wood anatomists, whose expertise requires years of training (Lens et al. 2020), and allow non-experts such as customs officers to apply these wood identification methods to efficiently detect and prevent fraud (Wiedenhoeft et al. 2019). The process of CVWID involves two crucial steps: feature extraction and classification. In feature extraction, computer algorithms identify the critical parts of the image, while classification employs additional computer algorithms to differentiate images based on these extracted features. By comparing the image features from an unknown species with image features from known species embedded into the reference database, the computer calculates the probability of the unknown image with that of a known species. This is referred to as image-to-species recognition.
Compilation of traditional machine learning studies in wood-based recognition, showing the relatively low number of species (often less than 100) and low number of images (typically less than 100 per species), along with information about type of images, feature extraction algorithm, classifier used, and the (artificially high) species recognition (top) accuracy for each study.
Citation: IAWA Journal 45, 4 (2024) ; 10.1163/22941932-bja10157
Despite significant progress in CVWID (Filho et al. 2014; Ibrahim et al. 2017; Rosa da Silva et al. 2017, 2022; de Andrade et al. 2020; He et al. 2020; Lens et al. 2020; Verly Lopes et al. 2021; Figueroa-Mata et al. 2022; Ravindran & Wiedenhoeft 2022; Ravindran et al. 2022a,b; Bello et al. 2023; Yang et al. 2023), the development of a globally accessible ML-based wood identification platform remains challenging. Three main reasons can be cited. First, the current image-to-species recognition methods are not suited to identify a large number of timber species. This is because these methods heavily rely on the time-consuming creation of a reference database, including representative images of as many individuals per species as possible for the more than a thousand different timber species that are being traded (Mark et al. 2014; Lens et al. 2020). Unfortunately, available image databases do not meet these criteria for most species (see Table 1). Second, the current recognition methods are not designed to recognize species that are not included in the reference database. Third, the results achieved by image-to-species methods are expressed as probabilities that an image is correctly classified, but the exact features or regions in the images that lead to such a classification generally remain unclear. This creates confusion among wood anatomists who have a hard time being convinced that the outcome of ML-based wood recognition methods is trustworthy and generally applicable when adding more species. This is referred to as ‘feature incompatibility’ and is further exacerbated by the fact that the meaning of the word ‘feature’ in ML and ‘feature’ in wood anatomy is different (Fig. 1). An ML feature refers to mathematical representations of high-dimensional vectors used for numerical calculations, while a wood anatomical feature reflects the type, shape, and arrangement of the cellular structure in wood.
To explore how ML can overcome these issues and to develop a global CVWID method that can reliably recognize hundreds or even thousands of timber species throughout the world based on wood anatomy, we need to gain insight into the history of ML-based wood identification. Here, we review the ML-based wood recognition process, including image acquisition, feature extraction and classification, summarize the main pitfalls and opportunities, and put forward a new idea for a global CVWID pipeline that does not require an extensive reference database at the species level.
Overview of machine learning image-to-species models for wood recognition. In the model training component (left), known reference images (De, Dc) are first processed; computer algorithms such as GLCM or LBP are then used to extract computer features from these images (feature extraction), after which a classifier trains the feature extraction (FX) model. On the right-hand side, the DL model training process is almost the same, with the only difference being that the CNN or Transformer algorithms incorporate both the feature extraction and classifier steps. Both FX and DL models can be used to predict whether an unknown image belongs to reference data (De or Dc) or not. If the unknown image does not belong to the reference data, there may be an error recognition result.
Citation: IAWA Journal 45, 4 (2024) ; 10.1163/22941932-bja10157
Image acquisition
Although a variety type of images (e.g. scanning electron microscope images, X-ray images, fluorescence microscope images) are available for CVWID, this paper only focuses on images derived from light microscopic sections and macroscopic surfaces of wood. At the microscopic level, high-resolution images can be captured in bulk with a slide scanner in the lab or individually with a light microscope equipped with a digital camera (Martins et al. 2013; Rosa da Silva et al. 2017). At the macroscopic level (Ruffinatto et al. 2015; Ruffinatto & Crivellaro 2019), images can be obtained using portable imaging equipment to develop real-time applications on-site or in the field (Tang et al. 2018; de Andrade et al. 2020; Arévalo et al. 2021; Figueroa-Mata et al. 2022). Examples are the Xyloscope (Hermanson et al. 2019), a digital imaging system for viewing and recording macroscopic images of wood surfaces for the open-source XyloTron platform (Ravindran & Wiedenhoeft 2020), and the more affordable XyloPhone (Wiedenhoeft 2020) and Xylorix (Tay 2019) that is adaptable to a camera of the smartphone (Ravindran et al. 2019, 2020, 2021; Arévalo et al. 2021). In addition to the global platforms like XyloTron, there are also more local platforms that are working well at a smaller scale within countries, such as iWood (He et al. 2021) in China and AIKO (Arifin et al. 2020) in Indonesia.
Images captured at various magnifications can influence machine learning algorithms for feature recognition due to the distinct textures observed at different scales, although Nguyen-Trong (2023) advocates that a 20× magnification yields favorable outcomes. When making microscopic wood sections or macroscopic surface images in the field, it is crucial to attain a specific quality threshold to enable precise identification. The challenge arises from the inherent limitations of field or on-site identification, which often restricts complex sample processing. The processing for wood surfaces may involve the use of sandpaper with varying grain sizes ranging from coarse to fine-grit to eliminate scratches and grooves left by the blade. Further polishing of the wood surface with finer grain sizes does not yield statistically better outcomes in CVWID once the quality threshold is surpassed (Owens et al. 2023). The same quality threshold probably also applies to quick and dirty microscope sections that are made in the field compared to more standardized lab conditions, until future algorithms can accommodate for these cutting artefacts. We envision that the future of wood identification will most likely rely on a combination of macro CVWID tools in the field or on-site complemented with microscopic tools in the lab. However, the accuracy levels of the current wood identification tools should always be cautiously interpreted, since any accurate global ML model relies on a good quality (microscopic and macroscopic) reference database, which is not yet available at the global scale for most traded timbers (Lens et al. 2020).
Regardless of the images acquired by digital imaging devices, the basic unit of computer storage is the pixel point (with 0–255 numbers indicating color brightness). A color image is created by arranging multiple pixel points both vertically and horizontally to form three RGB (Red, Green, Blue) channels. The density of these pixels determines the resolution of the image. For example, a common microscopic image taken by an optical microscope with a digital camera has a resolution of 2592 × 1944 pixels times 3 RGB channels, which leads to a computer image file having 14.7 × 106 pixel points. If multiple image data were calculated by classification algorithms directly, the computational capacity of common computers would quickly be exceeded. This leads to the first paradox: wood anatomists tend to use high-resolution images to show more details of the anatomical features, while computers prefer lower-resolution images covering a larger surface to fit their processing ability. A combined effort between experts in wood anatomy and computer vision is the way forward to finding the right balance.
To represent inter-species and intra-species variability accurately, the number of images also plays an essential role. If an image-to-species recognition approach is employed, it means that many images are required for each tree species in the reference database. But how many individuals and images per species are enough? Wood anatomists typically consider images from more than 20 individuals per species sufficient to reach statistical significance (Lens et al. 2020), although this minimum threshold likely needs to be upscaled for species with a wide distribution range. For ML and especially DL approaches, many more training images per category (hundreds or thousands) are required depending on the model’s structure and the number of parameters (Cho et al. 2016; Shahinfar et al. 2020). Unfortunately, as shown in Table 1 and Table 2, the number of species or images per species that are being used in ML studies is relatively small (often less than 100), making present reference databases unfit for accurate wide-scale wood identification based on an image-to-species recognition approach. Indeed, few images per species typically lead to an overestimation of the recognition accuracy values in CVWID (Ravindran & Wiedenhoeft 2022). That is also the reason why the number of images is often artificially augmented for ML analyses using a wide array of image pre-processing approaches, such as binary transformation, cropping, rotation, shrinkage, magnification, and sub-images that lead to several smaller images based on a single original image (Fabijańska et al. 2021; He et al. 2021; Yang et al. 2023). The downside of this pre-processing is, however, that some smaller-scale traits in the image may be lost or are no longer recognizable. Therefore, various strategies, such as incorporating overlapping regions during the cropping process, are employed to mitigate this issue. Another issue with these original reference images is that they are generally derived from only transverse orientation planes and have a similar low magnification that further impedes observations of small-scale wood anatomical traits, such as vessel pits.
