Deep Learning-based In-situ Coniferous Wood Identification of Components in Heritage Architectures of China

Deep Learning-based In-situ Coniferous Wood Identification of Components in Heritage Architectures of China | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Deep Learning-based In-situ Coniferous Wood Identification of Components in Heritage Architectures of China Chang Zheng, Lichao Jiao, Tuo He, Yang Lu, Shoujia Liu, Tianxiao Li, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6178164/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 28 Oct, 2025 Read the published version in npj Heritage Science → Version 1 posted 10 You are reading this latest preprint version Abstract Wood identification of structural components is crucial for heritage architecture conservation, elucidating utilization patterns of forest resources and revealing evolution of civilization in the history. This study first proposed a computer vision-based in-situ identification method using 63 wooden components and 4050 digital images obtained from nine representatives of historical heritage architectures in China. The optimal algorithm, RepLKNet, which was developed on the training dataset constructed by collected images from xylarium specimens of coniferous wood ( Abies , Larix , Picea , and Pinus ), achieved a wood identification accuracy of 96.67%, with average sample precisions of 93.33% and 90% at confidence levels of 70% and 90% respectively for the components of heritage architectures. The minimum sample size requirements for constructing an effective model were determined to be 25 wood specimens and 1500 images per genus, as validated by real-world testing with an accuracy exceeding 90%. Meanwhile, this study investigated the impact of two common deterioration types in wooden components of heritage construction — decay and crack — as well as their severity, on the identification accuracy of the proposed method. The results demonstrate that crack exert a more significant impact on the wood recognition accuracy of historical components compared to decay. Specifically, when the cracked area exceeds 30% of the captured image area, the model’s identification accuracy experiences a sharp decline. Furthermore, the integrity of latewood features plays a crucial role in wood identification accuracy, particularly when compared to the earlywood region within the growth ring. The computer vision-driven methodology for in-situ identification and assessment of wooden components proposed and implemented in this investigation contributes to the advancement of structural preservation strategies, and preventive maintenance practices in heritage architectures. Heritage architecture Wood identification Genus Computer vision Latewood Wood anatomy Wood crack Wood decay Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 1 Introduction Wooden heritage architectures serve as crucial material embodiments of human ingenuity, encapsulating profound historical, artistic, and scientific significance 1 . Subject to the intrinsic properties of wood, environmental conditions, and anthropogenic influences, wooden heritage architectures, having endured centuries of service, are prone to various forms and extents of deterioration, including crack and decay, which compromise the structural integrity and safety. In accordance with the Principles for the Conservation of Wooden Built Heritage 2 and Chinese national standard GB/T 50165-2020 - Technical Standard for Maintenance and Strengthening of Historical Timber Buildings 3 , the conservation and reinforcement of heritage wooden structures must adhere to the principles of preserving the original form, structure, materials, and craftsmanship. When restoration or replacement of original wooden components is necessitated, the timber utilized should, as far as practicable, be of the same species as the original. Consequently, the precise identification for wooden components is foundational to the effective maintenance and safeguarding of heritage wooden structures 4 . Additionally, a comprehensive understanding of the timber species employed in the wooden frameworks of historic buildings can illuminate the principles governing timber selection and utilization across different epochs and regions, thereby offering insights into the evolution of societal civilization in the realms of forest resource exploitation and architectural construction techniques 5 . Currently, the wood identification of wooden heritage architectures predominantly relies on the wood anatomy method, which is primarily used for genus-level identification 6,7 . This method necessitates destructive sampling of wood components, which inevitably damages the selected elements. The process involves transporting samples to the laboratory, where they undergo a series of steps including sample preparation, softening, sectioning, staining, microscopic observation, and characteristic analysis. This procedure is not only highly specialized and labor-intensive but also time-consuming, often resulting in a prolonged identification cycle 8 . For historical wooden structures, comprehensive sampling of all components significantly increases on-site workload. Moreover, due to the unique historical value of many heritage structures, extensive destructive sampling is often impractical. Additionally, the accuracy of this method heavily depends on the expertise and subjective judgment of the appraisers, which can introduce variability and bias into the identification results. On the other hand, in-situ wood determination in heritage architecture remains a significant challenge, despite the gradual application of advanced techniques such as DNA barcoding 9-12 and chemical fingerprinting 13 . DNA barcoding, while promising, still requires a laboratory setting and involves complex procedures such as sampling, sample processing, and nucleic acid extraction. These requirements make it unsuitable for in-situ identification and non-sampled sampling of wood components 14 . Furthermore, wood components in historical buildings often undergo deterioration processes such as decay and degradation, which can alter their chemical composition. This poses a challenge for the effective application of chemical fingerprint markers, as the original chemical profiles may no longer be reliable 15 . Consequently, there is a pressing need for the development of non-sampled, efficient, and accurate methods for wood identification in the field of heritage conservation. The application of computer vision technology for wood identification has seen remarkable progress, showcasing significant potential in modern wood identification due to rapid advancements in hardware and artificial intelligence algorithms. This approach is characterized by its portability, accuracy, speed, and cost-effectiveness 16-18 . Typically, this method relies on the construction of a large-scale dataset. However, the unique cultural value of heritage architectures and the limited availability of wooden components pose significant challenges. The image data of wooden heritage architectures that can be collected is often extremely scarce, making it difficult to meet the data volume requirements essential for effective computer vision recognition 19 . Additionally, compared to modern wooden materials, the surfaces of wooden heritage architectures frequently exhibit defects such as decay, crack, and aging. These imperfections significantly hinder the image acquisition process and reduce the accuracy of component recognition 20, 21 . To the best of our knowledge, no existing research has yet explored the use of computer vision technology for the wood identification in heritage architectures, highlighting a critical gap in the field. Chinese traditional architecture boasts a long and illustrious history, characterized by a construction methodology that predominantly employs wood structures, often in combination with earth, masonry, and stone. Among these, wood structures stand out as one of the most distinctive and defining features of traditional Chinese architecture. The selection, processing, and utilization of wooden heritage architectures reflect the profound technical and cultural wisdom of ancient craftsmen, who adeptly leveraged the inherent properties of wood. This not only demonstrates the advanced wood construction technologies of the time but also embodies the social consciousness and aesthetic values of the era. Historical literature and field research have consistently indicated that the Pinaceae family is extensively utilized in Chinese wooden heritage architectures, with a notably high prevalence 6, 22-24 . In light of this, the present study focuses on four genera within the Pinaceae family— Abies , Larix , Picea , and Pinus —which are frequently identified in Chinese wooden heritage architectures. These genera were selected as the primary research subjects to explore the feasibility of in-situ wood identification using computer vision technology. By leveraging this innovative approach, the study aims to develop a non-sampled, efficient, and accurate method for identifying wood in heritage architectures, thereby contributing to the preservation and understanding of these invaluable cultural treasures. To address the challenge of large-scale image collection and database construction for wooden components in heritage architectures, this study proposes a novel deep learning-based identification model. This model was initially trained using an image database of xylarium wood specimens. Subsequently, a test set comprising images of wooden heritage architectures was established to evaluate the model's performance in wood identification. The objectives of this study are as follows: • To investigate the viability employing xylarium wood specimen models to identify components in wooden heritage architectures. • To explores the minimum sample size required for xylarium wood specimens to achieve accurate identification of the components in wooden heritage architectures. • To propose a grading criterion for wood crack and decay based on image analysis, and determine the impact of these defects on model recognition accuracy. 2 Materials and methods The specific process is shown in Fig. 1 . It consists of four steps: dataset establishment, model construction, on-site image capture, and wood identification. Furthermore, the degradation level of the model is specified in order to explore the impact of varying levels of degradation on the model's accuracy 2.1 Data preparation The training dataset is constructed by collected images of xylarium wood specimens from the Wood Specimen Resource Center of the National Forestry and Grassland Administration. Transverse end surfaces of the specimens were polished with 240, 400, 600, and 800 sandings in turn to obtain clear surfaces for image collection. The cross-section images of 2048 × 2048 pixels, 8-bit RGB in PNG format, representing 6.35 × 6.35 mm of tissue, were obtained using iWood 25 . Wood defects, including surface crack, blue staining, and knots, were avoided during image collection of xylarium wood specimens. The training dataset contains four genera of Pinaceae family which are Abies , Larix , Picea and Pinus , 481 xylarium wood specimens, and in total 38208 images, distributed across 25 provinces of China [unpublished data]. The test dataset covers 63 sampling components from nine heritage architectures of China, containing 4,050 images of absence of wood defects. As shown in Table 1 , the selected architectural sites include: Jiexiu Houtu Temple in Shanxi (JHT, established 457 AD), Pagoda of Fogong Temple in Shanxi (PFT, established 1056 AD), Chongshan Temple in Shanxi (CT, established 1383 AD), the Forbidden City (FC, established 1420 AD), Dahui Temple in Beijing (DT, established 1513 AD), Chunyang Palace in Shanxi (CP, established 1573 AD), Wanshou Temple in Beijing (WT, established 1577 AD), Financial Street in Beijing (FS, established 1912 AD), and Xuanwu Hospital in Beijing (XH, established 1958 AD). Comprehensive dataset specifications are presented in Table 2 . Table 1 Wood sample sources of historical heritage architectures for image collection. Genus No. of samples Architecture Component type Region Abies 3 FS, XH - Beijing Larix 15 CT a , CP a , FT a , FS a , XH, WT a , FC Rafter, purlin Beijing, Shanxi Province Picea 4 CT a flying rafter Shanxi Province Pinus 41 JHT a , CT a , DT a , XH, FC a , Column, Eaves board, rafter, beam, tile fillet, purlin Beijing, Shanxi Province Note: a National important cultural relic of China. CT = Chongshan Temple, CP = Chunyang Palace, DT = Dahui Temple, FC = the Forbidden City, FS = Financial Street, FT = Pagoda of Fogong Temple, JHT = Jiexiu Houtu Temple, XH = Xuanwu Hospital, WT = Wanshou Temple. Wooden components from heritage architectures undergo specialized sanding procedures that differ from standard xylarium wood specimen preparation to minimize structural damage. The in-situ preparation process involves sequential polishing of transverse end surfaces using a specialized 1 cm diameter grinder with 180, 240, 400, 600 and 800 sanding. This process removes the surface material of the wood components with the thickness of approximately 0.5-1.0 mm, ensuring optimal visibility of wood anatomical features while maintaining structural integrity. Image acquisition was performed using the iWood, which specifically designed to reduce damage as much as possible to the heritage wooden components. All samples were systematically identified to genus level through wood anatomical analysis. The resulting classification data were then used to annotate the acquired images for subsequent analytical processing and deep learning model development. Table 2 Sampling size of images collected from the components of heritage architectures. Genus Samples numbers Image numbers Absent of defect Crack Decay Abies 3 198 12 0 Larix 15 707 103 114 Picea 4 215 59 24 Pinus 41 2930 686 392 Total 63 4050 860 530 2.2 Deterioration classification The primary forms of deterioration observed in wooden heritage architectures encompass surface changes 26 , crack 21 , mechanical deformation 27 , mechanical damage, insect infestation 27 , decay 28 , and biological growth. During the cross-sectional preparation process required for image acquisition, certain deterioration patterns — particularly insect infestation, decay, and crack — may become more pronounced in the collected data. Given the prevalence and structural significance of crack and decay as key degradation mechanisms in wooden components, these specific deterioration types were selected for detailed analysis. The wood identification accuracy of components is significantly influenced by varying degrees of deterioration. While previous studies have established comprehensive grading systems for decay states 29 , these classifications are primarily designed for macroscopic