Wood science has long been shaped by a fundamental challenge: the properties that determine material performance are intrinsically heterogeneous, anisotropic, scale-dependent, and sensitive to environmental and processing conditions. Density, moisture, thermal behavior, surface quality, bonding performance, internal integrity, and species identity cannot be treated as isolated attributes, because they emerge from interactions among anatomical structure, chemical composition, processing history, and service environment. Consequently, the advancement of wood-based industries depends not only on measuring these properties more accurately, but also on transforming measurements into reliable models, interpretable predictions, and technically relevant decisions. This complexity is fundamentally rooted in the hierarchical structure of wood, which spans from macroscopic anatomical features such as tracheids and vessels down to cellulose microfibrils and molecular architecture, governing its mechanical, thermal, and fluidic behavior [1].
While density accounts for a substantial portion of inter-species variability, typically ranging from 100 to 1200 kg/m3, it remains insufficient to predict properties dominated by cellular morphology, such as transverse shrinkage [2]. Moisture further amplifies this complexity by playing a central role in wood performance, affecting decomposition resistance, mechanical properties, and dimensional stability. From a modeling perspective, the prediction of wood quality has evolved to include both empirical and process-based approaches, particularly for standing trees, and is increasingly integrated into decision-support systems for forest management. At the structural scale, orthotropy remains a defining characteristic that must be explicitly considered in design, alongside factors such as load duration, moisture content, temperature, and size effects [3].
This was the premise of the Special Issue “Wood Properties: Measurement, Modeling, and Future Needs”, which sought to bring together innovative approaches involving ultrasound, imaging, spectroscopy, machine learning, and predictive modeling while maintaining a clear connection with the practical demands of construction, furniture production, forest management, and wood processing. Taken together, the six contributions published in this Special Issue reveal a broader transition in the field. Research on the properties of wood is moving from isolated measurements toward integrated workflows in which experimental acquisition, metrological control, computational modeling, and decision support are progressively connected. This transition should not be interpreted as the replacement of physical experimentation by computation. Rather, the collection demonstrates that the value of modeling depends directly on the quality, representativeness, and interpretability of the underlying measurements. Likewise, the value of advanced measurement techniques increases when the resulting data can be converted into predictions, classifications, risk assessments, or process adjustments. The central scientific message of this Special Issue is therefore that the future of wood characterization lies in the convergence of measurement reliability and data-driven inference.
This convergence is particularly evident when the articles are considered along a continuum extending from the determination of physical properties to the prediction of complex performance indicators. At the measurement end of this continuum, Meng et al. [4] addressed the reliability of the transient plane source method for determining the thermal conductivity of wood. Their study demonstrated that measurement accuracy cannot be separated from parameter selection. Power output, heating time, and the time window used for analysis directly affected the stability of the results, while indicators such as total-to-characteristic time, probe-temperature increase, mean deviation, and temperature-drift behavior provided criteria for judging measurement quality. The recommended conditions—power outputs of 0.05 or 0.1 W, a heating time of 120 s, and a time window covering 60%–80% of the heating period—reduced standard uncertainty by 18.9%–59.5% relative to unoptimized measurements. The work also showed that the transient method produced higher thermal-conductivity values than the steady-state method in all grain directions, reinforcing the need to interpret measurements in relation to both the physical assumptions and operational settings of the selected technique.
This contribution is important beyond thermal conductivity itself. It illustrates a requirement that applies to the entire field: data used for modeling or comparison must be generated under conditions whose uncertainties are understood. Wood is especially sensitive to methodological inconsistency because anisotropy, density variation, moisture gradients, specimen geometry, and grain direction can alter measured responses. The development of faster techniques is therefore insufficient unless accompanied by parameter validation, uncertainty analysis, and explicit criteria for data acceptance. This metrological perspective forms the foundation upon which predictive wood science must be built.
The same principle emerges in the internal assessment of living trees. Esteban et al. [5] compared acoustic tomography generated using ultrasonic devices operating at different frequencies and a stress-wave device, while using drilling resistance to refine the interpretation of internal defects in urban trees. Tests on cross-sections from Robinia pseudoacacia, Platanus × hybrida, Ulmus pumila, and Populus alba showed that higher-frequency ultrasonic measurements, particularly those obtained at 45 kHz, produced more accurate representations of defects. However, the study also demonstrated that tomographic imaging alone may not reliably distinguish among cavities, cracks, and deteriorated zones. Drilling resistance profiles provided complementary local information and improved defect identification. The combination of spatial imaging and localized mechanical verification therefore offered a more robust diagnosis than either method in isolation.