Compilation of DL studies in wood-based recognition, with information about the number of species and images, type of images and feature extraction AI model (standard CNN structures (LeNet, VGGs, ResNets, AlexNet, DenesNets, Inception, Xception, RCNN)).
Citation: IAWA Journal 45, 4 (2024) ; 10.1163/22941932-bja10157
It is interesting to see the discrepancy between the small image datasets that are currently being used in ML-based recognition studies on the one hand. On the other hand, the long tradition of wood anatomy has been generating hundreds of thousands of wood anatomical slides in institute collections worldwide and has led to a wealth of published wood anatomical descriptions and images (Wheeler 2011; Koch et al. 2018; Sugiyama et al. 2020; Haneca et al. 2022). Part of the available wood anatomical data is publicly available via InsideWood (InsideWood 2004-onwards; Wheeler 2011), the largest and best-known online wood anatomical database in the world — containing over 50 000 images that cover more than 10 000 species, along with wood descriptions following the standardized IAWA hardwood and softwood lists (Fig. 2; Wheeler et al. 2020).
73 IAWA hardwood features with more than 1000 images on the InsideWood website (extracted June 2022). Of the 163 hardwood features, 156 have more than 100 images, and 99 have more than 500 images. This set of images is sufficient to make multiple mutually exclusive datasets for deep learning.
Citation: IAWA Journal 45, 4 (2024) ; 10.1163/22941932-bja10157
Feature extraction and classification
Handcrafted feature extraction
Computer vision features can be identified by handcrafted algorithms that can be used to analyze and classify different features of wood images. These algorithms are designed by computer experts who came up with processes of manually designing and selecting specific image features, such as edges, corners or textures. Gray-Level Co-occurrence Matrices (GLCM) (Haralick et al. 1973) algorithm is a popular handcrafted algorithm for feature extraction of wood anatomical images (see references in Table 1). GLCM analyzes the spatial relationships in images based on pixel intensities, allowing the algorithm to detect larger-scale wood anatomical traits, such as wood growth ring boundaries, vessels, and rays as viewed in transverse sections. The Local Binary Pattern (LBP) algorithm (Ojala et al. 1994) is another popular algorithm used in wood anatomy (Table 1) and changes the 0–255 color values around a central pixel into a 0/1 binary value making it effective in capturing and representing the features that have repetitive patterns or distinctive textures. While GLCM and LBP focus more on larger, pattern-like traits in wood images, the Scale-Invariant Feature Transform (SIFT) algorithm (Lowe 1999) detects smaller-scale anatomical traits (like septate fibres, vessel pits, mineral inclusions) by focusing on local parts of an image (Hwang et al. 2018, 2020a,b).
We can draw two conclusions from these general handcrafted feature extraction algorithms that were not initially designed to extract features from wood images. First, they can still be used to identify a number of anatomical traits, indicating that a comprehensive analysis and further research on manually extracted features could offer an opportunity to expand our domain knowledge and cultivate a deeper understanding of wood. Second, the discrepancy between handcrafted feature extraction and IAWA features remains, making it hard for wood anatomists to trust the recognition results.
Features classification
In computer vision, a classifier is a computer algorithm that analyzes images and assigns them to predefined categories. Depending on whether or not there is human expertise guidance available, classification can be divided into unsupervised (no human expertise involved) or supervised (image annotation by human experts). Unsupervised classification employs algorithms to automatically identify patterns and structures in images without relying on pre-existing labels or training examples. Supervised classification requires a large amount of labeled training data to identify new, unlabeled images from species that are included in the reference database. Below, we only discuss the three main supervised classifiers based on the algorithm properties: statistical-based classifiers, rule-based classifiers, and perceptron classifiers.
Some data are simple enough to be classified using a linear function to assign images to categories (linear classifier), while other data are more complex and require non-linear classifiers to model the relationships between image features and the output categories. Support Vector Machine (SVM) (Cortes & Vapnik 1995; Table 1) is among the common statistical-based classifiers used in wood studies. This classifier achieves excellent separation of high-dimensional, non-linear data through a series of complex computations, which can reach high species recognition accuracy in wood identification (Rosa da Silva et al. 2017, 2022; Hwang et al. 2020b).
Classification can also be implemented by rules. The Decision Tree (DT) algorithms (Quinlan 1986; Salzberg 1994; Lewis 2000), which are composed of many ‘if-else’ rules to explain the classification principle, have good explainability due to the clear rules. Nevertheless, they are prone to overfitting, meaning that the accuracy is often artificially high, which makes DT not that convenient for wood recognition (Pan & Kudo 2011). Multiple decision trees can be combined to form a Random Forest (RF) (Ho 1998), which afterward, can be computationally strengthened via the AdaBoost algorithm that can improve the accuracy of weak classifiers by several orders of magnitude (Freund & Schapire 1997). Both RF and AdaBoost can be used to assess the importance of anatomical features based on their classification rules (Hwang et al. 2020a), and sometimes can even outperform SVM with respect to classification accuracy in wood identification studies (Sun et al. 2015).
Perceptron classifiers (Rosenblatt 1958) work by imitating how human brain neurons receive and process information. If a perceptron classifier initially obtains a wrong classification, it can adjust its parameters until it obtains the correct result. Stacking MultiLayer Perceptrons (MLP) (Werbos 1974) as a tool to build a classifier, also known as Artificial Neuron Network (ANN) classifier, forms the foundation of modern deep learning. However, ANN for wood recognition does not reach the high accuracies obtained by SVM (Yadav et al. 2013; Yuliastuti et al. 2013; Zhao et al. 2014; Zamri et al. 2016), because the full potential of ANN cannot be reached without high computational power and a suitable ANN network model.
To summarize, classifiers are applicable to wood classification, but some perform much better than others. Since the classifiers only classify the data into categories, i.e. species in case of image-to-species wood identification, they do not solve the unexplainability issue in wood recognition. In addition, classifiers are not designed to classify species that are not present in the reference database.
Deep learning-based wood recognition
During the last two decades, machine learning has entered the era of deep learning (LeCun et al. 2015) due to a combination of massive increase in data, rapid development of computer hardware (mainly GPU: Graphic Processing Units), and major advances in algorithms (Hinton & Salakhutdinov 2006). For instance, deep Convolutional Neural Networks (CNNs) exemplify a process where image features are initially extracted through mathematical convolution operations (Lecun et al. 1998). Subsequently, these features are input into a network structure comprising stacked layers of neurons, known for their deep architecture. Training this network through multiple rounds can result in more efficient feature extraction and enhanced classification accuracies compared to traditional machine learning methods (Lens et al. 2020). This superiority is attributed to the universal approximation capabilities inherent in neural networks, as established by Hornik et al. (1989, 1990) and Leshno et al. (1993).
The evolution of CNN structures includes various architectures such as the VGG series (Simonyan & Zisserman 2015), which emphasizes low-level features and simplicity for ease of implementation. The ResNet series (He et al. 2016) has gained prominence due to its ease of training, while the EfficientNet series (Tan & Le 2019) boasts fewer parameters yet remains effective in capturing features. However, CNNs are often viewed as black boxes, meaning it is difficult to understand the exact details (neuron parameters) of how the network arrived at its decision. In other words, it is difficult to interpret the thought process of the network as we lack an understanding of how it learns, extracts and classifies the features. Despite our incomplete knowledge of the process involved, humans can still assign a specific task to the algorithm to identify species based on the images (He et al. 2020; Lens et al. 2020).