assessment and prove inadequate for analyzing small-scale areas captured in individual images. To address this limitation, a comprehensive five-level classification system was implemented according to crack and decay characteristics. The status and description of crack and decay classification are shown in Table 3 , and the illustration of crack and decay relevant to various levels are shown in Fig. 2 . Table 3 Criterion for crack and decay grading of image data acquired from wooden heritage architecture. Deterioration Crack Decay Level Level of crack Features Level of decay Features Level-0 c0: Material in good condition no crack was found within the view of the collector d0: Material in good condition no decay was found within the view of the collector Level-1 c1: Minor crack the area of crack not exceeding 10% of the area within the collector's field of view d1: Minor decay the area of decay not exceeding 10% of the area within the collector's field of view Level-2 c2: Obviously crack the area of crack between 10% and 30% of the collector's field of view d2: Obviously decay the area of decay between 10% and 30% of the collector's field of view Level-3 c3: Serious crack the area of crack between 30% and 60% of the collector's field of view d3: Serious decay the area of decay between 30% and 60% of the collector's field of view Level-4 c4: Damaged crack the area of crack of 60% or more of the area within the collector's field of view d4: Damaged decayed the area of decay of 60% or more of the area within the collector's field of view 2.3 Calculation of degradation area and degradation simulation Traditional image area quantification typically relies on pixel value analysis of target regions 30 , conventionally performed through manual measurement — a process characterized by significant time requirements. The inherent morphological irregularity of wood deterioration patterns poses substantial challenges for accurate area quantification using conventional methods. Semi-automated annotation techniques offer a viable solution for enhancing computational efficiency while maintaining measurement precision. For the deterioration data, we used ISAT (An Interactive Semi-Automatic Annotation Tool) 31 for semi-automatic labelling. The method is used in conjunction with the SAM (Segment Anything Model) 32 to rapidly and accurately identify instances of deterioration types. Instance segmentation outputs were processed through a threshold-based binarization algorithm, with deterioration area quantification achieved through pixel analysis using numpy.sum (threshold = 255) in Python. This computational pipeline enabled precise calculation of deterioration area percentages based on white pixel distribution within segmented regions. The methodological workflow is illustrated in Fig. 3 , while the distribution of sample and image number across different deterioration levels is presented in Table 4 . Table 4 Distribution of sample and image numbers across different deterioration levels of crack and decay for the components in heritage architecture. Sample numbers Image numbers Crack Decay Crack Decay Level-1 39 38 631 245 Level-2 27 29 157 136 Level-3 5 20 19 91 Level-4 - 12 - 52 total - - 860 530 The dataset comprises 860 crack images and 530 decay images of wood cross-sections, as detailed in Table 4 . While the decay image collection is numerically smaller, its distribution across severity levels (c1-c4) demonstrates greater uniformity compared to crack images, which predominantly cluster at c1 and c2 levels with significant underrepresentation at c3 and c4. To investigate the impact of cracking on wood identification accuracy, a subset of 3,403 cross-section images from 52 samples was selected for simulated crack generation at c3 and c4 severity levels. As illustrated in Fig. 2 , crack features typically manifest as irregular black patterns in cross-sectional images. The simulation process employed instance segmentation region with crack morphology replicated by setting instance segmentation region pixels to 0 while preserving surrounding areas, as demonstrated in Fig. 4 . 2.4 model construction for wood identification The resurgence of large kernel models in computer vision, facilitated by advancements in computational hardware, has demonstrated superior predictive accuracy in recent studies 33 . Conventional deep models exhibit limited effective receptive fields 33, 34 . In contrast, large kernel architectures provide substantially expanded receptive fields that more closely approximate human perceptual characteristics. The deterioration features of wood will affect the model's judgment to varying levels. A large receptive field is more likely to detect the length correlation between wood features, thus facilitating accurate judgement. As a representative implementation, RepLKNet 33 exemplifies the large kernel convolutional neural network architecture, with its structural configuration illustrated in Fig. 5 . The model structure of RepLKNet is relatively simple. Following the input of image data, the image is processed by the stem module, which consists two convolutional layers and two depthwise separable convolutions. In the following four Stages, there is a preponderance of RepLK Blocks and ConvFFN modules. The majority of the large kernel reflected in the RepLK Block. Finally, the model downsamples through the Transition module. 2.5 Evaluation standards In target identification, the criterion for evaluating a model is its accuracy. The following equations 1 exemplify this: TP (True Positive) indicates that the target was identified correctly as a positive sample, and FP (False Positive) suggests that the target was identified as a positive sample but was a negative sample. FN (False Negative) indicates that the target was a negative sample, but was a positive sample. TN (True Negative) suggests that the target was identified correctly as a negative sample. However, accuracy alone does not sufficiently evaluate model performance in this context, as wooden component identification typically requires multiple image acquisitions per sample. To address this requirement, we introduced confidence metrics to quantify model performance at the sample level. Sample confidence and precision were mathematically defined in Equations 2 and 3, respectively. A classification was considered correct when sample confidence values exceeded empirically determined thresholds of 0.7 or 0.9. 2.6 Experimental setup All the above steps were implemented on a workstation (CPU: Intel Xeon Silver 4210R @ 2.4 GHz, RAM: 64 GB, and GPUs: NVIDIA GeForce GTX 3090). All the implementations of models are based on python 3.8, cuda 11.8 and PyTorch 2.0. 3. Results and discussion 3.1 In-situ wood identification of components in heritage architectures The development of computer vision-based identification methods for wooden heritage architectures presents significant challenges due to two primary constraints: the difficult destructive sampling properties due to unique cultural values, and the scarcity of wooden components. To address these limitations, we implemented a wood identification model initially developed using xylarium wood specimen image databases. This model, which has demonstrated robust generalization capabilities through validation with unknown modern wood samples [unpublished data], was subsequently adapted and applied to the wood identification of components in heritage architectures. The experimental results demonstrate satisfactory performance across multiple evaluation metrics. Figure 4 presents the confusion matrix and corresponding confidence levels for the RepLKNet classification outcomes. Notably, the model achieved exceptional recall rates exceeding 95% for Pinus and Larix (Fig. 6 a), two genera predominantly utilized in Chinese heritage architecture. In contrast, classification performance for Abies and Picea yielded lower recall rates of 58.08% and 70.70%, respectively. This performance discrepancy can be primarily attributed to misclassifications associated with specific samples, as detailed in Table 5 . Table 5 Wood identification results with an accuracy rate of less than 90% for the components of heritage architectures. Samples Genus Image numbers Confidence Distribution of predicted results Abies Larix Picea Pinus CSS02 Pinus 16 56.25% 0 7 0 9 CSS03 Pinus 14 78.57% 0 3 0 11 WA01 Picea 28 89.28% 0 0 25 3 WA05 Picea 41 60.97% 0 0 25 16 MY02 Abies 92 39.13% 36 0 53 3 MY13 Abies 68 61.76% 42 0 23 3 Table 6 Sample precision of the RepLKNet model at the confidence of 70% and 90% respectively for the components of heritage architectures. Genus Sample precision 90% confidence 70% confidence Abies 33.33% 33.33% Larix 100% 100% Picea 50% 75% Pinus 95.12% 97.56% Average 90% 93.33% In practical identification, the reliability of the model cannot be evaluated on accuracy alone. Comprehensive performance analysis necessitates individual sample confidence calculations, as demonstrated in Fig. 6 (b). As shown in Table 6 , among 60 samples with absence of deterioration, 54 (90%) achieved confidence levels exceeding 90%, while sample precision increased to 93.33% when applying a 70% confidence threshold. In the identification result of Picea , a total of 55 images were erroneously identified as Pinus , while other 8 images were incorrectly identified as Larix . For sample WA05, it had the lowest confidence level below 70% among Picea samples, and a total of 16 images were incorrectly identified as Pinus (Table 5 ). It is speculated that the reason may be due to the less difference in wood structure characteristics of Pinaceae wood in cross-section, and the change in the morphology of the tracheids during the transition from earlywood to latewood in the cross section of Picea and soft pine (Subgen. Strobus ), one type of genus Pinus , is similar, and both of them are gradual. Therefore, the identification of Picea and soft pine (Subgen. Strobus ) woods need to be carried out with the help of the main characteristics of cross-field pitting pattern on the radial section. The erroneous data for the Abies mainly comes from two samples, MY02 and MY13. From the identification results, it can be concluded that Abies is mainly prone to confusion with Picea . In addition, only two of the 41 Pinus wood samples showing confidence levels below 90%. Given the high sample precision obtained in this study, the RepLKNet model indicates strong potential for application to the accurate wood identification of wooden heritage architectures. This scheme will effectively avoid the difficulties of computer vision methods that require the establishment of large-scale image databases of wooden components of wooden heritage architectures, and at the same time, it can also realize the rapid in-situ wood identification for wooden heritage architectures, and reduce the destructive sampling and the cycle of wood identification. 3.2 Minimum number of samples and images for establishing an effective wood identification model While image data from xylarium wood specimens have proven effective for wood identification in heritage architecture applications, this methodology typically requires substantial specimen collections. The limited availability of xylarium wood specimens presents significant challenges for model implementation and methodological application. A critical research question is regarding the minimum specimen and image requirements for achieving reliable wood identification accuracy (> 90%). Previous studies have discussed the number of xylarium wood specimens and the number of images separately 15 , but we believe that there should be a relative relationship between the minimum number of xylarium wood specimens and the minimum number of images. Therefore, this section will discuss the interrelationship between them. Prior to determining the minimum number of xylarium wood specimens required, establishing an adequate image dataset is essential. The number of image acquisition potential varies significantly with specimen dimensions, necessitating initial categorization of specimens by genus and corresponding image quantity. Then, select the xylarium wood specimens in sequence and establish a model, and test the final results, as shown in Fig. 7 . The analysis revealed a positive correlation between specimen quantity and model accuracy, achieving optimal performance (96.2%) at 40 xylarium wood specimens per genus. Notably, accuracy consistently exceeded 90% when xylarium wood specimen numbers surpassed 20 per genus, while the rate of precision improvement diminished beyond this threshold. Consequently, 20 xylarium wood specimens per genus are established as the minimum requirement for reproducible results in this study. While model accuracy generally improves with increased training data volume, this relationship is not strictly linear, particularly for anisotropic wood materials. The observed trend in Fig. 7 demonstrates a deviation from the expected correlation between image quantity and model accuracy, warranting further investigation to establish a definitive relationship. To examine this relationship systematically, comparative analyses were conducted using 20, 25, 30, and 35 xylarium specimens with corresponding training datasets of 1,500, 2,000, 2,500, 3,000, and 3,500 images. Given the physical size variability among xylarium wood specimens, individual specimens may yield fewer images than the calculated average. In such cases, when selecting a 2,000-image dataset from 20 xylarium wood specimens, for instance, specimens with limited image availability (< 100 images) were supplemented by random image selection from other xylarium wood specimens within the same genus to maintain dataset integrity. As shown in Fig. 8 , the number of xylarium wood specimens in the model directly affects the accuracy of the image. By controlling the number of images in each category and comparing different xylarium wood specimen sizes, we found that the overall accuracy increases with the increase of xylarium wood specimens, and the overall accuracy of the model also increases with the increase of image numbers. Notably, at image numbers of 2,000 and 2,500, models trained on 25 xylarium wood specimens marginally outperformed those using 30 xylarium wood specimens. However, this relationship reversed at higher image quantities, with 30-specimen models demonstrating superior accuracy. Analysis of fixed specimen quantities revealed that models trained on 20 or 25 xylarium wood specimens reached performance plateaus at approximately 2,000 images, suggesting this image volume sufficiently captures the representative features of these specimen sets. Beyond this threshold, it is difficult to improve accuracy with additional images without incorporating new xylarium wood specimens. The 20-specimen model exhibited overfitting beyond 2,500 images, despite achieving > 90% accuracy. Based on these findings, we recommend a minimum dataset of 1,500 images from 25 xylarium wood specimens per genus as optimal for maintaining > 90% accuracy while preventing overfitting. The recommended dataset of 1,500 images from 25 xylarium wood specimens should preferentially absence of deterioration samples, without strict image quantity constraints per individual specimen. Previous research has controlled the number of images collected, and trained models on a large number of xylarium wood specimens, but collected very few images per specimen 35 . The restricted image sampling may inadequately represent specimen characteristics, as collectors cannot reliably determine whether minimal images sufficiently capture a specimen's full feature. Moreover, this methodology necessitates substantial xylarium wood specimen resources and proves challenging to replicate. Since it is difficult to obtain xylarium wood specimens that have a clear background in plant taxonomy, we believe that the principle of "collecting as many images as possible" should be followed to reduce the need to collect a large number of xylarium wood specimens. Although excessive acquisition of images could theoretically induce overfitting, the practical threshold for such occurrences remains substantially high and scales with specimen quantity. This image acquisition principle enhances the reproducibility of computer vision-based wood identification methods, ensuring more consistent and reliable application outcomes. 