This result represents a broader methodological lesson. Future wood assessment will increasingly depend on multimodal strategies rather than on a single “best” instrument. Acoustic propagation provides information over an entire cross-section, whereas drilling resistance interrogates a limited path with higher local specificity. Their strengths are complementary, as are their limitations. The integration of techniques operating at different spatial scales can reduce ambiguity, improve confidence in interpretation, and support more defensible decisions in urban-tree management. Similar logic applies to structural timber, heritage elements, standing forests, and engineered wood products, where the combination of global and local measurements may be more valuable than marginal improvements in any isolated technique.
A related transition from physical signal acquisition to intelligent interpretation is evident in the work of Silva et al. [6], who combined near-infrared spectroscopy with machine learning classifiers to discriminate four economically relevant wood species from an Amazonian floodplain forest. The study used 32 increment cores, from which 640 spectra were obtained across the bark-to-pith direction. Although the mean spectra displayed broadly similar profiles, differences in absorbance intensity contained sufficient information for classification. Principal component analysis explained 98.4% of total variance in the first two components, but also revealed substantial overlap among several species, demonstrating that visual or unsupervised separation alone was insufficient. Once supervised classification was applied, partial least squares discriminant analysis achieved a mean accuracy and F1 score above 98%, with recognition rates exceeding 96% for all species and reaching 100% for Pseudobombax munguba. By contrast, support vector machine and k-nearest-neighbor models achieved substantially lower performance under the selected validation procedure.
This study illustrates how chemical fingerprints that appear similar at first inspection can become operationally useful when combined with appropriate statistical learning. Its implications extend beyond laboratory classification. Timber identification in highly diverse tropical forests is constrained by the similarity of macroscopic features, the need for specialized anatomical knowledge, and the limited feasibility of detailed microscopic examination at large scales. NIR-based classification offers a rapid and comparatively non-destructive route that could contribute to traceability, legal timber verification, species-specific processing, and the reduction of inadvertent substitution. At the same time, the contrast among classifier performances is a reminder that machine learning is not a generic layer that automatically improves any dataset. Model selection, preprocessing, blocking, cross-validation, and prevention of data leakage are decisive components of the scientific method. The authors’ use of blocked repeated cross-validation, ensuring that spectra from the same increment core were not divided between training and validation sets, is therefore as important as the reported accuracy itself.
The predictive modeling contributions of the Special Issue extend this progression from classification to the forecasting of process performance. Three studies focused on heat-treated wood and employed backpropagation neural networks optimized through metaheuristic algorithms. Considered together, these articles demonstrate both the potential and the current methodological limitations of intelligent modeling in wood processing.
He et al. [7] proposed an improved Harris Hawks Optimization algorithm integrating Sobol-sequence initialization, Whale Optimization, and t-distribution perturbation. The resulting IHHO-BP model predicted the bonding strength of heat-treated wood as a function of temperature, treatment time, feed rate, cutting speed, and abrasive grit size. Compared with conventional BP, WOA-BP, and HHO-BP models, the improved approach reduced MAE, RMSE, and MAPE by at least 51.16%, 40.38%, and 51.93%, respectively, while increasing R2 by at least 10.85%. The contribution demonstrated that the random initialization and local-minimum sensitivity of conventional BP networks can be substantially mitigated by optimization strategies designed to improve population diversity, global exploration, and local refinement.
Cao et al. [8] subsequently compared four more recent metaheuristic approaches—Hippopotamus Optimization, Parrot Optimization, Osprey Optimization, and Goose Optimization—for tuning BP networks used to predict the adhesive strength of heat-treated Chinese pine. The HO-BP model performed best, reaching MAE, RMSE, MAPE, and R2 values of 0.0822 MPa, 0.1024 MPa, 1.1317%, and 0.9358%, respectively. It also outperformed linear regression, extreme learning machine, support vector regression, GA-SVR, and a previously developed CPA-BP model. Importantly, XGBoost-based feature-importance analysis indicated that cutting speed and treatment duration had the greatest influence on the predictions, whereas treatment temperature had the lowest importance within the investigated domain.
The latter result has practical and conceptual significance. Heat treatment is often discussed primarily in terms of temperature, but model interpretation indicated that surface preparation and processing history may govern bonding performance more strongly than temperature over a restricted experimental range. Such findings show how predictive models can contribute not merely by reproducing test values, but by reorganizing the relative importance attributed to process variables. Nevertheless, importance scores must be interpreted as model- and dataset-dependent quantities rather than universal causal rankings. Their validity is bounded by the experimental domain, variable ranges, sample size, and potential correlations among inputs.
