Multiclass Oversampling via Optimum-Path Forest for Tree Species Classification from Street-view Perspectives

• Danilo Samuel Jodas UNESP / USP
• Leandro Aparecido Passos University of Wolverhampton
• Giuliana Del Nero Velasco USP
• Mariana Hortelani Carneseca Longo USP
• Aline Ribeiro Machado USP
• João Paulo Papa UNESP

Resumo

Urban forest surveillance relies on several aspects that involve the analysis of green area preservation and the monitoring of individual trees. Urban trees are essential to maintain the good quality of the cities and reduce the effects of carbon dioxide emissions in the atmosphere. In this sense, one can cite the tree species diversity as essential to ensuring the preservation and proper functioning of the urban ecosystem and the conservation of the wildlife species in the urban forest environment. Furthermore, tree species play an essential role in assessing the tree risk of falling since the species are related to the wood density, thus providing further details for the tree structural analysis. However, tree species classification involves a time-consuming process that requires allocating human resources for fieldwork. Also, the tree species are quite imbalanced in the urban landscape, requiring a more efficient approach to provide accurate results for minority species. Therefore, computer-aided methods are helpful to support the rapid analysis of the tree species for tasks involving inventory and analysis of the tree conditions. This paper proposes a multiclass extension of the O2 PF, an Optimum-Path Forest-based oversampling method, to generate synthetic samples based on features extracted from images of five urban tree species. Further, we present the so-called “Street Level Tree Species Classification”, a novel dataset for tree species classification based on tree images from the ground-view perspective. Four variants of the multiclass O2 PF were tested and compared to several state-of-the-art oversampling methods found in the literature. The obtained results confirm the effectiveness and superior accuracy of the proposed approaches in most cases.

Palavras-chave: Graphics, Surveillance, Wildlife, Urban areas, Green products, Ecosystems, Vegetation, Optimum-Path Forest, deep learning, convolutional neural networks, tree classification, urban forest