Abstract
Tree recognition depends on the presence of the plant’s morphological characteristics, such as leaves, flowers, roots, stems, and trunks. However, the transformation of a tree into processed timber eliminates these botanical characteristics. In this context, wood anatomy analysis becomes an alternative method for recognition when other characteristics are absent. The objective of this study was to demonstrate the performance of a convolutional neural network in recognizing, through wood anatomy, seven specimens identified at the species level (Bertholletia excelsa Bonpl., Cariniana estrellensis (Raddi) Kuntze, Erisma uncinatum Warm., Manilkara huberi (Ducke) A. Chev., Mezilaurus itauba (Meisn.) Taub. ex Mez, Grevillea robusta A. Cunn. ex R. Br., and Melia azedarach L.) and three taxa identified at the genus level (Hymenaea sp. Mart., Dipteryx sp. Aubl., and Machaerium sp. Pers.). The images used were collected from the “Forest Species Database - Macroscopic,” provided by the Robotics Vision and Image Laboratory of the Federal University of Paraná (UFPR). A total of 550 images of wood anatomy in JPG format, with a resolution of 3264 × 2448 pixels, were used. Through tests conducted with the developed convolutional neural network, confusion matrix, precision, recall, and F1-score metrics were obtained. The results demonstrated the feasibility of recognizing timber forest species using artificial intelligence (AI) through convolutional neural networks, achieving an accuracy of 90% in the conducted tests, demonstrating this technology as a promising tool for timber forest species recognition.