Advanced wood species identification based on multiple anatomical sections and using deep feature transfer and fusion

TL;DR

This work tackles wood species identification by addressing limitations of traditional morphology-based methods and costly alternative techniques. It advances a transfer-learning framework that uses two feature-extraction strategies, GAP and RADAM, on a Congo timber dataset across three anatomical sections (transversal, tangential, radial). RADAM, in particular, demonstrates superior robustness and accuracy, outperforming a prior LPQ-based approach across multiple datasets and fusion schemes, with substantial gains on larger, more varied data. The results suggest a practical, scalable pathway toward automated, texture-based wood identification to support forest conservation and regulatory enforcement, with future potential for integrating higher-resolution imagery and Vision Transformer architectures. Key contributions include multi-view fusion of deep features, randomized autoencoder-based feature extraction, and a comprehensive cross-validation analysis demonstrating strong generalization across sections and datasets.

Abstract

In recent years, we have seen many advancements in wood species identification. Methods like DNA analysis, Near Infrared (NIR) spectroscopy, and Direct Analysis in Real Time (DART) mass spectrometry complement the long-established wood anatomical assessment of cell and tissue morphology. However, most of these methods have some limitations such as high costs, the need for skilled experts for data interpretation, and the lack of good datasets for professional reference. Therefore, most of these methods, and certainly the wood anatomical assessment, may benefit from tools based on Artificial Intelligence. In this paper, we apply two transfer learning techniques with Convolutional Neural Networks (CNNs) to a multi-view Congolese wood species dataset including sections from different orientations and viewed at different microscopic magnifications. We explore two feature extraction methods in detail, namely Global Average Pooling (GAP) and Random Encoding of Aggregated Deep Activation Maps (RADAM), for efficient and accurate wood species identification. Our results indicate superior accuracy on diverse datasets and anatomical sections, surpassing the results of other methods. Our proposal represents a significant advancement in wood species identification, offering a robust tool to support the conservation of forest ecosystems and promote sustainable forestry practices.

Figures (9)

• Figure 1: Representation of each anatomical section of a wood sample from the species Afzelia africana. The arrows point from the respective regions on the sample (transversal, tangential, and radial) to their corresponding images. Each image offers unique structural characteristics that are crucial for wood species identification.
• Figure 2: Illustration of the data augmentation techniques employed for each sub-dataset. (a) 1000 $\times$ 1000 original dataset, (b) 1000 $\times$ 500 dataset with each image halved, (c) 500 $\times$ 500 dataset with each original image split into four equal parts, and (d) 500 $\times$ 500 - OGRN dataset demonstrating the applied transformations including original (O), Gaussian smoothing (G), 90-degree rotation (R), and salt-and-pepper noise addition (N).
• Figure 3: Methodological process employed in this study. Anatomical sections are input to a pre-trained backbone to extract feature blocks. These blocks are then fused and classified using an SVM to produce the final wood species prediction.
• Figure 4: Confusion matrix for the GAP results on the transversal anatomical section in the original dataset. Class b) Chrysophyllum africanum is often misclassified with d) Pentaclethra eetveldeana, which is also confused with c) Mammea africana class.
• Figure 5: Confusion matrix for the GAP results on the tangential anatomical section in the original dataset. The model often misclassifies c) Lophira alata with b) Klainedoxa gabonensis, which is in general correctly classified. Also, class d) Pericopsis elata is often confused with other several species.