Segmentação e classificação de espécies florestais do bioma Amazônico por meio de imagens digitais macroscópicas

Abstract

Deforestation and forest degradation contribute to global warming, making their preservation essential for reducing greenhouse gas levels. One way to achieve this reduction is through forest monitoring by inspection agents. However, this is a complex task due to the extensive biodiversity of forests, with the Amazon rainforest standing out in this regard. To assist in the identification of forest species, various computational techniques have been used, mostly involving images of species along with other elements, which can reduce classification accuracy. To optimize this task, this study proposes the use of image segmentation techniques, such as YOLO11, to isolate the wood before classification using the ResNet-50 model. For training and segmentation, a dataset of 20 species was developed, each containing 50 images. The methodology of this study involved dataset selection, the choice and training of the segmen tation model, image annotation, species classification using both the original images (without segmentation) and the segmented images (generated from YOLO11 segmentation masks), and the analysis of results using evaluation metrics such as Intersection over Union (IoU), Precision, Recall, F1-score, and Dice coefficient. The experiments demonstrated that, although YOLO11 performed well in segmentation, background removal did not result in an improve ment in the final classification accuracy. Thus, the model using the original images achieved superior performance compared to the segmented images, with an accuracy of 91% versus 87%.