Reconhecimento de espécies florestais do bioma Amazônico, utilizando técnicas de deep learning

Abstract

Tropical forest biomes, found in South America, Africa, and Asia, are fundamental ecosystems due to their biodiversity, warm climate, and high humidity. The Amazon Rainforest, shared by nine South American countries, is one of the most notable examples of this biome. However, it faces increasing threats due to deforestation and wildfires, driven by economic expansion and illegal logging. This degradation compromises the forest’s ability to regulate the global climate, capture carbon dioxide (CO2), and preserve its biodiversity. Accurate identification of illegally exploited tree species is essential to mitigate these impacts, but conventional methods are time-consuming and prone to errors. In this context, machine learning techniques have been explored for automated species recognition. However, the similarity between species presents an additional challenge. This study evaluates the use of Deep Learning (DL) for forest species identification in the Amazon, employing the YOLOv8 model. The model was trained with macroscopic transverse and tangential images, combined with organoleptic characteristics, aiming to improve classification accuracy. The results indicate that fusing these data reduced critical prediction errors and increased model reliability. The lowest performance was observed in the Tangential model, with an accuracy of 0,992, while the Hybrid Model with Organoleptic Characteristics achieved the best result, reaching 0,999. The model’s robustness was confirmed even under adverse conditions, such as noise and lighting variations, highlighting its applicability in environmental monitoring. The proposed technology represents a significant advancement in automating environmental enforcement, contributing to the fight against illegal deforestation and the conservation of the Amazon Rainforest.