In contrast to CNNs, the Vision Transformer (Vaswani et al. 2017) adopts a different strategy for feature extraction, treating image components similarly to how words are handled in Transformers designed for Natural Language Processing (NLP) (Dosovitskiy et al. 2021; Radford et al. 2021). Transformer is a neural network architecture that has achieved breakthroughs in ML-based NLP (e.g., ChatGPT). The Vision Transformer excels at capturing broader correlations in images compared to CNNs, particularly in the case of microscopic wood anatomical images. These images typically lack a prominent central object, but are characterized by intricate patterns composed of points, lines, edges, and so forth (Zhang et al. 2021).
Both supervised CNNs and Vision Transformers heavily depend on a priori data, which is typically available as labeled data provided by experts. These labels typically consist of a large amount of data, which is divided into a small number of categories. The commonly used dataset, ImageNet-1K (Deng et al. 2009), is a dataset of more than 1.2 million images consisting of 1000 common object categories with more than 1000 images per category. Microscope wood image datasets for ML recognition, however, are considerably smaller (as shown in Table 2). That is why the wood datasets require data augmentation, which involves generating smaller sub-images from the original image, as well as transfer learning, which makes training small datasets easier (Thrun & Pratt 2012). Recent CNN-based recognition systems for macroscopic images that trained image-to-species models by ImageNet transfer learning resulted in a series of successful applications (Ravindran et al. 2019; de Andrade et al. 2020; He et al. 2020; Ravindran et al. 2020, 2021, 2022a,b; Arévalo et al. 2021; Kirbas & Çifci 2022; Bello et al. 2023; and see Table 2).
In addition to the advancements in transfer learning for extracting features from wood images, one can leverage feature visualization techniques such as Class Activation Mapping (CAM) (Zhou et al. 2016), Gradient-weighted Class Activation Mapping (Grad-CAM) (Selvaraju et al. 2017), and Grad-CAM
Too soon to celebrate — major bottlenecks of ML-based wood recognition
Although new DL methods are becoming more powerful and offer advantages compared to the more traditional ML methods, there are still several bottlenecks to overcome before obtaining an ML-based wood recognition pipeline that yields reliable accuracy rates for hundreds or thousands of timber species. There are two fundamental challenges we need to tackle. First and foremost, a new unified reference wood image dataset needs to be developed, containing a sufficient number of high-quality images (both within a single species and across species) from correctly identified species (linked to a herbarium voucher) to serve as a basis for all image-to-species ML studies. As long as the image-to-species reference database is far from complete, there will always be many tree species that are not in the database. This means that the ML algorithms will always have limited applications at a global scale without an inclusive database. As the original high-resolution images of wood anatomy cannot be directly fed to ML algorithms because of the current computational limitations, one way to increase the number of images per species is to generate smaller, preferably overlapping sub-images from the original images to avoid losing smaller-scale traits (Fabijańska et al. 2021; He et al. 2021; Yang et al. 2023; see also section on image acquisition). Another way we could increase the number of images is by incorporating images from all three orientation planes, rather than utilizing only one. While it is acknowledged that transverse section images generally exhibit superior accuracy (Barmpoutis et al. 2018; Figueroa-Mata et al. 2018; de Geus et al. 2020; Yang et al. 2023), longitudinal section images offer additional diagnostic features that help to identify species (Wheeler & Baas 1998; Gasson 2011; Lens et al. 2020; Wu et al. 2021; Rosa da Silva et al. 2022). Therefore, integrating images from all three orientations will enhance the effectiveness of ML identification (Kirbas & Çifci 2022). It is important, however, to always use precise orientation planes, especially for the radial planes, because a light deviation from the radial planes generates much shorter rays. Ultimately, filling the major gaps in this reference dataset, both in terms of number of images per species and number of species, will require a joint effort from all the major wood collection institutes in the world to digitize their microscopic slide collections and upload them in the same open-access cloud infrastructure (Lens et al. 2020). This will prove advantageous for CNN models and, notably, Vision Transformer models. The Transformer architecture, with its increased parameter count, necessitates a substantial volume of data — ranging from tens of thousands to even millions of images — for effective training, surpassing the requirements of CNNs.
A second challenge we need to overcome is that the fundamental principle of DL remains a black box. As explained before, heatmaps generated by visualization technology can be used to better understand how the DL model is actually working, which offers the chance for wood anatomy experts to validate the outcome of the model.
Towards a robust, GLOBAL wood identification pipeline
Recent developments in ML reveal promising solutions for current bottlenecks in wood recognition
Recent trends in ML show an ever-increasing demand for data that are generated by different techniques, which are referred to as multimodal data. Multimodal deep learning (Ramachandram & Taylor 2017) will likely become a game changer in wood recognition. This cutting-edge methodology may allow analyzing both macroscopic and microscopic images of wood from the same species along with the published wood description in the form of a text file. An interesting way forward may be the self-supervised learning model SimCLR (Chen et al. 2020), which has the potential to generate an automatic textual description for images, known as captioning. The downside of this model, though, is (1) that training data for image captioning should include captions, which entails significantly more effort compared to standard annotation tasks, and (2) SimCLR requires substantial computational resources. Another trend that will become more prominent in the near future is that ML algorithms will be able to directly process raw, high-resolution anatomical images due to the increasing computing power of GPUs (He et al. 2020).
From an annotation perspective, ML is developing from supervised learning (time-consuming annotation is required by wood anatomists) toward unsupervised learning (expert knowledge not required). The so-called contrast learning approach has made remarkable progress in recent years and states that annotation of data is not always necessary (Wu et al. 2018; van den Oord et al. 2019; Chen et al. 2020). This new self-supervised ML method does not require a large amount of well-labeled data and is considered a promising new development for future ML (Yeh et al. 2022), especially since these kinds of new methods have been proven to gain better recognition accuracy than those based on supervised learning (Chen et al. 2020). The common method of self-supervision is to process unsupervised data by introducing supervised signals from other modalities. For example, by using a large amount of text description information as a supervised signal from wood descriptions belonging to rare or endangered species that are poorly represented in reference image databases, more accurate recognition of these species may be achieved.
Another recent promising aspect of ML is the ability to recognize data not included in the reference database, called zero-shot learning (Pourpanah et al. 2022). Simply put, zero-shot learning allows computers to apply what they have learned from images in the reference database to images from unknown species outside the database, making them good at figuring out the unknown. Importantly, to achieve recognition of unknown categories of data outside the reference dataset, zero-shot learning requires pre-training on images that are highly relevant to the target classification, meaning that a large-scale wood image database is a prerequisite for this pre-training. Since these pre-training options are not possible for a global CVWID model, it has not yet been developed to such a level that it can be practically applied in wood identification, but the concept of this zero-shot learning approach has proven to be theoretically feasible (Fei-Fei et al. 2006).
In summary, in recent years, ML has moved toward larger models and unsupervised domains. By pre-training a large amount of unlabeled data and then performing transfer learning on a smaller dataset, it is possible to achieve more accurate results than the more traditional supervised methods. Regarding model architecture, Vision Transformer has better global feature extraction capability than CNN. Interestingly, Vision Transformer’s intrinsic advantage of multimodal processing can lead to unprecedented applications in wood recognition by (1) identifying timber species that are not yet included in the reference database, and (2) by lowering the necessity of having a huge, well-labeled reference dataset with a substantial number of images per species for as many species as possible. Evidently, a high-quality, representative, global reference collection will always remain important for any future application in wood recognition.
