Identificação de espécie arbórea invasora em imagens panorâmicas urbanas por aprendizagem profunda

Abstract

Mapping tree vegetation contributes to urban planning and facilitates the management of green areas in cities. Identifying tree species helps in environmental diagnosis and allows us to verify whether the species present in urban areas contribute to the balance and provide ideal ecosystem services. Green areas are not always composed of the recommended number of native species, and may contain exotic tree species capable of harming the urban environment, as certain species become invasive. In this context, the leucaena tree species (Leucaena leucocephala) is frequently present in several cities and corresponds to an invasive exotic species in several countries, including Brazil. Remote sensing data processed by artificial intelligence algorithms (especially deep learning) can be an efficient and effective strategy for identifying Leucaena in urban environments. However, this is a persistent gap in the literature. In this work, we seek to explore the potential of deep learning algorithms (detection and semantic segmentation) to identify the Leucaena tree species in urban regions using panoramic images from Google Street View and cell phones. These images are being explored to identify Leucenes both from the perspective of semantic segmentation and object detection. The SAM model (Segment Anything Model), launched in April 2023, was tested on these images to initially differentiate tree vegetation without taking the species into account. The partial results indicate SAM's ability in this task, reaching an F1-score of 99%. In tests with SAM for the Leucaena species, the model proved to be more advantageous in Street View images when compared to cell phone images, with an average accuracy of 92% and 63% respectively. From an object detection perspective, tests were carried out with the YOLOv8 and DeTR networks. Launched in January 2023, YOLOv8 is one of the most recent options for the task of object detection in images, while DeTR corresponds to a transformer-based network and has the potential for several tasks. The results indicate a promising capacity of YOLOv8 in the proposed task of detecting the Leucaena species in panoramic images. In addition to the tests, a scientometric review was constructed to investigate how the problem of tree mapping in general, and by species, has been addressed using remote sensing data and machine learning (shallow and deep). The tests carried out reinforce the originality of this work, showing a lack of studies that explore this type of data with advanced deep learning techniques, to identify the presence of one of the 100 worst invasive species in the world, the Leucaena.