3.3 The impact of wood deterioration on identifying model accuracy During prolonged service duration, wood architectural heritage components exhibit inherent susceptibility to hygrothermal fluctuations, sustained creep deformation under mechanical loads, and microbiological degradation mechanisms, cumulatively manifesting as characteristic deterioration patterns including crack propagation, dimensional instability, and biodeterioration of wooden components 20 . Table 7 Identification accuracy for different levels of deterioration crack decay Level-1 93.66% 98.37% Level-2 92.99% 90.44% Level-3 81.58% 90.87% Level-4 71.17% 88.45% The confidence levels of deteriorated samples were systematically evaluated and compared against non-deteriorated specimens, as illustrated in Fig. 9 and Fig. 10 . Samples exhibiting severe crack (c3 and c4) demonstrated substantially reduced confidence levels, significantly impacting identification accuracy of the model, as shown in Table 7 . In contrast, decayed samples showed minimal confidence reduction. These findings align with previous research identifying tracheid morphology transitions between earlywood and latewood as primary identification features 37 . The current study further emphasizes the critical role of latewood anatomical feature integrity within growth rings for accurate wood identification. Wood decay primarily refers to the biochemical degradation of wood mediated by microbial enzymatic activity 38 . Among these microorganisms, fungi constitute the most significant degradative agents, categorized into white rot and brown rot based on their distinct decay mechanisms 39 . Research on coniferous wood decomposition has established characteristic patterns: brown rot fungi selectively depolymerize cellulose while leaving a modified lignin matrix, whereas white rot fungi demonstrate simultaneous lignocellulosic degradation through oxidative enzymatic systems 40 . Growth of fungi causing rot decay was limited and slower in latewood than in earlywood due to the narrow cell lumen, thicker wall, and higher density of latewood 41 . Meanwhile, when sanding the surface of wooden components of heritage architectures, due to the shallow sanding depth, the latewood cells with higher strength are more easily exposed than the earlywood cells. It is for these reasons that the state of preservation and presentation of latewood cells within a growth ring of wood components are more favourable, and also explains why the deterioration type of wood decay has less influence on the accuracy of identification. The impact of crack on identification accuracy is particularly significant, primarily attributable to the compromised integrity of latewood features. Most of the crack in wooden components of heritage architectures are dry shrinkage cracking, which is caused by the dry shrinkage and wet swelling characteristics, and anisotropy of wood 42 . Wood is a natural anisotropic material with hygroscopic and desorptive capabilities, and undergoes drying and wetting with changes in environmental temperature and humidity. At the microscopic level, drying crack often appear first in wood ray tissues, because ray tissues are mostly thin-walled cells, with low strength, and is not connected closely enough with the surrounding cells. When the wood shrinks, the ray cells undergo significant deformation and are damaged by the drying stresses, resulting in crack 43,44 . The cell wall of the latewood is usually thicker, and the amount of drying deformation is larger than that of the earlywood cells. The ray tissue of the latewood is first destroyed by the shrinkage stress that produce small crack, which then gradually expend along the ray tissues, leading to a gradual increase in depth and width 44,46 . This phenomenon results in frequent crack formation within latewood regions, consequently compromising the identification accuracy of the model. 4. Conclusion In this study, a deep learning-based in-situ wood identification framework of components in wooden heritage architecture is firstly proposed. The constructed identification model based on image database of xylarium wood specimens were directly applied to the wooden heritage architecture, which effectively avoids the database construction problems of wooden components that are difficult to achieve large-scale image acquisition. At the same time, the model can be successfully applied to the rapid in-situ identification of wood in wooden heritage architectures, which reduces the destructive sampling and the identification cycle. The optimal algorithm RepLKNet achieves 96.67% accuracy in wood identification. The sample precisions were 93.33% and 90% at 70% and 90% confidence levels, respectively. The minimum xylarium wood specimen size and number of images required to build an effective model are 25 xylarium wood specimens and 1500 images per genus, respectively, for real-world test accuracies exceeding 90%. In addition, wood deterioration can have an impact on the wood identification accuracy of wooden heritage architectures. When the cracked area exceeds 30% of the collected image area, the recognition accuracy of the model decreases dramatically. Future work will focus on optimizing method efficiency and accuracy, particularly for components with varying types and levels of deterioration, while extending applicability from softwood to hardwood component specimens. This approach aims to establish adaptive conservation protocols for historic timber structures, ensuring scientific preservation and targeted restoration strategies across diverse material conditions. Declarations Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Acknowledgements This study was supported financially by the National Key Research and Development Program of China (Grant No. 2023YFF0906301) and the National Science & Technology Fundamental Resources Investigation Program (Grant No. 2023FY101400). The authors would like to express our gratitude to Professor Xiaomei Jiang of the Research Institute of wood Industry, the Chinese Academy of Forestry for her valuable academic advice, and Professor Yongping Chen, Professor Juan Guo, Mrs. Mingkun Xu, Mr. Yonggang Zhang and Mr. Yu Sun of the Research Institute of wood Industry, the Chinese Academy of Forestry for their technical supports. Author contributions CZ, LJ, HT and YY designed the experiments. CZ, TL and YL prepared the samples for imaging and imaged the specimens. TL, YL, SL and CZ curated the collected dataset. LJ, RY, HZ and YY provided research sources. CZ developed the deep learning models. CZ, YY, SL and LJ analyzed the results. CZ, LJ and YY wrote, reviewed and edited the paper. LJ and YY conducted project administration and supervision. All authors read and approved the final manuscript. References Zhu, H. et al. Wood-derived materials for green electronics, biological devices, and energy applications. Chem. Rev. 116 , 9305-9374 (2016). International Council on Monuments and Sites. Guidelines for the Conservation of Wooden Built Heritage https://jianzhuyichan.tongji.edu.cn/info/1007/1543.htm (2017) Sichuan Institute of building research & Zhongke construction. Technical Standard for Maintenance and Reinforcement of Wooden Structures in Ancient Buildings. GBT501652020 https://zjj.sm.gov.cn/xxgk/fgwj/jsbz/202011/t20201113_1589718.htm (2020) Jiang, X., Yin, Y., & Liu, B. Current Status Development and Prospect of Wood Identification Technology. China Wood Industry 24 , 36-39 (2010). (in Chinese) Jiao, L. et al. 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 , 696-709 (2022). Gasson, P. How precise can wood identification be? Wood anato my's role in support of the legal timber trade, especially CITES. IAWA J. 32 , 137–154. (2011). Dong, M. et al. Wood used in ancient timber architecture in Shanxi Province, China. IAWA J. 38 , 182-200, doi:10.1163/22941932-20170167 (2017). Dormontt, E. E. et al. Forensic timber identification: It's time to integrate disciplines to combat illegal logging. Biol. Conserv. 191 , 790-798 (2015). Hartvig, I., Czako, M., Kjær, E. D., Nielsen, L. R. & Theilade, I. The use of DNA barcoding in identification and conservation of rosewood (Dalbergia spp.). Plos One 10 , e0138231 (2015). Jiao, L. et al. DNA Barcode Authentication and Library Development for the Wood of Six Commercial Pterocarpus Species: the Critical Role of Xylarium Specimens. Sci. Rep-Uk. 8 , doi:10.1038/s41598-018-20381-6 (2018). Yu, M. et al. DNA barcoding of vouchered xylarium wood specimens of nine endangered Dalbergia species. Planta 246 , 1165-1176, doi:10.1007/s00425-017-2758-9 (2017). Lu Y. et al. DNA Methods for Identifying Wood in Ancient Timber Architecture. Chinese Journal of Wood Science and Technology 37 , 12-18 (2023). (in Chinese) Domínguez-Delmás, M. Seeing the forest for the trees: New approaches and challenges for dendroarchaeology in the 21st century. Dendrochronologia 62 , 125731 (2020). Jiao, L., Lu, Y., He, T., Guo, J. & Yin, Y. DNA barcoding for wood identification: Global review of the last decade and future perspective. IAWA J. 41 , 620-643 (2020). Traoré, M., Kaal, J. & Cortizas, A. M. Chemometric tools for identification of wood from different oak species and their potential for provenancing of Iberian shipwrecks (16th-18th centuries AD). J. Archaeol Sci. 100 , 62-73 (2018). He, T. et al. Developing deep learning models to automate rosewood tree species identification for CITES designation and implementation. Holzforschung 74 , 1123-1133 (2020). Ravindran, P., Thompson, B. J., Soares, R. K. & Wiedenhoeft, A. C. The XyloTron: Flexible, Open-Source, Image-Based Macroscopic Field Identification of Wood Products. Front. Plant Sci. 11 , doi:10.3389/fpls.2020.01015 (2020). Ravindran, P., Costa, A., Soares, R. & Wiedenhoeft, A. C. Classification of CITES-listed and other neotropical Meliaceae wood images using convolutional neural networks. Plant Methods 14 , doi:10.1186/s13007-018-0292-9 (2018). Russakovsky, O. et al. Imagenet large scale visual recognition challenge. Int. J. Comput. Vision. 115 , 211-252 (2015). Stratigaki, M. Autofluorescence for the Visualization of Microorganisms in Biodeteriorated Materials in the Context of Cultural Heritage. ChemPlusChem 89 , e202400170 (2024). Li, X., Qian, W. & Chang, L. Analysis of the density of wooden components in ancient buildings by micro-drilling resistance, using information diffusion. BioResources 14 , 5777-5787 (2019). Yin, Y. et al. Research on the Identification of Wood Species Used for Wooden Structures in Southeastern Shanxi Province. World of Antiquity 04 , 33-36 (2010) (in Chinese) Zhang, Q. An Analysis of the History of the Palace of Compassion and Tranquility Complex from the Aspect of Wood Species of Timber Members, The Forbidden City. Heritage Architecture 04 , 1-12 (2020). (in Chinese) Li, S. et al. Research on the identification and configuration of wood components for the main hall of Jianshui Zhilin Temple Cultural Relics Protection and Archaeological Science. Sciences of Conservation and Archaeology 32 , 91-98 (2020). (in Chinese) He, T. et al. iWood: An Automated Wood Identification System for Endangered and Precious Tree Species Using Convolutional Neural Networks. Scientia Silvae Sincae 57 ,152-159 (2021). (in Chinese) Tan, Y. et al. Inspection and Evaluation of Wood Components of Ancient Buildings in the South-Three Courts of the Forbidden City. BioResources. 17 (2022). Ma, X. et al. 3D structural deformation monitoring of the archaeological wooden shipwreck stern investigated by optical measuring techniques. J. Cult. Herit. 59 , 102-112 (2023). Venugopal, P., Junninen, K., Linnakoski, R., Edman, M. & Kouki, J. Climate and wood quality have decayer-specific effects on fungal wood decomposition. Forest. Ecol. Manag. 360 , 341-351 (2016). China Academy of Forestry Wood Industry Research Institute, et al. LY/T 2014-2024 Non-destructive testing method and defects classification for wooden components of ancient buildings. National Forestry and Grassland Administration https://std.samr.gov.cn/hb/search/stdHBDetailed?id=1E5DB4EC8381C2FFE06397BE0A0A7B72 (in Chinese) Ji, M., Zhang, W., Wang, G., Wang, Y. & Miao, H. Online Measurement of Outline Size for Pinus densiflora Dimension Lumber: Maximizing Lumber Recovery by Minimizing Enclosure Rectangle Fitting Area. Forests 13 , 1627 (2022). Ji, S. & Zhang, H. ISAT with Segment Anything: An Interactive Semi-Automatic Annotation Tool v1.10 https://github.com/yatengLG/ISAT_with_segment_anything (2025) Kirillov, A. et al. Segment anything[C]//in Proceedings of the IEEE/CVF International Conference on Computer Vision. 4015-4026 (2023). Ding, X., Zhang, X., Han, J. & Ding, G. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns[C]//in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11963-11975 (2022). Ding, X. et al. UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition[C]//in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5513-5524 (2024). Ravindran, P., Owens, F. C., Wade, A. C., Shmulsky, R. & Wiedenhoeft, A. C. Towards sustainable North American wood product value chains, part I: computer vision identification of diffuse porous hardwoods. Front. Plant Sci. 12 , 758455 (2022). Zhu, Q., Zhou, X., Tan, J. & Guo, L. Knowledge base reasoning with convolutional-based recurrent neural networks. Ieee T. Knowl. Data En. 33 , 2015-2028 (2019). Pieter, B. et al. IAWA List of microscopic features for softwood identification. IAWA J. 25 , 1-70 (2004). Ayrilmis, N., Kaymakci, A. & Güleç, T. Potential use of decayed wood in production of wood plastic composite. Ind. Crop Prod. 74 , 279-284 (2015). Green, F. & Highley, T. L. Mechanism of brown-rot decay: paradigm or paradox. Int. Biodeter. Biodegr. 