Reconhecimento de espécies florestais brasileiras ameaçadas com aprendizagem de máquina

Abstract

This work presents a model for the recognition of Brazilian forest species from macroscopic images of wood textures. The approach combines Inception v3 as a feature extractor and Random Forest as the classifier. The main goal of this work was to evaluate augmentation and preprocessing techniques, with especially good results on splitting the images into smaller patches in order to reduce distortions caused by resizing and enrich the dataset. The model achieved an average F1-Score of 97.41%. It is expected that this study will contribute to combating the illegal harvesting of endangered species and support environmental conservation efforts through ecosystem monitoring.