Conclusions and suggestions for further research
In this ML-based wood identification paper, we discuss a number of traditional feature extraction and classifier algorithms that have been regularly used in wood identification studies, and state that they are often outcompeted by the more advanced deep learning methods that integrate feature extraction and classification in a complex mathematical way. Recent advances in DL methodologies and increased computational power offer unprecedented opportunities to improve wood recognition. Examples of important recent breakthroughs are, among others, (1) multimodal deep learning where different types of data (text in wood descriptions combined with digital images of wood sections or surfaces) can be integrated, (2) self-supervised contrast learning enabling wood recognition based on a largely unsupervised (i.e., not labeled or not annotated by wood anatomists) image dataset, 3) applying methodologies in ML models via image heatmaps that allow wood anatomists to link computer vision features with anatomical features, and 4) zero-shot learning with the potential to recognize species that are not included in the reference database. However, despite all these advances in image-to-species ML algorithms that assign an unknown wood image to the correct species name, we are still struggling to overcome the lack of one global, labelled, digital image reference database for thousands of timber species (Lens et al. 2020; Hwang & Sugiyama 2021; Verly Lopes et al. 2021; De Blaere et al. 2023). To further improve ML-based wood identification in a short time frame, we propose to focus on the following three aspects:
(1) One way to move CVWID forward is to build image-to-features ML model(s) that focuses on automatically detecting anatomical wood features as an intermediate step, and then employ a features-to-species search via for instance InsideWood (InsideWood 2004-onwards; Wheeler 2011; Wheeler et al. 2020). This can be achieved by re-labeling existing wood anatomical images to obtain suitable image datasets for most of the IAWA features (IAWA Committee 1989, 2004). Or perhaps even better, we might combine the image-to-features ML model with the features-to-species model and develop a Feature Recognition AI (FRAI) in CVWID. The main advantage of this FRAI model over existing image-to-species ML models is that it does not rely on the existing incomplete species-based image reference databases, but instead would rely on (novel) trait-based reference image databases that easily allow more than 1000 images per trait for most IAWA features (Fig. 2). Therefore, this commentary paper shows that there is a strong theoretical basis for developing an image-to-features-to-species ML model with heatmap visualization that could be a promising new avenue to boost wood identification in the near future.
(2) More attention should be paid to the problem of ‘feature incompatibility’ since wood anatomists need to understand which parts in images are used for species recognition by machine learning. In addition to heatmaps generated by feature visualization, there are several ways to visualize traits. Some examples include exploring the potential of applying computer vision to extract anatomical features through the investigation of feature extraction methods like SIFT (a widely used local feature detection algorithm), studying the explainability of ML-based wood recognition methods such as CNN or Vision Transformer, examining the fitting ability of neural networks to recognize anatomical features of wood, and conducting mathematical modeling and evaluations to assess feature compatibility.
(3) Publicly available wood description and image data are an essential first step to further develop unsupervised recognition methods to also identify rare species for which there are insufficient wood samples available in collection institutes worldwide. More specifically, we can apply the principle of self-supervised multimodal contrast learning to further improve wood recognition, by combining descriptive text and images to increase the classification accuracy of unknown or poorly sampled species that are not necessarily included in the ‘classical’ image reference database. The vast amount of standardized wood anatomical descriptions contain a hidden treasure of information for ML methods, especially in the field of wood anatomy that reached a consensus about a standardized set of traits used to identify timbers (IAWA Committee 1989, 2004). Much of the older literature, published in journals or books that are inaccessible via the internet, will likely be unlocked by ongoing digitization efforts that make these wood anatomical descriptions applicable for global wood identification pipelines. At the same time, our IAWA community should take action to boost the creation of a global shared image dataset containing a sufficient number of high-quality images for each species for at least all the traded timbers. As suggested by Lens et al. (2020), this global image dataset can only be compiled in the coming years if the large wood collection institutes join forces and digitize their slide collection in an appropriate, standardized way. This global effort would boost the development of a reliable, global wood identification pipeline that should become freely accessible for all (non-) academic stakeholders in their endeavors to identify wood samples from past times (e.g. fossils, samples in archaeological sites or natural history collections), or to inspect modern imported logs that may have been illegally cut.
Corresponding authors; emails: qiujian@swfu.edu.cn; frederic.lens@naturalis.nl
Acknowledgements
This work was supported by a special fundamental research project of Natural Science Foundation of Yunnan province (202001AS070044), by the Yunnan Provincial Education Department Scientific Research project (2020J0400), and by the Chinese Scholarship Council (202108535013). We would like to express our gratitude to Pieter Baas, Naturalis Biodiversity Center, for the help and guidance during the whole project, and to Yafang Yin, Chinese Academy of Forestry, for his guidance of algorithms in wood AI recognition.
References
Arévalo R, Pulido REN, Solórzano GJF, Soares R, Ruffinatto F, et al.2021. Imaged based identification of Colombian timbers using the Xylotron: a proof of concept international partnership. Colomb. For. 24(1): 5–16. DOI: 10.14483/2256201x.16700.
Arifin MRPE, Sugiarto B, Wardoyo R, Rianto Y. 2020. Development of mobile-based application for practical wood identification. IOP Conf. Ser.: Earth Environ. Sci. 572(1): 012040. DOI: 10.1088/1755-1315/572/1/012040.
Barmpoutis P, Dimitropoulos K, Barboutis I, Grammalidis N, Lefakis P. 2018. Wood species recognition through multidimensional texture analysis. Comput. Electron. Agric. 144: 241–248. DOI: 10.1016/j.compag.2017.12.011.
Bello R-W, Oluigbo CU, Tiebiri CB, Ogidiga BD, Moradeyo OM. 2023. Wood species identification using mask RCNN-residual network approach. Pro Ligno 19(1): 41–51.
Bergo MCJ, Pastore TCM, Coradin VTR, Wiedenhoeft AC, Braga JWB. 2016. NIRS identification of Swietenia macrophylla is robust across specimens from 27 countries. IAWA J. 37(3): 420–430. DOI: 10.1163/22941932-20160144.
Braga JWB, Pastore TCM, Coradin VTR, Camargos JAA, da Silva AR. 2011. The use of near infrared spectroscopy to identify solid wood specimens of Swietenia macrophylla (CITES appendix II). IAWA J. 32(2): 285–296. DOI: 10.1163/22941932-90000058.
Brandes AF das N, Novello BQ, de Lemos DN, do Nascimento LB, Albuquerque RP, Tamaio N, Barros CF. 2020. Wood anatomy of endangered Brazilian tree species. IAWA J. 41(4): 510–576. DOI: 10.1163/22941932-bja10037.
Braun B. 2013. Wildlife detector dogs — a guideline on the training of dogs to detect wildlife in trade. WWF Germany, Berlin.
Chandra R. 2022. Wood classification for efficiency in preventing illegal logging using k-nearest neighbor. J. Mantik 6(1): 494–501.
Chattopadhyay A, Sarkar A, Howlader P, Balasubramanian VN. 2018. Grad-CAM++: improved visual explanations for deep convolutional networks. 839–847. DOI: 10.1109/WACV.2018.00097.
Chen T, Kornblith S, Norouzi M, Hinton G. 2020. A simple framework for contrastive learning of visual representations. In: International conference on machine learning, 21 November: 1597–1607.
Cho J, Lee K, Shin E, Choy G, Do S. 2016. How much data is needed to train a medical image deep learning system to achieve necessary high accuracy? arXiv:. DOI: 10.48550/arXiv.1511.06348.
CITES COP 19. 2022. Summary of the 19th meeting of the CITES conference of the parties: 14–25 November 2022. Earth Negot. Bull. 21(112): 1–21.
Cortes C, Vapnik V. 1995. Support-vector networks. Mach. Learn. 20(3): 273–297. DOI: 10.1007/BF00994018.
de Andrade BG, Basso VM, Latorraca JV de F. 2020. Machine vision for field-level wood identification. IAWA J. 41(4): 681–698. DOI: 10.1163/22941932-bja10001.
De Blaere R, Lievens K, Van Hassel D, Deklerck V, De Mil T, et al.2023. SmartWoodID — an image collection of large end-grain surfaces to support wood identification systems. Database: baad034. DOI: 10.1093/database/baad034.
De Geus AR, da Silva SF, Gontijo AB, Silva FO, Batista MA, Souza JR. 2020. An analysis of timber sections and deep learning for wood species classification. Multimed. Tools Appl. 79(45): 34513–34529. DOI: 10.1007/s11042-020-09212-x.