39 , 113-124 (1997). Blanchette, R. A. Wood decay: a submicroscopic view. J. Forest 78 , 734-737 (1980). Besma, B., Ahmed, K., & Yves, B. Effects of biodegradation by brown-rot decay on selected wood properties in eastern white cedar (Thuja occidentalis L.). Int. Biodeterior. Biodegrad. 87 , 87-98 (2014). Chen, Y., Huang, A., Meng, S., & Ni, L. Research Progress in Influencing Factors to Wood Deformation and Cracking and Anti-cracking Measures. World Forestry Research 37 , 71-76 (2024) (in Chinese) Wang, H. H. & Youngs, R. L. Drying stress and check development in the wood of two oaks. IAWA J. 17 , 15-30 (1996). Yamamoto, H., Sakagami, H., Kijidani, Y. & Matsumura, J. Dependence of microcrack behavior in wood on moisture content during drying. Adv. Mater. Sci. Eng. 2013 , 802639 (2013). Gao, Y. et al. The formation mechanism of microcracks and fracture morphology of wood during drying. Dry. Technol. 41 , 1268-1277 (2023). Sakagami, H. Microcrack propagation in transverse surface from heartwood to sapwood during drying. J. Wood Sci. 65 , 33 (2019). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 28 Oct, 2025 Read the published version in npj Heritage Science → Version 1 posted Editorial decision: Revision requested 22 May, 2025 Reviews received at journal 09 May, 2025 Reviews received at journal 07 May, 2025 Reviewers agreed at journal 28 Apr, 2025 Reviewers agreed at journal 27 Apr, 2025 Reviewers agreed at journal 25 Apr, 2025 Reviewers invited by journal 17 Mar, 2025 Editor assigned by journal 16 Mar, 2025 Submission checks completed at journal 16 Mar, 2025 First submitted to journal 07 Mar, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6178164","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":433200934,"identity":"a0ec5d4a-21bc-4757-8146-0aa4d5daab04","order_by":0,"name":"Chang Zheng","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Chang","middleName":"","lastName":"Zheng","suffix":""},{"id":433200935,"identity":"7884e10f-88c7-41e2-9a11-0e3cd16b91a0","order_by":1,"name":"Lichao Jiao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAu0lEQVRIiWNgGAWjYBACxmYg8YDBhoGB+QApWhIY0hgY2BJIsSqB4TAJWpjbmR8+SKg5L8/fxsD44QeDXR4RDmMzNkg4dttwxjEGZskehuRiIrQwmEkkNtxOYLjfwCDNwHAgsYGwFvZvQC3nEuSBtvwmUgsPyJYDCQbHGNiItYWnGOiXZMONxxjbLHsMkglrMew/vvHBhxo7ebljzIdv/KiwI0ILQgUjkGlASD0QyBOhZhSMglEwCkY6AACafjg9ilktMQAAAABJRU5ErkJggg==","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":true,"prefix":"","firstName":"Lichao","middleName":"","lastName":"Jiao","suffix":""},{"id":433200936,"identity":"716503e0-1276-4f4d-a2da-353d8d7b117d","order_by":2,"name":"Tuo He","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Tuo","middleName":"","lastName":"He","suffix":""},{"id":433200939,"identity":"f9e268a9-6335-432d-9fa4-0ee4d79e56f8","order_by":3,"name":"Yang Lu","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Lu","suffix":""},{"id":433200942,"identity":"1cfc5f08-518f-43a8-99d2-9e9164381d66","order_by":4,"name":"Shoujia Liu","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Shoujia","middleName":"","lastName":"Liu","suffix":""},{"id":433200945,"identity":"214e2f9e-b694-44a4-8fdc-0adc868f7450","order_by":5,"name":"Tianxiao Li","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Tianxiao","middleName":"","lastName":"Li","suffix":""},{"id":433200947,"identity":"ff765c7b-8170-40fb-91e3-01ca8912ffcd","order_by":6,"name":"Ruochen Ye","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Ruochen","middleName":"","lastName":"Ye","suffix":""},{"id":433200948,"identity":"e86eab0e-98a0-40d0-9540-39aa191de737","order_by":7,"name":"Haibin Zhou","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Haibin","middleName":"","lastName":"Zhou","suffix":""},{"id":433200949,"identity":"75397658-ff52-45ee-b433-db5171fb8dcc","order_by":8,"name":"Yafang Yin","email":"","orcid":"","institution":"Chinese Academy of Forestry","correspondingAuthor":false,"prefix":"","firstName":"Yafang","middleName":"","lastName":"Yin","suffix":""}],"badges":[],"createdAt":"2025-03-07 12:08:29","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6178164/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6178164/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s40494-025-02120-z","type":"published","date":"2025-10-28T15:58:18+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":79259132,"identity":"dc32a893-20d0-4939-9cc9-c51cd377960b","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":217014,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of wood identification for the components of heritage architecture\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/4422a53b40f118da0cb2011f.png"},{"id":79259133,"identity":"721a1027-070c-4481-95ec-84a900692a23","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":596035,"visible":true,"origin":"","legend":"\u003cp\u003eImage illustration of wood crack and decay at different levels collected from the components of wooden heritage architectures.\u003c/p\u003e\n\u003cp\u003eNote: a-d represent crack levels 1-4 (c1-c4); e-f represent decay levels 1-4 (d1-d4)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/a1a4cacfa6f11cd0c88b104e.png"},{"id":79259131,"identity":"8740ebbd-1c87-4e2b-b30b-d57c6f4935c2","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":94720,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of calculating the proportion of wood deterioration area (crack as illustration) using an Interactive Semi-Automatic Annotation Tool (ISAT).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/47985e1aab8090ea53a8ca4a.png"},{"id":79259137,"identity":"f557a37c-7363-454f-b159-46862998a4a2","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":203936,"visible":true,"origin":"","legend":"\u003cp\u003eImage simulation of wood crack.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/d9b9baa7146577494d43d869.png"},{"id":79259145,"identity":"bbc3d2fd-7ea4-4426-b736-562cb721ae2d","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":150270,"visible":true,"origin":"","legend":"\u003cp\u003eThe network architecture diagram of RepLKNet.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/073eb1d715bb6faca6e4deeb.png"},{"id":79259142,"identity":"177f1594-d690-43d3-9b2c-a41358f608a3","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":81043,"visible":true,"origin":"","legend":"\u003cp\u003eWood identification results of the RepLKNet model. (a) Confusion matrix; (b) Confidence for each specimen\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/6269a696e5af2063442cad27.png"},{"id":79260707,"identity":"0e124edc-266c-41be-90ba-43965f01e057","added_by":"auto","created_at":"2025-03-26 09:25:32","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":81252,"visible":true,"origin":"","legend":"\u003cp\u003eThe relationship between wood identification accuracy and sample numbers.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/48fe803159a0a2861acc3f9d.png"},{"id":79259135,"identity":"b2330dfe-0719-442f-a554-4780dcb5be86","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":130009,"visible":true,"origin":"","legend":"\u003cp\u003eThe relationship between wood identification accuracy and image numbers.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/e95249fa00d08cc8fad62229.png"},{"id":79259144,"identity":"1274f466-095c-4dc1-8397-a34cc4cf57b6","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":215815,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of sample confidence between crack and absence of deterioration.\u003c/p\u003e\n\u003cp\u003eNote: The crack area and non-deterioration area are from the same specimen and are represented by the same specimen number. c1: Crack level-1; c2: Crack level-2; c3: Crack level-3; c4: Crack level-4.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/2fdff6a1ff8d0b1703e78ee2.png"},{"id":79259151,"identity":"d4c7859d-6570-45da-836f-821c33a5098b","added_by":"auto","created_at":"2025-03-26 09:17:32","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":150053,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of sample confidence between decay and absence of deterioration.\u003c/p\u003e\n\u003cp\u003eNote: The decay area and non-deterioration area are from the same specimen and are represented by the same specimen number. d1: Decay level-1; d2: Decay level-2; d3: Decay level-3; d4: Decay level-4. The black triangle represents that the specimen is only collected images with defect (decay).\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/99aacb7dbfb4ee94f2b68b46.png"},{"id":95040027,"identity":"7b77f24e-1686-44ae-a6ac-6e829177cb61","added_by":"auto","created_at":"2025-11-03 16:07:35","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3116030,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6178164/v1/d3342552-319b-487f-a515-afff77c18e65.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning-based In-situ Coniferous Wood Identification of Components in Heritage Architectures of China","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eWooden heritage architectures serve as crucial material embodiments of human ingenuity, encapsulating profound historical, artistic, and scientific significance\u003csup\u003e1\u003c/sup\u003e. Subject to the intrinsic properties of wood, environmental conditions, and anthropogenic influences, wooden heritage architectures, having endured centuries of service, are prone to various forms and extents of deterioration, including crack and decay, which compromise the structural integrity and safety. In accordance with the Principles for the Conservation of Wooden Built Heritage\u003csup\u003e2\u003c/sup\u003e and Chinese national standard GB/T 50165-2020 - Technical Standard for Maintenance and Strengthening of Historical Timber Buildings\u003csup\u003e3\u003c/sup\u003e, the conservation and reinforcement of heritage wooden structures must adhere to the principles of preserving the original form, structure, materials, and craftsmanship. When restoration or replacement of original wooden components is necessitated, the timber utilized should, as far as practicable, be of the same species as the original. Consequently, the precise identification for wooden components is foundational to the effective maintenance and safeguarding of heritage wooden structures\u003csup\u003e4\u003c/sup\u003e. Additionally, a comprehensive understanding of the timber species employed in the wooden frameworks of historic buildings can illuminate the principles governing timber selection and utilization across different epochs and regions, thereby offering insights into the evolution of societal civilization in the realms of forest resource exploitation and architectural construction techniques\u003csup\u003e5\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eCurrently, the wood identification of wooden heritage architectures predominantly relies on the wood anatomy method, which is primarily used for genus-level identification\u003csup\u003e6,7\u003c/sup\u003e. This method necessitates destructive sampling of wood components, which inevitably damages the selected elements. The process involves transporting samples to the laboratory, where they undergo a series of steps including sample preparation, softening, sectioning, staining, microscopic observation, and characteristic analysis. This procedure is not only highly specialized and labor-intensive but also time-consuming, often resulting in a prolonged identification cycle\u003csup\u003e8\u003c/sup\u003e. For historical wooden structures, comprehensive sampling of all components significantly increases on-site workload. Moreover, due to the unique historical value of many heritage structures, extensive destructive sampling is often impractical. Additionally, the accuracy of this method heavily depends on the expertise and subjective judgment of the appraisers, which can introduce variability and bias into the identification results.\u003c/p\u003e\n\u003cp\u003eOn the other hand, in-situ wood determination in heritage architecture remains a significant challenge, despite the gradual application of advanced techniques such as DNA barcoding\u003csup\u003e9-12\u003c/sup\u003e and chemical fingerprinting\u003csup\u003e13\u003c/sup\u003e. DNA barcoding, while promising, still requires a laboratory setting and involves complex procedures such as sampling, sample processing, and nucleic acid extraction. These requirements make it unsuitable for in-situ identification and non-sampled sampling of wood components\u003csup\u003e14\u003c/sup\u003e. Furthermore, wood components in historical buildings often undergo deterioration processes such as decay and degradation, which can alter their chemical composition. This poses a challenge for the effective application of chemical fingerprint markers, as the original chemical profiles may no longer be reliable\u003csup\u003e15\u003c/sup\u003e. Consequently, there is a pressing need for the development of non-sampled, efficient, and accurate methods for wood identification in the field of heritage conservation.\u003c/p\u003e\n\u003cp\u003eThe application of computer vision technology for wood identification has seen remarkable progress, showcasing significant potential in modern wood identification due to rapid advancements in hardware and artificial intelligence algorithms. This approach is characterized by its portability, accuracy, speed, and cost-effectiveness\u003csup\u003e16-18\u003c/sup\u003e. Typically, this method relies on the construction of a large-scale dataset. However, the unique cultural value of heritage architectures and the limited availability of wooden components pose significant challenges. The image data of wooden heritage architectures that can be collected is often extremely scarce, making it difficult to meet the data volume requirements essential for effective computer vision recognition\u003csup\u003e19\u003c/sup\u003e. Additionally, compared to modern wooden materials, the surfaces of wooden heritage architectures frequently exhibit defects such as decay, crack, and aging. These imperfections significantly hinder the image acquisition process and reduce the accuracy of component recognition\u003csup\u003e20, 21\u003c/sup\u003e. To the best of our knowledge, no existing research has yet explored the use of computer vision technology for the wood identification in heritage architectures, highlighting a critical gap in the field.