Redução da amostragem em inventários florestais de eucalipto com redes neurais artificiais e sensoriamento remoto
Date
2020Author
Paes, Thuliany Fernandes Araujo
Metadata
Show full item recordSubject
Levantamentos florestaisEucalipto - Cultivo - Sensoriamento remoto
Redes neurais (Computação)
Florestas - Medição
xmlui.dri2xhtml.METS-1.0.item-type
Monografia Especialização DigitalAbstract
Resumo: O inventário florestal tradicional é dispendioso e moroso. Quanto maior o número de amostras, maior serão os custos dessa atividade. Com isso, é crescente o desenvolvimento de estudos que reduzam as atividades de mensuração de parcelas e cubagem de árvores em campo. As redes neurais artificiais já têm se consolidado como ferramentas para redução do esforço amostral e o sensoriamento remoto têm se mostrado como alternativa para a quantificação de biomassa florestal. Essas ferramentas, no entanto, demandam mão de obra específica para verificação de métodos mais apropriados para alcançar resultados. Com isso, o presente trabalho propõe a utilização de imagens gratuitas do satélite Landsat 8 juntamente com técnicas de inteligência artificial para estimação do volume de madeira em plantios de eucalipto na região de Três Lagoas, MS. Foram selecionadas as cenas das áreas estudadas, adicionado o shapefile de parcelas e extraído os dados das bandas 1, 2, 3, 4, 5, 6, 7 e 8 dos valores dos pixels e calculado os índices de vegetação NDVI, SAVI, EVI, GNDVI, ARVI e SR. Treinou-se algumas configurações de redes neurais, a de melhor resultado foi a configuração de talhão como variável categórica, as bandas e índices de vegetação como variáveis numéricas e o volume de madeira com casca (m³cc/ha) como saída. Essa rede foi simulada com a redução do esforço amostral em 28,7%, 43,2% e 64,2% das parcelas e apresentaram erro percentual de 98% no intervalo entre -5% e 5%. A aplicação teve mais de 60% dos valores estimados em nível de talhão se encontram entre ±10% de erro, isso considerando redução de 64,2% das amostras. Com a redução de 28,7% dos dados, essa precisão aumenta para 77,3%. No entanto, é importante ressaltar que são necessários mais estudos quando se trabalha com bases muito extensas de número de parcelas e variação de sítios. Abstract: The traditional forest inventory is expensive and time-consuming. The greater the number of samples, the greater the costs of this activity. With this, the development of studies that reduce the activities of measuring plots and cubing trees in the field is growing. Artificial neural networks have already consolidated themselves as tools for reducing sampling effort and remote sensing has shown to be an alternative for the quantification of forest biomass. These tools, however, require specific labor to verify the most appropriate methods to achieve results. Thus, the present work proposes the use of free images from the Landsat 8 satellite together with artificial intelligence techniques to estimate the volume of wood in eucalyptus plantations in the region of Três Lagoas, MS. Scenes from the studied areas were selected, the plot shapefile was added and data from bands 1, 2, 3, 4, 5, 6, 7 and 8 were extracted from the pixel values and the NDVI, SAVI, EVI, GNDVI, ARVI and SR. Some neural network configurations were trained, the best result was the configuration of stand as a categorical variable, bands and vegetation indexes as numerical variables and the volume of wood with bark (m³cc / ha) as an output. This network was simulated with the reduction of the sampling effort in 28.7%, 43.2% and 64.2% of the plots and presented a percentage error of 98% in the range between -5% and 5%. The application had more than 60% of the estimated values in field level are between ± 10% of error, considering a reduction of 64.2% of the samples. With a 28.7% reduction in data, this accuracy increases to 77.3%. However, it is important to emphasize that more studies are needed when working with very extensive bases of number of plots and variation of sites.