Avaliação da invasão de Hedychium coronarium J. König (Zingiberaceae) em florestas ripárias usando algoritmos de aprendizagem de máquina e imagens de veículo aéreo não tripulado (VANT)

• Ano de defesa: 2020
• Autor(a) principal: Arantes, Janaina Goreth
• Orientador(a): Matos, Dalva Maria da Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Campus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ecologia e Recursos Naturais - PPGERN
• Palavras-chave em Português: Plantas invasoras; UAV; Mapeamento
• Palavras-chave em Inglês: Invasive plants; Mapping
• Área do conhecimento CNPq: CIENCIAS BIOLOGICAS::ECOLOGIA::ECOLOGIA APLICADA
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/13378

The riparian zones have undergone great changes due to anthropic activities, among them, the introduction of invasive species. The invasions monitoring in these areas can be complex due to flooded environments and often not reachable, such as the case of monitoring invasions of the species Hedychium coronarium. In this study, a low-cost and easy-to-use integrated methodology for early detection and mapping of this species was proposed, using a technology with the capacity to provide high temporal and spatial resolution data. Two distinct areas were chosen where the images were captured (visible RGB light spectrum) at two times of the year by an unmanned aerial vehicle (UAV). The generated images were classified using machine learning algorithms present in the Dzetsaka ML plugin: the Gaussian Mixture Model, K-Nearest Neighbors and Random Forest. To compare the efficiency of the classifiers according to the form of sampling, different types were used, varying the size of the polygons (6-8 m2 and 12-16 m2) and the number of classes (5 and 8 classes). Classified in 5 classes, with a Kappa index of 78.8% and 80% in June and November, respectively, in the first area and a Kappa index of 72.6% in the second area. The best time for the classification of images was the month of November, possibly due to the better distinction between species and vegetation. We demonstrate promising results in the creation of maps of invaded areas using an adopted methodology, which can subsidize dynamic geospatial models to identify distribution patterns of the studied species and the damage caused. It is also possible to subsidize studies on other invasive species.