Inteligência artificial na análise de dados de um plantio comercial de Eucalyptus saligna Smith por meio de imagens multiespectrais

• Autor(a) principal: Fantinel, Roberta Aparecida
• Data de Publicação: 2024
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Manancial - Repositório Digital da UFSM
• Texto Completo: http://repositorio.ufsm.br/handle/1/31734
• topic: Aprendizado profundo; Aprendizado de máquina; Aeronave remotamente pilotada; Deep learning; Machine learning; Remotely piloted aircraft; CNPQ::CIENCIAS AGRARIAS::RECURSOS FLORESTAIS E ENGENHARIA FLORESTAL

The advancements in Remotely Piloted Aircraft (RPA), along with artificial intelligence, have been facilitating and enhancing field data collection in terms of temporal and spatial accuracy, with the possibility of creating customized datasets according to specific needs. The main goal of this thesis is to apply artificial intelligence techniques for identification and count estimation of Eucalyptus saligna Smith using high spatial resolution multispectral images onboard a Remotely Piloted Aircraft (RPA). The study area is located in the municipality of Eldorado do Sul - RS. Spectral reflectance curves of leaves were obtained using the Analytical Spectral Devices (ASD) FieldSpec® 3 spectroradiometer. For each of the three selected plant species, five measurements were taken, resulting in a total of 15 readings for each species. For this analysis, the wavelength ranges of the FieldSpec® 3 and those restricted to the four bands (Green, Red, RedEdge, and NIR) of the Parrot Sequoia® multispectral sensor were considered. Subsequently, the data were subjected to the Shapiro-Wilk test to check for normality at the 95% significance level. Given the absence of normal distribution, the non-parametric Kruskal-Wallis test was chosen. The results obtained from the wavelengths of the FieldSpec® 3 and the Parrot Sequoia® sensor demonstrated the ability to characterize and distinguish E. saligna from most spontaneous plants (SP), particularly in the RedEdge and NIR range. At another instance, images of the forest plantation at 180 days post-planting were also acquired via the RPA. Machine learning algorithms (Random Forest - RF) and deep learning (YOLOv8n) were used to identify, detect, and count E. saligna, respectively. The RF algorithm proved to be efficient in identifying E. saligna and the other thematic classes analyzed (SP, exposed soil, and shadow), achieving an overall accuracy of 93%. The YOLOv8n model showed promising results both in detection (recall of 0.93) and count estimation (0.915), demonstrating excellent performance. The results suggest that the use of RPA, multispectral images, and advanced technologies are highly promising for the forestry sector.