The prediction of glossiness reported by Cao et al. [9] broadened this modeling framework from bonding performance to an aesthetic and commercially relevant surface property. Their improved Whale Optimization Algorithm combined tent-chaos initialization, sine–cosine operations, and adaptive inertia weights to optimize a BP neural network. The model used wood species, treatment temperature, treatment pressure, grain direction, and measurement angle as inputs and was trained using 216 observations from alder, beech, birch, and pine. Relative to the conventional BP model, IWOA-BP reduced test-set MAE, MAPE, and RMSE by 66.02%, 64.21%, and 69.60%, respectively, while increasing R2 by 12.87%. The optimized model attained an R2 of 0.9885 on the test set. Feature-importance analysis identified grain direction as the dominant predictor, followed by treatment temperature and pressure, whereas wood species had the lowest importance within the dataset.
Read collectively, the three neural-network studies show that the predictive performance of a familiar architecture can be substantially improved through different metaheuristic search strategies. However, they also expose an important future need. Continued proliferation of bio-inspired optimization algorithms will add scientific value only when improvements are demonstrated under rigorous and comparable conditions. The field must move beyond reporting the performance of a newly proposed optimizer against a limited set of alternatives. Stronger evidence will require repeated nested validation, uncertainty intervals, statistical comparison of errors, external datasets, assessment of computational cost, and testing across species, laboratories, instruments, and processing conditions. The article comparing four metaheuristic algorithms explicitly recognized that stronger statistical testing and broader data coverage remain necessary. The glossiness study similarly identified missing variables—including density, moisture content, treatment duration, age, growth environment, and harvest season—as constraints on model generalizability.
These limitations are not peripheral; they define the next stage of development. High coefficients of determination within a specific dataset do not by themselves establish industrial reliability. Wood datasets are frequently small, structurally dependent, and derived from narrow combinations of species and conditions. Random train–test division can overestimate generalization when measurements from closely related specimens are distributed across both subsets. In contrast, the blocked validation adopted in the NIR classification study provides an example of a more conservative design. Future predictive studies should increasingly separate data by tree, batch, species, production line, geographic origin, or laboratory, depending on the intended application. Models should be tested not only on unseen observations but also on genuinely unseen sources of variability.
Interpretability must also develop alongside accuracy. Feature-importance analysis was included in the studies of glossiness and adhesive strength, representing a useful step toward explaining model behavior. Yet future work should distinguish between variable importance, sensitivity, interaction, and causality. Techniques such as permutation importance, partial-dependence analysis, SHAP values, global sensitivity analysis, and physically constrained modeling could help establish whether a model is learning meaningful material relationships or merely exploiting correlations specific to a dataset. In wood science, physically implausible predictions may appear statistically accurate within a restricted domain. Incorporating knowledge of anisotropy, moisture transport, thermal degradation, wave propagation, and interfacial adhesion into computational models could reduce this risk.
The six studies also reveal that digitalization in wood science will not follow a single technological pathway. In one case, the key advance was better selection of experimental parameters for thermal measurement. In another, it was the fusion of tomography and drilling resistance. Elsewhere, spectral data were converted into species classifications, and experimental process variables were converted into forecasts of glossiness or bonding strength. These are different manifestations of the same underlying transformation: measurements are becoming components of decision systems rather than isolated endpoints.
For industrial implementation, this transformation will require interoperable datasets, standardized metadata, traceable calibration, and clearly defined domains of applicability. Data should record not only the target property, but also specimen origin, anatomical direction, moisture condition, environmental history, equipment configuration, preprocessing, and uncertainty. Without these elements, datasets from different laboratories may be numerically compatible but scientifically non-equivalent. Open and well-documented databases would enable more rigorous benchmarking, external validation, transfer learning, and the development of models that are less dependent on a single experimental campaign.
Real-time and in-line applications are a logical extension of the research presented here. NIR instruments can potentially support field identification and production-line sorting. Acoustic systems may contribute to continuous inspection of trees or timber elements. Predictive models could estimate surface or bonding performance before destructive testing, allowing processing conditions to be adjusted earlier. Thermal property measurements could support better control of drying, pressing, and heat-treatment operations. However, real-time implementation will require attention to model latency, sensor drift, uncertainty propagation, and decision thresholds. A prediction should not be considered operational merely because it can be calculated rapidly; it must also be reliable enough to justify the action that follows.