Deklerck V. 2023. Timber origin verification using mass spectrometry: challenges, opportunities, and way forward. Forensic Science International: Animals and Environments 3: 100057. DOI: 10.1016/j.fsiae.2022.100057.
Deng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L. 2009. ImageNet: a large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition: 248–255. DOI: 10.1109/CVPR.2009.5206848.
Dormontt EE, Boner M, Braun B, Breulmann G, Degen B, et al.2015. Forensic timber identification: it’s time to integrate disciplines to combat illegal logging. Biol. Conserv. 191: 790–798. DOI: 10.1016/j.biocon.2015.06.038.
Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, et al.2021. An image is worth 16 × 16 words: transformers for image recognition at scale. arXiv:. DOI: 10.48550/arXiv.2010.11929.
European Commission. 2003. Forest Law Enforcement, Governance and Trade (FLEGT) — Proposal for an EU Action Plan. Communication from the Commission COM 251. European commission, Brussels.
European Commission. 2010. Regulation (EU) No 995/2010 of the European Parliament and of the Council of 20 October 2010 laying down the obligations of operators who place timber and timber products on the market. OJ. L 295: 23–34.
European Commission. 2023. Regulation (EU) 2023/1115 of the European Parliament and of the Council of 31 May 2023 on the making available on the Union market and the export from the Union of certain commodities and products associated with deforestation and forest degradation and repealing Regulation (EU) No 995/2010. OJ. L 150: 206–247.
Evans PD, Mundo IA, Wiemann MC, Chavarria GD, McClure PJ, Voin D, Espinoza EO. 2017. Identification of selected CITES-protected Araucariaceae using DART TOFMS. IAWA J. 38(2): 266–281. DOI: 10.1163/22941932-20170171.
Fabijańska A, Danek M, Barniak J. 2021. Wood species automatic identification from wood core images with a residual convolutional neural network. Comput. Electron. Agric. 181: 105941. DOI: 10.1016/j.compag.2020.105941.
Fei-Fei L, Fergus R, Perona P. 2006. One-shot learning of object categories. IEEE Trans. Pattern Anal. Mach. Intell. 28(4): 594–611. DOI: 10.1109/TPAMI.2006.79.
Figueroa-Mata G, Mata-Montero E, Valverde-Otarola JC, Arias-Aguilar D. 2018. Using deep convolutional networks for species identification of xylotheque samples 2018. In: IEEE International Work Conference on Bioinspired Intelligence (IWOBI), San Carlos, Costa Rica: 1–9. DOI: 10.1109/IWOBI.2018.8464216.
Figueroa-Mata G, Mata-Montero E, Valverde-Otárola JC, Arias-Aguilar D, Zamora-Villalobos N. 2022. Using deep learning to identify Costa Rican native tree species from wood cut images. Front. Plant Sci. 13: 789227. DOI: 10.3389/fpls.2022.789227.
Filho PLP, Oliveira LS, Nisgoski S, Britto AS. 2014. Forest species recognition using macroscopic images. Mach. Vis. Appl. 25(4): 1019–1031. DOI: 10.1007/s00138-014-0592-7.
Freund Y, Schapire RE. 1997. A decision-theoretic generalization of on-line learning and an application to boosting. J. Comput. Syst. Sci. 55(1): 119–139. DOI: 10.1006/jcss.1997.1504.
Gasson P. 2011. How precise can wood identification be? Wood anatomy’s role in support of the legal timber trade, especially CITES. IAWA J. 32(2): 137–154. DOI: 10.1163/22941932-90000049.
Gasson P, Miller R, Stekel DJ, Whinder F, Zieminska K. 2010. Wood identification of Dalbergia nigra (CITES appendix I) using quantitative wood anatomy, principal components analysis and naive Bayes classification. Ann. Bot. 105(1): 45–56. DOI: 10.1093/aob/mcp270.
Hafemann LG, Oliveira LS, Cavalin P. 2014. Forest species recognition using deep convolutional neural networks. In: 22nd International Conference on Pattern Recognition (ICPR), Stockholm, Sweden: 1103–1107. DOI: 10.1109/ICPR.2014.199.
Haneca K, Deforce K, Van Der Laan J, Nicolaij S, Lange S. 2022. WOODAN: an online database of archaeological wooden objects. Veget. Hist. Archaeobot. 31(5): 541–547. DOI: 10.1007/s00334-022-00868-z.
Haralick RM, Shanmugam K, Dinstein I. 1973. Textural features for image classification. IEEE Trans. Syst. Man Cybern. SMC-3(6): 610–621. DOI: 10.1109/TSMC.1973.4309314.
He J, Sun Y, Yu C, Cao Y, Zhao Y, Du G. 2022. An improved wood recognition method based on the one-class algorithm. Forests 13(9): 1350. DOI: 10.3390/f13091350.
He K, Zhang X, Ren S, Sun J. 2016. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition: 770–778.
He T, Jiao L, Yu M, Guo J, Jiang X, Yin Y. 2019. DNA barcoding authentication for the wood of eight endangered Dalbergia timber species using machine learning approaches. Holzforschung 73(3): 277–285. DOI: 10.1515/hf-2018-0076.
He T, Lu Y, Jiao L, Zhang Y, Jiang X, Yin Y. 2020. Developing deep learning models to automate rosewood tree species identification for CITES designation and implementation. Holzforschung 74(12): 1123–1133. DOI: 10.1515/hf-2020-0006.
He T, Shoujia S, Lu Y, Zhang Y, Jiao L, Yin Y. 2021. iWood: an automated wood identification system for endangered and precious tree species using convolutional neural networks. Sci. Silv. Sin. 57(9): 152–159.
Helmling S, Olbrich A, Heinz I, Koch G. 2018. Atlas of vessel elements: identification of Asian timbers. IAWA J. 39(3): 249–352. DOI: 10.1163/22941932-20180202.
Hermanson JC, Dostal D, Destree JC, Wiedenhoeft AC. 2019. The Xyloscope: a field-deployable macroscopic digital imaging device for wood. United States Department of Agriculture, Forest Service, Forest Products Laboratory, Beltsville, MD.
Hinton GE, Salakhutdinov RR. 2006. Reducing the dimensionality of data with neural networks. Science 313(5786): 504–507. DOI: 10.1126/science.1127647.
Ho TK. 1998. The random subspace method for constructing decision forests. IEEE Trans. Pattern Anal. Mach. Intell. 20(8): 832–844. DOI: 10.1109/34.709601.
Hornik K, Stinchcombe M, White H. 1989. Multilayer feedforward networks are universal approximators. Neural Netw. 2(5): 359–366. DOI: 10.1016/0893-6080(89)90020-8.
Hornik K, Stinchcombe M, White H. 1990. Universal approximation of an unknown mapping and its derivatives using multilayer feedforward networks. Neural Netw. 3(5): 551–560. DOI: 10.1016/0893-6080(90)90005-6.
Hwang SW, Sugiyama J. 2021. Computer vision-based wood identification and its expansion and contribution potentials in wood science: a review. Plant Methods 17(1): 47. DOI: 10.1186/s13007-021-00746-1.
Hwang SW, Kobayashi K, Zhai S, Sugiyama J. 2018. Automated identification of Lauraceae by scale-invariant feature transform. J. Wood Sci. 64(2): 69–77. DOI: 10.1007/s10086-017-1680-x.
Hwang SW, Kobayashi K, Sugiyama J. 2020a. Evaluation of a model using local features and a codebook for wood identification. IOP Conf. Ser. Earth Environ. Sci. 415(1): 012029. DOI: 10.1088/1755-1315/415/1/012029.
Hwang SW, Kobayashi K, Sugiyama J. 2020b. Detection and visualization of encoded local features as anatomical predictors in cross-sectional images of Lauraceae. J. Wood Sci. 66(1): 16. DOI: 10.1186/s10086-020-01864-5.
IAWA Committee. 1989. IAWA list of microscopic features for hardwood identification. IAWA Bull. 10(3): 219–332.
