Comparação entre o método de segmentação supervisionada e não supervisionada de modelos deep learning para identificação de indivíduos Pinus spp. Invasores em área de campo úmido

Abstract

The introduction of Pinus spp. in Brazil, it causes a series of environmental disorders and can lead to the impoverishment of biodiversity through its invasive potential. Thinking about the difficulty of obtaining quantification and geolocation data of individuals, this work aims to compare two deep learning models for automated segmentation of Pinus spp. crowns. The comparison was carried out between a model manually trained by the Mask R-CNN algorithm (supervised classification) and an existing model, SAM (unsupervised classification). Mask R-CNN presented 81% intersection over the union of polygons and 88% overlap rate between the manually delimited Masks and the masks segmented by the algorithm, while SAM presented results of 72% and 77%, respectively. The results found reinforce the importance of studies related to new remote sensing technologies for monitoring native vegetation as a biological conservation tool.