Desenvolvimento de aplicativo mobile multilíngue para reconhecimento de espécies florestais de diferentes biomas

Abstract

Illegal logging represents one of the main threats to forest ecosystems, demanding rigorous enforcement actions. However, this process requires visual analysis of the macroscopic characteristics of wood, as documentation can be falsified. Such identification demands trained professionals, whose training takes time, and who are often scarce and unable to meet the high demand. In this context, accurate identification of forest species is essential to combat deforestation and illegal timber trade. A portable solution capable of assisting in this identification can significantly support environmental enforcement efforts. This work proposes the use of deep learning algorithms for wood image classification, with subsequent integration of the model into a mobile application, aiming to serve as a field support tool. Two deep learning architectures were evaluated: YOLOv11 and MobileViT. Both were trained on a unified dataset composed of three distinct image collections. Originally comprising 79 species, the dataset was curated and refined, resulting in a total of 66 species used in the main experiments. The tests showed that the YOLOv11 model achieved the best performance, with an accuracy of 0.994 and an F1-score of 0.996, surpassing MobileViT. Based on these results, YOLOv11 was integrated into the mobile application developed in Java, using the ONNX Runtime library for model execution. It is concluded that developing a tool capable of recognizing forest species directly on mobile devices is feasible and can become a valuable asset in enforcement operations.