\u003c/p\u003e\n\u003cp\u003eChinese traditional architecture boasts a long and illustrious history, characterized by a construction methodology that predominantly employs wood structures, often in combination with earth, masonry, and stone. Among these, wood structures stand out as one of the most distinctive and defining features of traditional Chinese architecture. The selection, processing, and utilization of wooden heritage architectures reflect the profound technical and cultural wisdom of ancient craftsmen, who adeptly leveraged the inherent properties of wood. This not only demonstrates the advanced wood construction technologies of the time but also embodies the social consciousness and aesthetic values of the era. Historical literature and field research have consistently indicated that the Pinaceae family is extensively utilized in Chinese wooden heritage architectures, with a notably high prevalence\u003csup\u003e6, 22-24\u003c/sup\u003e. In light of this, the present study focuses on four genera within the Pinaceae family—\u003cem\u003eAbies\u003c/em\u003e, \u003cem\u003eLarix\u003c/em\u003e, \u003cem\u003ePicea\u003c/em\u003e, and \u003cem\u003ePinus\u003c/em\u003e—which are frequently identified in Chinese wooden heritage architectures. These genera were selected as the primary research subjects to explore the feasibility of in-situ wood identification using computer vision technology. By leveraging this innovative approach, the study aims to develop a non-sampled, efficient, and accurate method for identifying wood in heritage architectures, thereby contributing to the preservation and understanding of these invaluable cultural treasures.\u003c/p\u003e\n\u003cp\u003eTo address the challenge of large-scale image collection and database construction for wooden components in heritage architectures, this study proposes a novel deep learning-based identification model. This model was initially trained using an image database of xylarium wood specimens. Subsequently, a test set comprising images of wooden heritage architectures was established to evaluate the model's performance in wood identification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe objectives of this study are as follows:\u003c/p\u003e\n\u003cp\u003e•\u0026nbsp;To investigate the viability\u0026nbsp;employing xylarium wood specimen models to identify components in wooden heritage architectures.\u003c/p\u003e\n\u003cp\u003e•\u0026nbsp;To explores the minimum sample size required for xylarium wood specimens \u0026nbsp;to achieve accurate identification of the components in wooden heritage architectures.\u003c/p\u003e\n\u003cp\u003e• To propose a grading criterion for wood crack and decay based on image analysis, and determine the impact of these defects on model recognition accuracy.\u003c/p\u003e"},{"header":"2 Materials and methods","content":"\u003cp\u003eThe specific process is shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. It consists of four steps: dataset establishment, model construction, on-site image capture, and wood identification. Furthermore, the degradation level of the model is specified in order to explore the impact of varying levels of degradation on the model\u0026apos;s accuracy\u003c/p\u003e\n\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1 Data preparation\u003c/h2\u003e\n \u003cp\u003eThe training dataset is constructed by collected images of xylarium wood specimens from the Wood Specimen Resource Center of the National Forestry and Grassland Administration. Transverse end surfaces of the specimens were polished with 240, 400, 600, and 800 sandings in turn to obtain clear surfaces for image collection. The cross-section images of 2048 \u0026times; 2048 pixels, 8-bit RGB in PNG format, representing 6.35 \u0026times; 6.35 mm of tissue, were obtained using iWood\u003csup\u003e25\u003c/sup\u003e. Wood defects, including surface crack, blue staining, and knots, were avoided during image collection of xylarium wood specimens. The training dataset contains four genera of Pinaceae family which are \u003cem\u003eAbies\u003c/em\u003e, \u003cem\u003eLarix\u003c/em\u003e, \u003cem\u003ePicea\u003c/em\u003e and \u003cem\u003ePinus\u003c/em\u003e, 481 xylarium wood specimens, and in total 38208 images, distributed across 25 provinces of China [unpublished data].\u003c/p\u003e\n \u003cp\u003eThe test dataset covers 63 sampling components from nine heritage architectures of China, containing 4,050 images of absence of wood defects. As shown in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, the selected architectural sites include: Jiexiu Houtu Temple in Shanxi (JHT, established 457 AD), Pagoda of Fogong Temple in Shanxi (PFT, established 1056 AD), Chongshan Temple in Shanxi (CT, established 1383 AD), the Forbidden City (FC, established 1420 AD), Dahui Temple in Beijing (DT, established 1513 AD), Chunyang Palace in Shanxi (CP, established 1573 AD), Wanshou Temple in Beijing (WT, established 1577 AD), Financial Street in Beijing (FS, established 1912 AD), and Xuanwu Hospital in Beijing (XH, established 1958 AD). Comprehensive dataset specifications are presented in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eWood sample sources of historical heritage architectures for image collection.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eGenus\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNo. of samples\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eArchitecture\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eComponent type\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRegion\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eAbies\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFS, XH\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBeijing\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eLarix\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCT\u003csup\u003ea\u003c/sup\u003e, CP\u003csup\u003ea\u003c/sup\u003e, FT\u003csup\u003ea\u003c/sup\u003e, FS\u003csup\u003ea\u003c/sup\u003e, XH, WT\u003csup\u003ea\u003c/sup\u003e, FC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRafter, purlin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBeijing, Shanxi Province\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePicea\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCT\u003csup\u003ea\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eflying rafter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eShanxi Province\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePinus\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eJHT\u003csup\u003ea\u003c/sup\u003e, CT\u003csup\u003ea\u003c/sup\u003e, DT\u003csup\u003ea\u003c/sup\u003e, XH, FC\u003csup\u003ea\u003c/sup\u003e,\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eColumn, Eaves board, rafter, beam, tile fillet, purlin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBeijing, Shanxi Province\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003eNote: \u003csup\u003ea\u003c/sup\u003eNational important cultural relic of China. CT\u0026thinsp;=\u0026thinsp;Chongshan Temple, CP\u0026thinsp;=\u0026thinsp;Chunyang Palace, DT\u0026thinsp;=\u0026thinsp;Dahui Temple, FC\u0026thinsp;=\u0026thinsp;the Forbidden City, FS\u0026thinsp;=\u0026thinsp;Financial Street, FT\u0026thinsp;=\u0026thinsp;Pagoda of Fogong Temple, JHT\u0026thinsp;=\u0026thinsp;Jiexiu Houtu Temple, XH\u0026thinsp;=\u0026thinsp;Xuanwu Hospital, WT\u0026thinsp;=\u0026thinsp;Wanshou Temple.\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eWooden components from heritage architectures undergo specialized sanding procedures that differ from standard xylarium wood specimen preparation to minimize structural damage. The in-situ preparation process involves sequential polishing of transverse end surfaces using a specialized 1 cm diameter grinder with 180, 240, 400, 600 and 800 sanding. This process removes the surface material of the wood components with the thickness of approximately 0.5-1.0 mm, ensuring optimal visibility of wood anatomical features while maintaining structural integrity. Image acquisition was performed using the iWood, which specifically designed to reduce damage as much as possible to the heritage wooden components. All samples were systematically identified to genus level through wood anatomical analysis. The resulting classification data were then used to annotate the acquired images for subsequent analytical processing and deep learning model development.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eSampling size of images collected from the components of heritage architectures.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eGenus\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eSamples numbers\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eImage numbers\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAbsent of defect\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCrack\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDecay\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eAbies\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e198\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eLarix\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e707\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e103\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e114\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePicea\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e215\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePinus\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2930\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e392\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e4050\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e860\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e530\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e\u003cem\u003e2.2 Deterioration classification\u003c/em\u003e\u003c/h2\u003e\n \u003cp\u003eThe primary forms of deterioration observed in wooden heritage architectures encompass surface changes\u003csup\u003e26\u003c/sup\u003e, crack\u003csup\u003e21\u003c/sup\u003e, mechanical deformation\u003csup\u003e27\u003c/sup\u003e, mechanical damage, insect infestation\u003csup\u003e27\u003c/sup\u003e, decay\u003csup\u003e28\u003c/sup\u003e, and biological growth. During the cross-sectional preparation process required for image acquisition, certain deterioration patterns \u0026mdash; particularly insect infestation, decay, and crack \u0026mdash; may become more pronounced in the collected data. Given the prevalence and structural significance of crack and decay as key degradation mechanisms in wooden components, these specific deterioration types were selected for detailed analysis.\u003c/p\u003e\n \u003cp\u003eThe wood identification accuracy of components is significantly influenced by varying degrees of deterioration. While previous studies have established comprehensive grading systems for decay states\u003csup\u003e29\u003c/sup\u003e, these classifications are primarily designed for macroscopic assessment and prove inadequate for analyzing small-scale areas captured in individual images. To address this limitation, a comprehensive five-level classification system was implemented according to crack and decay characteristics. The status and description of crack and decay classification are shown in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, and the illustration of crack and decay relevant to various levels are shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eCriterion for crack and decay grading of image data acquired from wooden heritage architecture.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDeterioration\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eCrack\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eDecay\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eLevel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eLevel of crack\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFeatures\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eLevel of decay\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFeatures\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ec0: Material in good condition\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eno crack was found within the view of the collector\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003ed0: Material in good condition\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eno decay was found within the view of the collector\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ec1: Minor crack\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of crack not exceeding 10% of the area within the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003ed1: Minor decay\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of decay not exceeding 10% of the area within the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ec2: Obviously crack\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of crack between 10% and 30% of the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003ed2: Obviously decay\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of decay between 10% and 30% of the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ec3: Serious crack\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of crack between 30% and 60% of the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003ed3: Serious decay\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of decay between 30% and 60% of the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ec4: Damaged crack\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of crack of 60% or more of the area within the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003ed4: Damaged decayed\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ethe area of decay of 60% or more of the area within the collector\u0026apos;s field of view\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003e2.3 Calculation of degradation area and degradation simulation\u003c/h2\u003e\n \u003cp\u003eTraditional image area quantification typically relies on pixel value analysis of target regions\u003csup\u003e30\u003c/sup\u003e, conventionally performed through manual measurement \u0026mdash; a process characterized by significant time requirements. The inherent morphological irregularity of wood deterioration patterns poses substantial challenges for accurate area quantification using conventional methods. Semi-automated annotation techniques offer a viable solution for enhancing computational efficiency while maintaining measurement precision.