Sustainability provides the wider context for these developments. More accurate measurement can prevent unnecessary rejection or replacement. Non-destructive assessment can extend the service life of trees and timber products. Species classification can support responsible forest management and timber traceability. Predictive models can reduce experimental consumption, energy use, production losses, and material waste. Yet the environmental value of digital tools should be demonstrated rather than assumed. Future studies should quantify whether model-supported optimization produces measurable reductions in energy demand, raw-material consumption, testing requirements, or product failure.
The contributions assembled in this Special Issue therefore define measurement, modeling, and future needs as inseparable elements. Reliable measurement provides the evidence base; modeling organizes complex relationships; validation establishes trust; and application converts scientific knowledge into improved management and production. The field is progressing from property determination toward integrated material intelligence, but this transition remains incomplete. Its success will depend less on the isolated novelty of individual instruments or algorithms than on the ability to connect robust experiments, transparent models, representative datasets, and technically meaningful decisions. The strategic priority for wood science is now clear: to build measurement-to-decision systems that retain the physical rigor of traditional wood characterization while gaining the speed, scalability, and predictive capacity of modern data science.
Author Contributions
Conceptualization, R.d.A.D. and R.H.D.; writing—original draft preparation, R.d.A.D. and T.A.d.S.; writing—review and editing, R.d.A.D., R.H.D. and T.A.d.S.; supervision, R.d.A.D. and R.H.D. All authors have read and agreed to the published version of the manuscript.
Conflicts of Interest
The authors declare no conflicts of interest.
References
- Chen, C.; Kuang, Y.; Zhu, S.; Burgert, I.; Keplinger, T.; Gong, A.; Li, T.; Berglund, L.; Eichhorn, S.J.; Hu, L. Structure–Property–Function Relationships of Natural and Engineered Wood. Nat. Rev. Mater. 2020, 5, 642–666. [Google Scholar] [CrossRef] [Scilit]
- Perré, P.; Almeida, G.; Ayouz, M.; Frank, X. New Modelling Approaches to Predict Wood Properties from Its Cellular Structure: Image-Based Representation and Meshless Methods. Ann. For. Sci. 2016, 73, 147–162. [Google Scholar] [CrossRef] [Scilit]
- Arriaga, F.; Wang, X.; Íñiguez-González, G.; Llana, D.F.; Esteban, M.; Niemz, P. Mechanical Properties of Wood: A Review. Forests 2023, 14, 1202. [Google Scholar] [CrossRef] [Scilit]
- Meng, H.; Yu, X.; Chen, B.; Ren, P.; Zhao, J. Parameter Estimations on Measurement Accuracy for Thermal Conductivity of Wood Using the Transient Plane Source Method. Forests 2024, 15, 1820. [Google Scholar] [CrossRef] [Scilit]
- Esteban, M.; Olvera-licona, G.; Virgen-cobos, G.H.; Bobadilla, I. Comparison of Acoustic Tomography and Drilling Resistance for the Internal Assessment of Urban Trees in Madrid. Forests 2025, 16, 1125. [Google Scholar] [CrossRef] [Scilit]
- Silva da Silva, W.D.; dos Santos, J.X.; Naide Acosta, T.L.; Souza, D.V.; Ferreira, A.P.S.; Reis, P.C.M.d.R.; Reis, L.P.; Vieira, H.C.; de Muñiz, G.I.B.; Nisgoski, S. Near-Infrared Spectroscopy and Machine Learning for Wood Species Discrimination in an Amazon Floodplain Forest Management Area. Forests 2025, 16, 984. [Google Scholar] [CrossRef] [Scilit]
- He, Y.; Wang, W.; Cao, Y.; Wang, Q.; Li, M. Prediction of Bonding Strength of Heat-Treated Wood Based on an Improved Harris Hawk Algorithm Optimized BP Neural Network Model (IHHO-BP). Forests 2024, 15, 1365. [Google Scholar] [CrossRef] [Scilit]
- Cao, Y.; Wang, W.; He, Y. Prediction of Heat-Treated Wood Adhesive Strength Using BP Neural Networks Optimized by Four Novel Metaheuristic Algorithms. Forests 2025, 16, 291. [Google Scholar] [CrossRef] [Scilit]
- Cao, Y.; Wang, W.; He, Y. Predicting Glossiness of Heat-Treated Wood Using the Back Propagation Neural Network Optimized by the Improved Whale Optimization Algorithm. Forests 2025, 16, 716. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.