Forest species identification using macroscopic images and convolutional neural networks

Abstract

Automatic identification of forest species from macroscopic wood images is a growing operational demand for environmental enforcement, customs control, and commercial verification, in the context of the fight against illegal timber trade and the conservation of forest biodiversity. This research develops a deep learning pipeline for this task, emphasizing the construction of a heterogeneous database, the systematic selection of a suitable architecture, and the in-depth analysis of the chosen model in terms of accuracy, robustness, and interpretability. A unified corpus was assembled from three distinct sources, covering 73 forest species, with scale calibration across capture devices and patch extraction at 448 × 448 pixels with controlled overlap. Five representative architectures (VGG16, ResNet50, InceptionV3, MobileNetV2, and YOLOv8) were compared under identical conditions, establishing the YOLO family as the basis for a benchmark of fifteen models spanning three generations (YOLOv8, YOLO11, and YOLO26) and five sizes (nano to extra-large). Four models achieved top-1 accuracy of 99.87% (MCC = 0.9987) and image-level majority voting yielded 100% accuracy on the 1,714 test images for YOLO26- medium, the configuration that combines the best balance between accuracy and computational efficiency, with inference in 2.13 ms per patch. Robustness was characterized by three complementary out-of-distribution detection methods on two distinct scenarios, revealing a dissociation between closed-set accuracy and rejection capability. Interpretability combined t-SNE projections and multiscale RISE saliency maps, evidencing that residual errors reflect anatomical ambiguities inherent at the macroscopic level. Complementary robustness analyses characterized the sensitivity to field image degradation, with sensor noise as the most critical factor, and the generalization across capture devices, showing that exposure to field data during training is decisive and that a few examples per species recover the performance through few-shot adaptation. The results establish a reference for macroscopic forest species identification and indicate the feasibility of deploying intermediate-sized models on edge devices, that is, portable, resourceconstrained hardware that runs inference locally, such as smartphones used in the field.