IAWA Committee. 2004. IAWA list of microscopic features for softwood identification. IAWA J. 25(1): 1–70. DOI: 10.1163/22941932-90000349.
Ibrahim I, Khairuddin ASM, Abu Talip MS, Arof H, Yusof R. 2017. Tree species recognition system based on macroscopic image analysis. Wood Sci. Technol. 51: 431–444. DOI: 10.1007/s00226-016-0859-4.
Ibrahim I, Khairuddin ASM, Arof H, Yusof R, Hanafi E. 2018. Statistical feature extraction method for wood species recognition system. Eur. J. Wood Wood Prod. 76(1): 345–356. DOI: 10.1007/s00107-017-1163-1.
InsideWood. 2004-onwards. InsideWood. Available online at http://insidewood.lib.ncsu.edu/search (accessed 28 June 2023).
Jahanbanifard M, Gravendeel B, Lens F, Verbeek F. 2019. Ebony wood identification to battle illegal trade. Biodivers. Inf. Sci. Stand. 3: e37084. DOI: 10.3897/biss.3.37084.
Jiao L, Liu X, Jiang X, Yin Y. 2015. Extraction and amplification of DNA from aged and archaeological Populus euphratica wood for species identification. Holzforschung 69(8): 925–931. DOI: 10.1515/hf-2014-0224.
Jiao L, Lu Y, Zhang M, Chen Y, Wang Z, et al.2022. Ancient plastid genomes solve the tree species mystery of the imperial wood “Nanmu” in the Forbidden City, the largest existing wooden palace complex in the world. Plants People Planet 4(6): 696–709. DOI: 10.1002/ppp3.10311.
Jiao L, Yu M, Wiedenhoeft AC, He T, Li J, Liu B, Jiang X, Yin Y. 2018. DNA barcode authentication and library development for the wood of six commercial Pterocarpus species: the critical role of xylarium specimens. Sci. Rep. 8(1): 1945. DOI: 10.1038/s41598-018-20381-6.
Johnson A, Laestadius L. 2011. New laws, new needs: the role of wood science in global policy efforts to reduce illegal logging and associated trade. IAWA J. 32(2): 125–136. DOI: 10.1163/22941932-90000048.
Khalid M, Yusof R, Mohd Khairuddin AS. 2011. Tropical wood species recognition system based on multi-feature extractors and classifiers. In: 2nd international conference on instrumentation control and automation: 6–11. DOI: 10.1109/ICA.2011.6130117.
Kirbas I, Çifci A. 2022. An effective and fast solution for classification of wood species: a deep transfer learning approach. Ecol. Inform. 69: 101633. DOI: 10.1016/j.ecoinf.2022.101633.
Kobayashi K, Kegasa T, Hwang S-W, Sugiyama J. 2019. Anatomical features of Fagaceae wood statistically extracted by computer vision approaches: some relationships with evolution. PLoS ONE 14(8): e0220762. DOI: 10.1371/journal.pone.0220762.
Koch G, Heinz I, Schmitt U, Richter H-G. 2018. Wood anatomy — the role of macroscopic and microscopic wood identification against illegal logging. In: Németh R, Teischinger A, Rademacher P, Bak M (eds.), 8th hardwood conference with special focus on “new aspects of hardwood utilization — from science to technology 8, Sopron, Hungary: 10–11.
Kwon O, Lee HG, Lee MR, Jang S, Yang SY, Park SY, Choi IG, Yeo H. 2017. Automatic wood species identification of Korean softwood based on convolutional neural networks. J. Kor. Wood Sci. Technol. 45(6): 797–808. DOI: 10.5658/WOOD.2017.45.6.797.
Kwon O, Lee HG, Yang SY, Kim H, Park SY, Choi IG, Yeo H. 2019. Performance enhancement of automatic wood classification of Korean softwood by ensembles of convolutional neural networks. J. Kor. Wood Sci. Technol. 47(3): 265–276. DOI: 10.5658/WOOD.2019.47.3.265.
LeCun Y, Bengio Y, Hinton G. 2015. Deep learning. Nature 521(7553): 436–444. DOI: 10.1038/NATURE14539.
LeCun Y, Bottou L, Bengio Y, Haffner P. 1998. Gradient-based learning applied to document recognition. Proc. IEEE 86(11): 2278–2324. DOI: 10.1109/5.726791.
Lens F, Liang C, Guo Y, Tang X, Jahanbanifard M, da Silva FSC, Ceccantini G, Verbeek FJ. 2020. Computer-assisted timber identification based on features extracted from microscopic wood sections. IAWA J. 41(4): 660–680. DOI: 10.1163/22941932-bja10029.
Leshno M, Lin VYa, Pinkus A, Schocken S. 1993. Multilayer feedforward networks with a nonpolynomial activation function can approximate any function. Neural Netw. 6(6): 861–867. DOI: 10.1016/S0893-6080(05)80131-5.
Lewis RJ. 2000. An introduction to classification and regression tree (CART) analysis. In: Annual meeting of the society for academic emergency medicine in San Francisco, California: 14.
Liu S, He T, Wang J, Chen J, Guo J, Jiang X, Wiedenhoeft AC, Yin Y. 2022. Can quantitative wood anatomy data coupled with machine learning analysis discriminate CITES species from their look-alikes? Wood Sci. Technol. 56: 1567–1583. DOI: 10.1007/s00226-022-01404-y.
Low MC, Schmitz N, Boeschoten LE, Cabezas JA, Cramm M, et al.2022. Tracing the world’s timber: the status of scientific verification technologies for species and origin identification. IAWA J. 44(1): 63–84. DOI: 10.1163/22941932-bja10097.
Lowe AJ, Wong K-N, Tiong Y-S, Iyerh S, Chew F-T. 2010. A DNA method to verify the integrity of timber supply chains; confirming the legal sourcing of Merbau timber from logging concession to sawmill. Silvae Genet. 59(1–6): 263–268. DOI: 10.1515/sg-2010-0037.
Lowe DG. 1999. Object recognition from local scale-invariant features. In: Proceedings of the seventh IEEE international conference on computer vision, vol.2: 1150–1157. DOI: 10.1109/ICCV.1999.790410.
Mark J, Newton A, Oldfield S, Rivers M. 2014. The international timber trade: a working list of commercial timber tree species. Botanic Gardens Conservation International, Richmond.
Martins J, Oliveira LS, Nisgoski S, Sabourin R. 2013. A database for automatic classification of forest species. Mach. Vis. Appl. 24(3): 567–578. DOI: 10.1007/s00138-012-0417-5.
Martins J, Oliveira LS, Sabourin R, Britto AS. 2018. Forest species recognition based on ensembles of classifiers. In: IEEE 30th International Conference on Tools with Artificial Intelligence (ICTAI), Volos, Greece: 371–378. DOI: 10.1109/ICTAI.2018.00065.
Martins JG, Oliveira LS, Sabourin R. 2012. Combining textural descriptors for forest species recognition. In: IECON 2012 — 38th annual conference on IEEE industrial electronics society: 1483–1488. DOI: 10.1109/IECON.2012.6388523.
Mata-Montero E, Arias-Aguilar D, Figueroa-Mata G, Valverde JC. 2018. Deep learning for forest species identification based on macroscopic images. Biodiversity Information Science and Standards 2: e25261. DOI: 10.3897/biss.2.25261.
Musah RA, Espinoza EO, Cody RB, Lesiak AD, Christensen ED, Moore HE, Maleknia S, Drijfhout FP. 2015. A high throughput ambient mass spectrometric approach to species identification and classification from chemical fingerprint signatures. Sci. Rep. 5(1): 11520. DOI: 10.1038/srep11520.
Nasirzadeh M, Khazael AA, Bin Khalid M. 2010. Woods recognition system based on local binary pattern. In: 2nd international conference on computational intelligence, communication systems and networks, Liverpool, UK: 308–313. DOI: 10.1109/CICSyN.2010.27.
Nguyen-Trong K. 2023. Evaluation of wood species identification using CNN-based networks at different magnification levels. Int. J. Adv. Comput. Sci. Appl. 14(4). DOI: 10.14569/IJACSA.2023.0140487.