\u003c/p\u003e\n \u003cp\u003eFor the deterioration data, we used ISAT (An Interactive Semi-Automatic Annotation Tool)\u003csup\u003e31\u003c/sup\u003e for semi-automatic labelling. The method is used in conjunction with the SAM (Segment Anything Model)\u003csup\u003e32\u003c/sup\u003e to rapidly and accurately identify instances of deterioration types. Instance segmentation outputs were processed through a threshold-based binarization algorithm, with deterioration area quantification achieved through pixel analysis using numpy.sum (threshold\u0026thinsp;=\u0026thinsp;255) in Python. This computational pipeline enabled precise calculation of deterioration area percentages based on white pixel distribution within segmented regions. The methodological workflow is illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, while the distribution of sample and image number across different deterioration levels is presented in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDistribution of sample and image numbers across different deterioration levels of crack and decay for the components in heritage architecture.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eSample numbers\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eImage numbers\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCrack\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDecay\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCrack\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDecay\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e631\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e245\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e157\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e136\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e91\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLevel-4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e52\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003etotal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e860\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e530\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eThe dataset comprises 860 crack images and 530 decay images of wood cross-sections, as detailed in Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. While the decay image collection is numerically smaller, its distribution across severity levels (c1-c4) demonstrates greater uniformity compared to crack images, which predominantly cluster at c1 and c2 levels with significant underrepresentation at c3 and c4. To investigate the impact of cracking on wood identification accuracy, a subset of 3,403 cross-section images from 52 samples was selected for simulated crack generation at c3 and c4 severity levels. As illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, crack features typically manifest as irregular black patterns in cross-sectional images. The simulation process employed instance segmentation region with crack morphology replicated by setting instance segmentation region pixels to 0 while preserving surrounding areas, as demonstrated in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4 model construction for wood identification\u003c/h2\u003e\n \u003cp\u003eThe resurgence of large kernel models in computer vision, facilitated by advancements in computational hardware, has demonstrated superior predictive accuracy in recent studies\u003csup\u003e33\u003c/sup\u003e. Conventional deep models exhibit limited effective receptive fields\u003csup\u003e33, 34\u003c/sup\u003e. In contrast, large kernel architectures provide substantially expanded receptive fields that more closely approximate human perceptual characteristics. The deterioration features of wood will affect the model\u0026apos;s judgment to varying levels. A large receptive field is more likely to detect the length correlation between wood features, thus facilitating accurate judgement. As a representative implementation, RepLKNet\u003csup\u003e33\u003c/sup\u003e exemplifies the large kernel convolutional neural network architecture, with its structural configuration illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eThe model structure of RepLKNet is relatively simple. Following the input of image data, the image is processed by the stem module, which consists two convolutional layers and two depthwise separable convolutions. In the following four Stages, there is a preponderance of RepLK Blocks and ConvFFN modules. The majority of the large kernel reflected in the RepLK Block. Finally, the model downsamples through the Transition module.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e2.5 Evaluation standards\u003c/h2\u003e\n \u003cp\u003eIn target identification, the criterion for evaluating a model is its accuracy. The following equations 1 exemplify this:\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n \u003cp\u003eTP (True Positive) indicates that the target was identified correctly as a positive sample, and FP (False Positive) suggests that the target was identified as a positive sample but was a negative sample. FN (False Negative) indicates that the target was a negative sample, but was a positive sample. TN (True Negative) suggests that the target was identified correctly as a negative sample.\u003c/p\u003e\n \u003cp\u003eHowever, accuracy alone does not sufficiently evaluate model performance in this context, as wooden component identification typically requires multiple image acquisitions per sample. To address this requirement, we introduced confidence metrics to quantify model performance at the sample level. Sample confidence and precision were mathematically defined in Equations 2 and 3, respectively. A classification was considered correct when sample confidence values exceeded empirically determined thresholds of 0.7 or 0.9.\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e2.6 Experimental setup\u003c/h2\u003e\n \u003cp\u003eAll the above steps were implemented on a workstation (CPU: Intel Xeon Silver 4210R @ 2.4 GHz, RAM: 64 GB, and GPUs: NVIDIA GeForce GTX 3090). All the implementations of models are based on python 3.8, cuda 11.8 and PyTorch 2.0.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3. Results and discussion","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 In-situ wood identification of components in heritage architectures\u003c/h2\u003e\n \u003cp\u003eThe development of computer vision-based identification methods for wooden heritage architectures presents significant challenges due to two primary constraints: the difficult destructive sampling properties due to unique cultural values, and the scarcity of wooden components. To address these limitations, we implemented a wood identification model initially developed using xylarium wood specimen image databases. This model, which has demonstrated robust generalization capabilities through validation with unknown modern wood samples [unpublished data], was subsequently adapted and applied to the wood identification of components in heritage architectures.\u003c/p\u003e\n \u003cp\u003eThe experimental results demonstrate satisfactory performance across multiple evaluation metrics. Figure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e presents the confusion matrix and corresponding confidence levels for the RepLKNet classification outcomes. Notably, the model achieved exceptional recall rates exceeding 95% for \u003cem\u003ePinus\u003c/em\u003e and \u003cem\u003eLarix\u003c/em\u003e (Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003ea), two genera predominantly utilized in Chinese heritage architecture. In contrast, classification performance for \u003cem\u003eAbies\u003c/em\u003e and \u003cem\u003ePicea\u003c/em\u003e yielded lower recall rates of 58.08% and 70.70%, respectively. This performance discrepancy can be primarily attributed to misclassifications associated with specific samples, as detailed in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab5\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eWood identification results with an accuracy rate of less than 90% for the components of heritage architectures.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"8\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eSamples\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eGenus\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eImage numbers\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eConfidence\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003eDistribution of predicted results\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eAbies\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eLarix\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePicea\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePinus\u003c/em\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCSS02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePinus\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e56.25%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCSS03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePinus\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e78.57%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWA01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePicea\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e89.28%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWA05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePicea\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e60.97%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e16\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMY02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eAbies\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e39.13%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e53\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMY13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eAbies\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e61.76%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e23\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv align=\"left\" class=\"colspec\"\u003e\u003cbr\u003e\u003c/div\u003e\u0026nbsp;\u003ctable id=\"Tab6\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eSample precision of the RepLKNet model at the confidence of 70% and 90% respectively for the components of heritage architectures.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eGenus\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eSample precision\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e90% confidence\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e70% confidence\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbies\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e33.33%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e33.33%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLarix\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePicea\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePinus\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e95.12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e97.56%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAverage\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e90%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e93.33%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eIn practical identification, the reliability of the model cannot be evaluated on accuracy alone. Comprehensive performance analysis necessitates individual sample confidence calculations, as demonstrated in Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e(b). As shown in Table \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, among 60 samples with absence of deterioration, 54 (90%) achieved confidence levels exceeding 90%, while sample precision increased to 93.33% when applying a 70% confidence threshold.\u003c/p\u003e\n \u003cp\u003eIn the identification result of \u003cem\u003ePicea\u003c/em\u003e, a total of 55 images were erroneously identified as \u003cem\u003ePinus\u003c/em\u003e, while other 8 images were incorrectly identified as \u003cem\u003eLarix\u003c/em\u003e. For sample WA05, it had the lowest confidence level below 70% among \u003cem\u003ePicea\u003c/em\u003e samples, and a total of 16 images were incorrectly identified as \u003cem\u003ePinus\u003c/em\u003e (Table \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e). It is speculated that the reason may be due to the less difference in wood structure characteristics of Pinaceae wood in cross-section, and the change in the morphology of the tracheids during the transition from earlywood to latewood in the cross section of \u003cem\u003ePicea\u003c/em\u003e and soft pine (Subgen. \u003cem\u003eStrobus\u003c/em\u003e), one type of genus \u003cem\u003ePinus\u003c/em\u003e, is similar, and both of them are gradual. Therefore, the identification of \u003cem\u003ePicea\u003c/em\u003e and soft pine (Subgen. \u003cem\u003eStrobus\u003c/em\u003e) woods need to be carried out with the help of the main characteristics of cross-field pitting pattern on the radial section. The erroneous data for the \u003cem\u003eAbies\u003c/em\u003e mainly comes from two samples, MY02 and MY13. From the identification results, it can be concluded that \u003cem\u003eAbies\u003c/em\u003e is mainly prone to confusion with \u003cem\u003ePicea\u003c/em\u003e. In addition, only two of the 41 \u003cem\u003ePinus\u003c/em\u003e wood samples showing confidence levels below 90%. Given the high sample precision obtained in this study, the RepLKNet model indicates strong potential for application to the accurate wood identification of wooden heritage architectures. This scheme will effectively avoid the difficulties of computer vision methods that require the establishment of large-scale image databases of wooden components of wooden heritage architectures, and at the same time, it can also realize the rapid in-situ wood identification for wooden heritage architectures, and reduce the destructive sampling and the cycle of wood identification.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Minimum number of samples and images for establishing an effective wood identification model\u003c/h2\u003e\n \u003cp\u003eWhile image data from xylarium wood specimens have proven effective for wood identification in heritage architecture applications, this methodology typically requires substantial specimen collections. The limited availability of xylarium wood specimens presents significant challenges for model implementation and methodological application. A critical research question is regarding the minimum specimen and image requirements for achieving reliable wood identification accuracy (\u0026gt;\u0026thinsp;90%). Previous studies have discussed the number of xylarium wood specimens and the number of images separately\u003csup\u003e15\u003c/sup\u003e, but we believe that there should be a relative relationship between the minimum number of xylarium wood specimens and the minimum number of images. Therefore, this section will discuss the interrelationship between them.