Nithaniyal S, Newmaster SG, Ragupathy S, Krishnamoorthy D, Vassou SL, Parani M. 2014. DNA barcode authentication of wood samples of threatened and commercial timber trees within the tropical dry evergreen forest of India. PLoS ONE 9(9): e107669. DOI: 10.1371/journal.pone.0107669.
Ojala T, Pietikainen M, Harwood D. 1994. Performance evaluation of texture measures with classification based on Kullback discrimination of distributions. In: Proceedings of 12th international conference on pattern recognition 1: 582–585. DOI: 10.1109/ICPR.1994.576366.
Olschofsky K, Köhl M. 2020. Rapid field identification of CITES timber species by deep learning. Trees, Forests and People 2: 100016. DOI: 10.1016/j.tfp.2020.100016.
Owens F, Ravindran P, Costa A, Rodrigues BP, Chavesta M, Montenegro R, Shmulsky R, Wiedenhoeft A. 2023. Evaluation of test specimen surface preparation on computer vision wood identification. Wood Fiber Sci. 55(2): 176–202. DOI: 0000-0002-5421-3269.
Pan S, Kudo M. 2011. Recognition of porosity in wood microscopic anatomical images. In: Perner P (ed.), Advances in data mining. Applications and theoretical aspects 11: 147–160. Springer, Berlin. DOI: 10.1007/978-3-642-23184-1_12.
Pan X, Qiu J, Yang Z. 2022. Identification of softwood species using convolutional neural networks and raw near-infrared spectroscopy. Wood Mater. Sci. Eng. 18(4): 1338–1348. DOI: 10.1080/17480272.2022.2130822.
Paul MGS, Clostre G. 2018. Deep learning applied to identification of commercial timber species from Peru. In: IEEE XXV international conference on electronics, electrical engineering and computing (INTERCON), Lima, Peru: 1–4. DOI: 10.1109/INTERCON.2018.8526457.
Paula Filho PL, Oliveira LS, Britto AS, Sabourin R. 2010. Forest species recognition using color-based features. In: 20th international conference on pattern recognition, Istanbul, Turkey: 4178–4181. DOI: 10.1109/ICPR.2010.1015.
Pourpanah F, Abdar M, Luo Y, Zhou X, Wang R, Lim CP, Wang X-Z, Wu QMJ. 2022. A review of generalized zero-shot learning methods. IEEE Trans. Pattern Anal. Mach. Intell. 45(4): 405–407. DOI: 10.1109/TPAMI.2022.3191696.
Price ER, Miles-Bunch IA, Gasson PE, Lancaster CA. 2021. Pterocarpus wood identification by independent and complementary analysis of DART-TOFMS, microscopic anatomy, and fluorescence spectrometry. IAWA J. 42(4): 397–418. DOI: 10.1163/22941932-bja10064.
Price ER, McClure PJ, Huffman AN, Voin D, Espinoza EO. 2022. Reliability of wood identification using DART-TOFMS and the ForeST©database: a validation study. Forensic Science International: Animals and Environments 2: 100045. DOI: 10.1016/j.fsiae.2022.100045.
Putz FE, Romero C, Sist P, Schwartz G, Thompson I, Roopsind A, Ruslandi, Medjibe V, Ellis P. 2022. Sustained timber yield claims, considerations, and tradeoffs for selectively logged forests. Proc. Natl. Acad. Sci. Nexus 1(3): pgac102. DOI: 10.1093/pnasnexus/pgac102.
Quinlan JR. 1986. Induction of decision trees. Mach. Learn. 1(1): 81–106. DOI: 10.1007/BF00116251.
Radford A, Kim JW, Hallacy C, Ramesh A, Goh G, et al.2021. Learning transferable visual models from natural language supervision. In: Proceedings of the 38th international conference on machine learning. PMLR 139: 8748–8763. DOI: 10.48550/arXiv.2103.00020.
Ramachandram D, Taylor GW. 2017. Deep multimodal learning: a survey on recent advances and trends. IEEE Signal Process. Mag. 34(6): 96–108. DOI: 10.1109/MSP.2017.2738401.
Ravindran P, Wiedenhoeft AC. 2020. Comparison of two forensic wood identification technologies for ten Meliaceae woods: computer vision versus mass spectrometry. Wood Sci. Technol. 54(5): 1139–1150. DOI: 10.1007/s00226-020-01178-1.
Ravindran P, Wiedenhoeft AC. 2022. Caveat emptor: on the need for baseline quality standards in computer vision wood identification. Forests 13(4): 632. DOI: 10.3390/f13040632.
Ravindran P, Costa A, Soares R, Wiedenhoeft AC. 2018. Classification of CITES-listed and other Neotropical Meliaceae wood images using convolutional neural networks. Plant Methods 14(1): 25. DOI: 10.1186/s13007-018-0292-9.
Ravindran P, Ebanyenle E, Asi A, Bonsu K, Lambog O, Soares R, Costa A, Wiedenhoeft AC. 2019. Image based identification of Ghanaian timbers using the XyloTron: opportunities, risks and challenges. arXiv:. DOI: 10.48550/arXiv.1912.00296.
Ravindran P, Thompson BJ, Soares RK, Wiedenhoeft AC. 2020. The XyloTron: flexible, open-source, image-based macroscopic field identification of wood products. Front. Plant Sci. 11: 1015. DOI: 10.3389/fpls.2020.01015.
Ravindran P, Owens FC, Wade AC, Vega P, Montenegro R, Shmulsky R, Wiedenhoeft AC. 2021. Field-deployable computer vision wood identification of Peruvian timbers. Front. Plant Sci. 12: 647515. DOI: 10.3389/fpls.2021.647515.
Ravindran P, Owens FC, Wade AC, Shmulsky R, Wiedenhoeft AC. 2022a. Towards sustainable North American wood product value chains, part I: computer vision identification of diffuse porous hardwoods. Front. Plant Sci. 12: 758455. DOI: 10.3389/fpls.2021.758455.
Ravindran P, Wade AC, Owens FC, Shmulsky R, Wiedenhoeft AC. 2022b. Towards sustainable North American wood product value chains, part 2: computer vision identification of ring-porous hardwoods. Can. J. For. Res. 52(7): 1014–1027. DOI: 10.1139/cjfr-2022-0077.
Rosa da Silva N, De Ridder M, Baetens JM, Van den Bulcke J, Rousseau M, Martinez Bruno O, Beeckman H, Van Acker J, De Baets B. 2017. Automated classification of wood transverse cross-section micro-imagery from 77 commercial Central-African timber species. Ann. For. Sci. 74: 1–14. DOI: 10.1007/s13595-017-0619-0.
Rosa da Silva N, Deklerck V, Baetens JM, Van den Bulcke J, De Ridder M, et al.2022. Improved wood species identification based on multi-view imagery of the three anatomical planes. Plant Methods 18(1): 79. DOI: 10.1186/s13007-022-00910-1.
Rosenblatt F. 1958. The perceptron: a probabilistic model for information storage and organization in the brain. Psychol. Rev. 65(6): 386.
Ruffinatto F, Crivellaro A. 2019. Atlas of macroscopic wood identification: with a special focus on timbers used in Europe and CITES-listed species. Springer International, Cham. DOI: 10.1007/978-3-030-23566-6.
Ruffinatto F, Crivellaro A, Wiedenhoeft AC. 2015. Review of macroscopic features for hardwood and softwood identification and a proposal for a new character list. IAWA J. 36(2): 208–241. DOI: 10.1163/22941932-00000096.
Salzberg SL. 1994. C4.5: Programs for machine learning by J. Ross Quinlan. Morgan Kaufmann Publishers, Inc., 1993. Mach. Learn. 16(3): 235–240. DOI: 10.1007/BF00993309.
Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. 2017. Grad-CAM: visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision: 618–626.
Shahinfar S, Meek P, Falzon G. 2020. “How many images do I need?” Understanding how sample size per class affects deep learning model performance metrics for balanced designs in autonomous wildlife monitoring. Ecological Informatics 57: 101085. DOI: 10.1016/j.ecoinf.2020.101085.
Simonyan K, Zisserman A. 2015. Very deep convolutional networks for large-scale image recognition. In: 3rd International Conference on Learning Representations (ICLR 2015): 1–14.
Sugiyama J, Hwang SW, Zhai S, Kobayashi K, Kanai I, Kanai K. 2020. Xylarium digital database for wood information science and education (XDD_001). Available online at https://repository.kulib.kyoto-u.ac.jp/dspace/handle/2433/250017. DOI: 10.14989/XDD_001.
Sun Y, Cao Y, Xiong F, Yue X, Qiu J, He X, Zhao F. 2015. The wood slice cell image identification algorithm based on singular value decomposition. J. Comput. Theor. Nanosci. 12(12): 5372–5378. DOI: 10.1166/jctn.2015.4529.
Tan M, Le Q. 2019. EfficientNet: rethinking model scaling for convolutional neural networks. In: International conference on machine learning: 6105–6114.
Tang XJ, Tay YH, Siam NA, Lim SC. 2018. MyWood-ID: automated macroscopic wood identification system using smartphone and macro-lens. In: Conference: CIIS 2018: proceedings of the 2018 international conference on computational intelligence and intelligent systems: 37–43. DOI: 10.1145/3293475.3293493.
Tay YH. 2019. Xylorix: an AI-as-a-service platform for wood identification. In: Proceedings of the IAWA-IUFRO international symposium for updating wood identification, Beijing, China, 20–22 May 2019. DOI: 10.13140/RG.2.2.32456.88325.
Thrun S, Pratt L. 2012. Learning to learn. Springer Science & Business Media, Dordrecht.
Tou JY, Lau PY, Tay YH. 2007. Computer vision-based wood recognition system. In: 2007 proceedings of international workshop on advanced image technology: 197–202.
Tou JY, Tay YH, Lau PY. 2009. A comparative study for texture classification techniques on wood species recognition problem. In: Fifth international conference on natural computation, Tianjian, China: 8–12. DOI: 10.1109/ICNC.2009.594.
UNODC Committee. 2016. Best practice guide for forensic timber identification. United Nations Office on Drugs and Crime, Vienna.
UNODC Committee. 2020. World Wildlife Crime Report 2020: trafficking in protected species. United Nations Office on Drugs and Crime, New York, NY.
van den Oord A, Li Y, Vinyals O. 2019. Representation learning with contrastive predictive coding. arXiv:. DOI: 10.48550/arXiv.1807.03748.
Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser Ł, Polosukhin I. 2017. Attention is all you need. Advances in Neural Information Processing Systems: 5998–6008.
Verly Lopes DJ, dos Santos Bobadilha G, Burgreen GW, Entsminger ED. 2021. Identification of North American softwoods via machine-learning. Can. J. For. Res. 51(9): 1245–1252. DOI: 10.1139/cjfr-2020-0416.
Werbos P. 1974. Beyond regression: new tools for prediction and analysis in the behavioral sciences. Harvard University Press, Cambridge, MA.
Wheeler EA. 2011. Insidewood — a web resource for hardwood anatomy. IAWA J. 32(2): 199–211. DOI: 10.1163/22941932-90000051.
Wheeler EA, Baas P. 1998. Wood identification — a review. IAWA J. 19(3): 241–264. DOI: 10.1163/22941932-90001528.
Wheeler EA, Gasson PE, Baas P. 2020. Using the InsideWood web site: potentials and pitfalls. IAWA J. 41(4): 412–462. DOI: 10.1163/22941932-bja10032.
Wiedenhoeft AC. 2020. The XyloPhone: toward democratizing access to high-quality macroscopic imaging for wood and other substrates. IAWA J. 41(4): 699–719. DOI: 10.1163/22941932-bja10043.
Wiedenhoeft AC, Simeone J, Smith A, Parker-Forney M, Soares R, Fishman A. 2019. Fraud and misrepresentation in retail forest products exceeds U.S. forensic wood science capacity. PLoS ONE 14(7): e0219917. DOI: 10.1371/journal.pone.0219917.
Wu F, Gazo R, Haviarova E, Benes B. 2021. Wood identification based on longitudinal section images by using deep learning. Wood Sci. Technol. 55(2): 553–563. DOI: 10.1007/s00226-021-01261-1.
Wu Z, Xiong Y, Yu SX, Lin D. 2018. Unsupervised feature learning via non-parametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition: 3733–3742.
Yadav AR, Dewal ML, Anand RS, Gupta S. 2013. Classification of hardwood species using ANN classifier. In: Fourth national conference on computer vision, pattern recognition, image processing and graphics (NCVPRIPG), Jodhpur, India: 1–5. DOI: 10.1109/NCVPRIPG.2013.6776231.
Yang X, Jiang H, Ma L, Yang W, Zhao X, Hu C, Ge Z. 2023. Micro image classification of 19 high-value hardwood species based on texture feature fusion. BioResources 18(2): 3373–3386. DOI: 10.15376/biores.18.2.3373-3386.
Yeh C-H, Hong C-Y, Hsu Y-C, Liu T-L, Chen Y, LeCun Y. 2022. Decoupled contrastive learning. In: European conference on computer vision: 668–684.
Yin Y, Wiedenhoeft AC, Donaldson L. 2020. Advancing wood identification — anatomical and molecular techniques. IAWA J. 41(4): 391–392. DOI: 10.1163/22941932-00002150.
Yuliastuti E, Suprijanto Sasi SR. 2013. Compact computer vision system for tropical wood species recognition based on pores and concentric curve. In: 3rd international conference on Instrumentation Control and Automation (ICA): 198–202. DOI: 10.1109/ICA.2013.6734071.
Yusof R, Khalid M, Khairuddin ASM. 2013. Application of kernel-genetic algorithm as nonlinear feature selection in tropical wood species recognition system. Comput. Electron. Agric. 93: 68–77. DOI: 10.1016/j.compag.2013.01.007.
Yusof R, Rosli NR, Khalid M. 2010. Using Gabor filters as image multiplier for tropical wood species recognition system. In: 12th international conference on computer modelling and simulation, Cambridge, UK: 289–294. DOI: 10.1109/UKSIM.2010.61.
Zamri MI, Aan P, Cordova F, Khairuddin ASM, Mokhtar N, Yusof R. 2016. Tree species classification based on image analysis using improved-basic gray level aura matrix. Comput. Electron. Agric. 124: 227–233. DOI: 10.1016/j.compag.2016.04.004.
Zhang M, Zhao G, Guo J, Wiedenhoeft AC, Liu CC, Yin Y. 2019. Timber species identification from chemical fingerprints using direct analysis in real time (DART) coupled to Fourier transform ion cyclotron resonance mass spectrometry (FTICR-MS): comparison of wood samples subjected to different treatments. Holzforschung 73(11): 975–985. DOI: 10.1515/hf-2018-0304.
Zhang P, Dai X, Yang J, Xiao B, Yuan L, Zhang L, Gao J. 2021. Multi-scale vision longformer: a new vision transformer for high-resolution image encoding. In: IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada: 2978–2988. DOI: 10.1109/ICCV48922.2021.00299.
Zhao P, Dou G, Chen G-S. 2014. Wood species identification using feature-level fusion scheme. Optik 125(3): 1144–1148. DOI: 10.1016/j.ijleo.2013.07.124.
Zhou B, Khosla A, Lapedriza A, Oliva A, Torralba A. 2016. Learning deep features for discriminative localization. In: 2016 IEEE conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA: 2921–2929. DOI: 10.1109/CVPR.2016.319.
Zhu M, Wang J, Wang A, Ren H, Emam M. 2021. Multi-fusion approach for wood microscopic images identification based on deep transfer learning. Appl. Sci. 11(16): 7639. DOI: 10.3390/app11167639.
Footnotes
Edited by Yafang Yin