\u003c/p\u003e\n \u003cp\u003ePrior to determining the minimum number of xylarium wood specimens required, establishing an adequate image dataset is essential. The number of image acquisition potential varies significantly with specimen dimensions, necessitating initial categorization of specimens by genus and corresponding image quantity. Then, select the xylarium wood specimens in sequence and establish a model, and test the final results, as shown in Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e. The analysis revealed a positive correlation between specimen quantity and model accuracy, achieving optimal performance (96.2%) at 40 xylarium wood specimens per genus. Notably, accuracy consistently exceeded 90% when xylarium wood specimen numbers surpassed 20 per genus, while the rate of precision improvement diminished beyond this threshold. Consequently, 20 xylarium wood specimens per genus are established as the minimum requirement for reproducible results in this study.\u003c/p\u003e\n \u003cp\u003eWhile model accuracy generally improves with increased training data volume, this relationship is not strictly linear, particularly for anisotropic wood materials. The observed trend in Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e demonstrates a deviation from the expected correlation between image quantity and model accuracy, warranting further investigation to establish a definitive relationship. To examine this relationship systematically, comparative analyses were conducted using 20, 25, 30, and 35 xylarium specimens with corresponding training datasets of 1,500, 2,000, 2,500, 3,000, and 3,500 images. Given the physical size variability among xylarium wood specimens, individual specimens may yield fewer images than the calculated average. In such cases, when selecting a 2,000-image dataset from 20 xylarium wood specimens, for instance, specimens with limited image availability (\u0026lt;\u0026thinsp;100 images) were supplemented by random image selection from other xylarium wood specimens within the same genus to maintain dataset integrity.\u003c/p\u003e\n \u003cp\u003eAs shown in Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e, the number of xylarium wood specimens in the model directly affects the accuracy of the image. By controlling the number of images in each category and comparing different xylarium wood specimen sizes, we found that the overall accuracy increases with the increase of xylarium wood specimens, and the overall accuracy of the model also increases with the increase of image numbers. Notably, at image numbers of 2,000 and 2,500, models trained on 25 xylarium wood specimens marginally outperformed those using 30 xylarium wood specimens. However, this relationship reversed at higher image quantities, with 30-specimen models demonstrating superior accuracy. Analysis of fixed specimen quantities revealed that models trained on 20 or 25 xylarium wood specimens reached performance plateaus at approximately 2,000 images, suggesting this image volume sufficiently captures the representative features of these specimen sets.\u003c/p\u003e\n \u003cp\u003eBeyond this threshold, it is difficult to improve accuracy with additional images without incorporating new xylarium wood specimens. The 20-specimen model exhibited overfitting beyond 2,500 images, despite achieving\u0026thinsp;\u0026gt;\u0026thinsp;90% accuracy. Based on these findings, we recommend a minimum dataset of 1,500 images from 25 xylarium wood specimens per genus as optimal for maintaining\u0026thinsp;\u0026gt;\u0026thinsp;90% accuracy while preventing overfitting.\u003c/p\u003e\n \u003cp\u003eThe recommended dataset of 1,500 images from 25 xylarium wood specimens should preferentially absence of deterioration samples, without strict image quantity constraints per individual specimen. Previous research has controlled the number of images collected, and trained models on a large number of xylarium wood specimens, but collected very few images per specimen\u003csup\u003e35\u003c/sup\u003e. The restricted image sampling may inadequately represent specimen characteristics, as collectors cannot reliably determine whether minimal images sufficiently capture a specimen\u0026apos;s full feature. Moreover, this methodology necessitates substantial xylarium wood specimen resources and proves challenging to replicate. Since it is difficult to obtain xylarium wood specimens that have a clear background in plant taxonomy, we believe that the principle of \u0026quot;collecting as many images as possible\u0026quot; should be followed to reduce the need to collect a large number of xylarium wood specimens. Although excessive acquisition of images could theoretically induce overfitting, the practical threshold for such occurrences remains substantially high and scales with specimen quantity. This image acquisition principle enhances the reproducibility of computer vision-based wood identification methods, ensuring more consistent and reliable application outcomes.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 The impact of wood deterioration on identifying model accuracy\u003c/h2\u003e\n \u003cp\u003eDuring prolonged service duration, wood architectural heritage components exhibit inherent susceptibility to hygrothermal fluctuations, sustained creep deformation under mechanical loads, and microbiological degradation mechanisms, cumulatively manifesting as characteristic deterioration patterns including crack propagation, dimensional instability, and biodeterioration of wooden components\u003csup\u003e20\u003c/sup\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab7\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eIdentification \u003cstrong\u003eaccuracy for different levels of deterioration\u003c/strong\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ecrack\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003edecay\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLevel-1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e93.66%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e98.37%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLevel-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e92.99%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e90.44%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLevel-3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e81.58%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e90.87%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLevel-4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e71.17%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e88.45%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eThe confidence levels of deteriorated samples were systematically evaluated and compared against non-deteriorated specimens, as illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e9\u003c/span\u003e and Fig. \u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003e. Samples exhibiting severe crack (c3 and c4) demonstrated substantially reduced confidence levels, significantly impacting identification accuracy of the model, as shown in Table \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e. In contrast, decayed samples showed minimal confidence reduction. These findings align with previous research identifying tracheid morphology transitions between earlywood and latewood as primary identification features\u003csup\u003e37\u003c/sup\u003e. The current study further emphasizes the critical role of latewood anatomical feature integrity within growth rings for accurate wood identification.\u003c/p\u003e\n \u003cp\u003eWood decay primarily refers to the biochemical degradation of wood mediated by microbial enzymatic activity\u003csup\u003e38\u003c/sup\u003e. Among these microorganisms, fungi constitute the most significant degradative agents, categorized into white rot and brown rot based on their distinct decay mechanisms\u003csup\u003e39\u003c/sup\u003e. Research on coniferous wood decomposition has established characteristic patterns: brown rot fungi selectively depolymerize cellulose while leaving a modified lignin matrix, whereas white rot fungi demonstrate simultaneous lignocellulosic degradation through oxidative enzymatic systems\u003csup\u003e40\u003c/sup\u003e. Growth of fungi causing rot decay was limited and slower in latewood than in earlywood due to the narrow cell lumen, thicker wall, and higher density of latewood\u003csup\u003e41\u003c/sup\u003e. Meanwhile, when sanding the surface of wooden components of heritage architectures, due to the shallow sanding depth, the latewood cells with higher strength are more easily exposed than the earlywood cells. It is for these reasons that the state of preservation and presentation of latewood cells within a growth ring of wood components are more favourable, and also explains why the deterioration type of wood decay has less influence on the accuracy of identification.\u003c/p\u003e\n \u003cp\u003eThe impact of crack on identification accuracy is particularly significant, primarily attributable to the compromised integrity of latewood features. Most of the crack in wooden components of heritage architectures are dry shrinkage cracking, which is caused by the dry shrinkage and wet swelling characteristics, and anisotropy of wood\u003csup\u003e42\u003c/sup\u003e. Wood is a natural anisotropic material with hygroscopic and desorptive capabilities, and undergoes drying and wetting with changes in environmental temperature and humidity. At the microscopic level, drying crack often appear first in wood ray tissues, because ray tissues are mostly thin-walled cells, with low strength, and is not connected closely enough with the surrounding cells. When the wood shrinks, the ray cells undergo significant deformation and are damaged by the drying stresses, resulting in crack\u003csup\u003e43,44\u003c/sup\u003e. The cell wall of the latewood is usually thicker, and the amount of drying deformation is larger than that of the earlywood cells. The ray tissue of the latewood is first destroyed by the shrinkage stress that produce small crack, which then gradually expend along the ray tissues, leading to a gradual increase in depth and width\u003csup\u003e44,46\u003c/sup\u003e. This phenomenon results in frequent crack formation within latewood regions, consequently compromising the identification accuracy of the model.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eIn this study, a deep learning-based in-situ wood identification framework of components in wooden heritage architecture is firstly proposed. The constructed identification model based on image database of xylarium wood specimens were directly applied to the wooden heritage architecture, which effectively avoids the database construction problems of wooden components that are difficult to achieve large-scale image acquisition. At the same time, the model can be successfully applied to the rapid in-situ identification of wood in wooden heritage architectures, which reduces the destructive sampling and the identification cycle. The optimal algorithm RepLKNet achieves 96.67% accuracy in wood identification. The sample precisions were 93.33% and 90% at 70% and 90% confidence levels, respectively. The minimum xylarium wood specimen size and number of images required to build an effective model are 25 xylarium wood specimens and 1500 images per genus, respectively, for real-world test accuracies exceeding 90%. In addition, wood deterioration can have an impact on the wood identification accuracy of wooden heritage architectures. When the cracked area exceeds 30% of the collected image area, the recognition accuracy of the model decreases dramatically. Future work will focus on optimizing method efficiency and accuracy, particularly for components with varying types and levels of deterioration, while extending applicability from softwood to hardwood component specimens. This approach aims to establish adaptive conservation protocols for historic timber structures, ensuring scientific preservation and targeted restoration strategies across diverse material conditions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported financially by the National Key Research and Development Program of China (Grant No. 2023YFF0906301) and the National Science \u0026amp; Technology Fundamental Resources Investigation Program (Grant No. 2023FY101400). The authors would like to express our gratitude to Professor Xiaomei Jiang of the Research Institute of wood Industry, the Chinese Academy of Forestry for her valuable academic advice, and Professor Yongping Chen, Professor Juan Guo, Mrs. Mingkun Xu, Mr. Yonggang Zhang and Mr. Yu Sun of the Research Institute of wood Industry, the Chinese Academy of Forestry for their technical supports.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCZ, LJ, HT and YY designed the experiments. CZ, TL and YL prepared the samples for imaging and imaged the specimens. TL, YL, SL and CZ curated the collected dataset. LJ, RY, HZ and YY provided research sources. CZ developed the deep learning models. CZ, YY, SL and LJ analyzed the results. CZ, LJ and YY wrote, reviewed and edited the paper. LJ and YY conducted project administration and supervision. All authors read and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eZhu, H.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Wood-derived materials for green electronics, biological devices, and energy applications. \u003cem\u003eChem. Rev.\u003c/em\u003e \u003cstrong\u003e116\u003c/strong\u003e, 9305-9374 (2016).\u003c/li\u003e\n \u003cli\u003eInternational Council on Monuments and Sites. Guidelines for the Conservation of Wooden Built Heritage https://jianzhuyichan.tongji.edu.cn/info/1007/1543.htm (2017)\u003c/li\u003e\n \u003cli\u003eSichuan Institute of building research \u0026amp; Zhongke construction. Technical Standard for Maintenance and Reinforcement of Wooden Structures in Ancient Buildings. GBT501652020 https://zjj.sm.gov.cn/xxgk/fgwj/jsbz/202011/t20201113_1589718.htm (2020)\u003c/li\u003e\n \u003cli\u003eJiang, X., Yin, Y., \u0026amp; Liu, B. Current Status Development and Prospect of Wood Identification Technology. \u003cem\u003eChina Wood Industry\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 36-39 (2010). (in Chinese)\u003c/li\u003e\n \u003cli\u003eJiao, L.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Ancient plastid genomes solve the tree species mystery of the imperial wood \u0026ldquo;Nanmu\u0026rdquo; in the Forbidden City, the largest existing wooden palace complex in the world. \u003cem\u003ePlants, People, Planet\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 696-709 (2022).\u003c/li\u003e\n \u003cli\u003eGasson, P. How precise can wood identification be? Wood anato my\u0026apos;s role in support of the legal timber trade, especially CITES. \u003cem\u003eIAWA J.\u003c/em\u003e \u003cstrong\u003e32\u003c/strong\u003e, 137\u0026ndash;154. (2011).\u003c/li\u003e\n \u003cli\u003eDong, M. et al. Wood used in ancient timber architecture in Shanxi Province, China. \u003cem\u003eIAWA J.\u003c/em\u003e \u003cstrong\u003e38\u003c/strong\u003e, 182-200, doi:10.1163/22941932-20170167 (2017).\u003c/li\u003e\n \u003cli\u003eDormontt, E. E. et al. Forensic timber identification: It\u0026apos;s time to integrate disciplines to combat illegal logging. \u003cem\u003eBiol. Conserv.\u003c/em\u003e \u003cstrong\u003e191\u003c/strong\u003e, 790-798 (2015).\u003c/li\u003e\n \u003cli\u003eHartvig, I., Czako, M., Kj\u0026aelig;r, E. D., Nielsen, L. R. \u0026amp; Theilade, I. The use of DNA barcoding in identification and conservation of rosewood (Dalbergia spp.). \u003cem\u003ePlos One\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, e0138231 (2015).\u003c/li\u003e\n \u003cli\u003eJiao, L. et al. DNA Barcode Authentication and Library Development for the Wood of Six Commercial \u003cem\u003ePterocarpus\u003c/em\u003e Species: the Critical Role of Xylarium Specimens. \u003cem\u003eSci. Rep-Uk.\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, doi:10.1038/s41598-018-20381-6 (2018).\u003c/li\u003e\n \u003cli\u003eYu, M. et al. DNA barcoding of vouchered xylarium wood specimens of nine endangered Dalbergia species. \u003cem\u003ePlanta\u003c/em\u003e \u003cstrong\u003e246\u003c/strong\u003e, 1165-1176, doi:10.1007/s00425-017-2758-9 (2017).\u003c/li\u003e\n \u003cli\u003eLu Y. et al. DNA Methods for Identifying Wood in Ancient Timber Architecture. \u003cem\u003eChinese Journal of Wood Science and Technology\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 12-18 (2023). (in Chinese)\u003c/li\u003e\n \u003cli\u003eDom\u0026iacute;nguez-Delm\u0026aacute;s, M. Seeing the forest for the trees: New approaches and challenges for dendroarchaeology in the 21st century. \u003cem\u003eDendrochronologia\u003c/em\u003e \u003cstrong\u003e62\u003c/strong\u003e, 125731 (2020).\u003c/li\u003e\n \u003cli\u003eJiao, L., Lu, Y., He, T., Guo, J. \u0026amp; Yin, Y. DNA barcoding for wood identification: Global review of the last decade and future perspective. \u003cem\u003eIAWA J.\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 620-643 (2020).\u003c/li\u003e\n \u003cli\u003eTraor\u0026eacute;, M., Kaal, J. \u0026amp; Cortizas, A. M. Chemometric tools for identification of wood from different oak species and their potential for provenancing of Iberian shipwrecks (16th-18th centuries AD). \u003cem\u003eJ. Archaeol Sci.\u003c/em\u003e \u003cstrong\u003e100\u003c/strong\u003e, 62-73 (2018).\u003c/li\u003e\n \u003cli\u003eHe, T. et al. Developing deep learning models to automate rosewood tree species identification for CITES designation and implementation. \u003cem\u003eHolzforschung\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 1123-1133 (2020).\u003c/li\u003e\n \u003cli\u003eRavindran, P., Thompson, B. J., Soares, R. K. \u0026amp; Wiedenhoeft, A. C. The XyloTron: Flexible, Open-Source, Image-Based Macroscopic Field Identification of Wood Products. \u003cem\u003eFront. Plant Sci.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, doi:10.3389/fpls.2020.01015 (2020).\u003c/li\u003e\n \u003cli\u003eRavindran, P., Costa, A., Soares, R. \u0026amp; Wiedenhoeft, A. C. Classification of CITES-listed and other neotropical Meliaceae wood images using convolutional neural networks. \u003cem\u003ePlant Methods\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, doi:10.1186/s13007-018-0292-9 (2018).\u003c/li\u003e\n \u003cli\u003eRussakovsky, O. et al. Imagenet large scale visual recognition challenge. \u003cem\u003eInt. J. Comput. Vision.\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e115\u003c/strong\u003e, 211-252 (2015).\u003c/li\u003e\n \u003cli\u003eStratigaki, M. Autofluorescence for the Visualization of Microorganisms in Biodeteriorated Materials in the Context of Cultural Heritage. \u003cem\u003eChemPlusChem\u003c/em\u003e \u003cstrong\u003e89\u003c/strong\u003e, e202400170 (2024).\u003c/li\u003e\n \u003cli\u003eLi, X., Qian, W. \u0026amp; Chang, L. Analysis of the density of wooden components in ancient buildings by micro-drilling resistance, using information diffusion. \u003cem\u003eBioResources\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 5777-5787 (2019).\u003c/li\u003e\n \u003cli\u003eYin, Y. et al. Research on the Identification of Wood Species Used for Wooden Structures in Southeastern Shanxi Province. \u003cem\u003eWorld of Antiquity\u003c/em\u003e \u003cstrong\u003e04\u003c/strong\u003e, 33-36 (2010) (in Chinese)\u003c/li\u003e\n \u003cli\u003eZhang, Q. An Analysis of the History of the Palace of Compassion and Tranquility Complex from the Aspect of Wood Species of Timber Members, The Forbidden City. \u003cem\u003eHeritage Architecture\u003c/em\u003e \u003cstrong\u003e04\u003c/strong\u003e, 1-12 (2020). (in Chinese)\u003c/li\u003e\n \u003cli\u003eLi, S. et al. Research on the identification and configuration of wood components for the main hall of Jianshui Zhilin Temple Cultural Relics Protection and Archaeological Science. \u003cem\u003eSciences of Conservation and Archaeology\u003c/em\u003e \u003cem\u003e32\u003c/em\u003e, 91-98 (2020). (in Chinese)\u003c/li\u003e\n \u003cli\u003eHe, T. et al. iWood: An Automated Wood Identification System for Endangered and Precious Tree Species Using Convolutional Neural Networks. \u003cem\u003eScientia Silvae Sincae\u003c/em\u003e \u003cstrong\u003e57\u003c/strong\u003e,152-159 (2021). (in Chinese)\u003c/li\u003e\n \u003cli\u003eTan, Y. et al. Inspection and Evaluation of Wood Components of Ancient Buildings in the South-Three Courts of the Forbidden City. \u003cem\u003eBioResources.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e (2022).\u003c/li\u003e\n \u003cli\u003eMa, X. et al. 3D structural deformation monitoring of the archaeological wooden shipwreck stern investigated by optical measuring techniques.\u003cem\u003e\u0026nbsp;J. Cult. Herit.\u003c/em\u003e \u003cstrong\u003e59\u003c/strong\u003e, 102-112 (2023).\u003c/li\u003e\n \u003cli\u003eVenugopal, P., Junninen, K., Linnakoski, R., Edman, M. \u0026amp; Kouki, J. Climate and wood quality have decayer-specific effects on fungal wood decomposition. \u003cem\u003eForest. Ecol. Manag.\u003c/em\u003e \u003cstrong\u003e360\u003c/strong\u003e, 341-351 (2016).\u003c/li\u003e\n \u003cli\u003eChina Academy of Forestry Wood Industry Research Institute, et al. LY/T 2014-2024 Non-destructive testing method and defects classification for wooden components of ancient buildings. National Forestry and Grassland Administration https://std.samr.gov.cn/hb/search/stdHBDetailed?id=1E5DB4EC8381C2FFE06397BE0A0A7B72 (in Chinese)\u003c/li\u003e\n \u003cli\u003eJi, M., Zhang, W., Wang, G., Wang, Y. \u0026amp; Miao, H. Online Measurement of Outline Size for Pinus densiflora Dimension Lumber: Maximizing Lumber Recovery by Minimizing Enclosure Rectangle Fitting Area. \u003cem\u003eForests\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 1627 (2022).\u003c/li\u003e\n \u003cli\u003eJi, S. \u0026amp; Zhang, H. ISAT with Segment Anything: An Interactive Semi-Automatic Annotation Tool v1.10 https://github.com/yatengLG/ISAT_with_segment_anything (2025)\u003c/li\u003e\n \u003cli\u003eKirillov, A. et al. Segment anything[C]//in Proceedings of the IEEE/CVF International Conference on Computer Vision. 4015-4026 (2023).\u003c/li\u003e\n \u003cli\u003eDing, X., Zhang, X., Han, J. \u0026amp; Ding, G. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns[C]//in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11963-11975 (2022).\u003c/li\u003e\n \u003cli\u003eDing, X. et al. UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition[C]//in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5513-5524 (2024).\u003c/li\u003e\n \u003cli\u003eRavindran, P., Owens, F. C., Wade, A. C., Shmulsky, R. \u0026amp; Wiedenhoeft, A. C. Towards sustainable North American wood product value chains, part I: computer vision identification of diffuse porous hardwoods. \u003cem\u003eFront. Plant Sci.\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 758455 (2022).\u003c/li\u003e\n \u003cli\u003eZhu, Q., Zhou, X., Tan, J. \u0026amp; Guo, L. Knowledge base reasoning with convolutional-based recurrent neural networks. \u003cem\u003eIeee T. Knowl. Data En.\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 2015-2028 (2019).\u003c/li\u003e\n \u003cli\u003ePieter, B. et al. IAWA List of microscopic features for softwood identification. \u003cem\u003eIAWA J.\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 1-70 (2004).\u003c/li\u003e\n \u003cli\u003eAyrilmis, N., Kaymakci, A. \u0026amp; G\u0026uuml;le\u0026ccedil;, T. Potential use of decayed wood in production of wood plastic composite. \u003cem\u003eInd. Crop Prod.\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e74\u003c/strong\u003e, 279-284 (2015).\u003c/li\u003e\n \u003cli\u003eGreen, F. \u0026amp; Highley, T. L. Mechanism of brown-rot decay: paradigm or paradox. \u003cem\u003eInt. Biodeter. Biodegr.\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, 113-124 (1997).\u003c/li\u003e\n \u003cli\u003eBlanchette, R. A. Wood decay: a submicroscopic view. \u003cem\u003eJ. Forest\u003c/em\u003e \u003cstrong\u003e78\u003c/strong\u003e, 734-737 (1980).\u003c/li\u003e\n \u003cli\u003eBesma, B., Ahmed, K., \u0026amp; Yves, B. Effects of biodegradation by brown-rot decay on selected wood properties in eastern white cedar (Thuja occidentalis L.). \u003cem\u003eInt. Biodeterior. Biodegrad.\u003c/em\u003e \u003cstrong\u003e87\u003c/strong\u003e, 87-98 (2014).\u003c/li\u003e\n \u003cli\u003eChen, Y., Huang, A., Meng, S., \u0026amp; Ni, L. Research Progress in Influencing Factors to Wood Deformation and Cracking and Anti-cracking Measures. \u003cem\u003eWorld Forestry Research\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 71-76 (2024) (in Chinese)\u003c/li\u003e\n \u003cli\u003eWang, H. H. \u0026amp; Youngs, R. L. Drying stress and check development in the wood of two oaks. \u003cem\u003eIAWA J.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 15-30 (1996).\u003c/li\u003e\n \u003cli\u003eYamamoto, H., Sakagami, H., Kijidani, Y. \u0026amp; Matsumura, J. Dependence of microcrack behavior in wood on moisture content during drying. \u003cem\u003eAdv. Mater. Sci. Eng.\u003c/em\u003e \u003cstrong\u003e2013\u003c/strong\u003e, 802639 (2013).\u003c/li\u003e\n \u003cli\u003eGao, Y. et al. The formation mechanism of microcracks and fracture morphology of wood during drying. \u003cem\u003eDry. Technol.\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 1268-1277 (2023).\u003c/li\u003e\n \u003cli\u003eSakagami, H. Microcrack propagation in transverse surface from heartwood to sapwood during drying. \u003cem\u003eJ. Wood Sci.\u003c/em\u003e \u003cstrong\u003e65\u003c/strong\u003e, 33 (2019).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Heritage architecture, Wood identification, Genus, Computer vision, Latewood, Wood anatomy, Wood crack, Wood decay","lastPublishedDoi":"10.21203/rs.3.rs-6178164/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6178164/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWood identification of structural components is crucial for heritage architecture conservation, elucidating utilization patterns of forest resources and revealing evolution of civilization in the history. This study first proposed a computer vision-based in-situ identification method using 63 wooden components and 4050 digital images obtained from nine representatives of historical heritage architectures in China. The optimal algorithm, RepLKNet, which was developed on the training dataset constructed by collected images from xylarium specimens of coniferous wood (\u003cem\u003eAbies\u003c/em\u003e, \u003cem\u003eLarix\u003c/em\u003e, \u003cem\u003ePicea\u003c/em\u003e, and \u003cem\u003ePinus\u003c/em\u003e), achieved a wood identification accuracy of 96.67%, with average sample precisions of 93.33% and 90% at confidence levels of 70% and 90% respectively for the components of heritage architectures. The minimum sample size requirements for constructing an effective model were determined to be 25 wood specimens and 1500 images per genus, as validated by real-world testing with an accuracy exceeding 90%. Meanwhile, this study investigated the impact of two common deterioration types in wooden components of heritage construction \u0026mdash; decay and crack \u0026mdash; as well as their severity, on the identification accuracy of the proposed method. The results demonstrate that crack exert a more significant impact on the wood recognition accuracy of historical components compared to decay. Specifically, when the cracked area exceeds 30% of the captured image area, the model\u0026rsquo;s identification accuracy experiences a sharp decline. Furthermore, the integrity of latewood features plays a crucial role in wood identification accuracy, particularly when compared to the earlywood region within the growth ring. The computer vision-driven methodology for in-situ identification and assessment of wooden components proposed and implemented in this investigation contributes to the advancement of structural preservation strategies, and preventive maintenance practices in heritage architectures.\u003c/p\u003e","manuscriptTitle":"Deep Learning-based In-situ Coniferous Wood Identification of Components in Heritage Architectures of China","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-26 09:17:27","doi":"10.21203/rs.3.rs-6178164/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-22T20:27:57+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-09T11:21:20+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-07T08:56:37+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"182147152434113419783886541854487139436","date":"2025-04-28T04:20:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"57259474008388028296277694243318739234","date":"2025-04-28T00:38:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"178495919516709512053545414630013602631","date":"2025-04-26T02:33:48+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-03-17T21:41:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-03-17T03:55:42+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-03-17T03:54:38+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj heritage science","date":"2025-03-07T12:06:28+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"951f7ede-d382-4773-9b58-f106bf7487ec","owner":[],"postedDate":"March 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-11-03T16:02:41+00:00","versionOfRecord":{"articleIdentity":"rs-6178164","link":"https://doi.org/10.1038/s40494-025-02120-z","journal":{"identity":"npj-heritage-science","isVorOnly":false,"title":"npj Heritage Science"},"publishedOn":"2025-10-28 15:58:18","publishedOnDateReadable":"October 28th, 2025"},"versionCreatedAt":"2025-03-26 09:17:27","video":"","vorDoi":"10.1038/s40494-025-02120-z","vorDoiUrl":"https://doi.org/10.1038/s40494-025-02120-z","workflowStages":[]},"version":"v1","identity":"rs-6178164","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6178164","identity":"rs-6178164","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}