Capacidade produtiva, estimativa de volume e biomassa em plantações de Schizolobium parahyba var. amazonicum com o uso de imagens sentinel 2

• Ano de defesa: 2019
• Autor(a) principal: Batista, Fábio de Jesus
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/26339/0013000013qxb
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Recursos Florestais e Engenharia Florestal; UFSM; Programa de Pós-Graduação em Engenharia Florestal; Centro de Ciências Rurais
• Palavras-chave em Português: Algoritmo de aprendizado de máquinas; Índice de vegetação; Paricá; Volume de madeira; Ulianópolis; Machine learning algorithm; Vegetation index; Volume of timber; CNPQ::CIENCIAS AGRARIAS::RECURSOS FLORESTAIS E ENGENHARIA FLORESTAL
• Link de acesso: http://repositorio.ufsm.br/handle/1/19024

The aim was to estimate the volume (VT) and biomass (BA) of Paricá plantations with machine learning from the images of MSI/SENTINEL-2A, in Ulianópolis, Pará. In three productive areas (PO2, CAP2, and PO) of the species, 56 sample units (UAs) were installed for the forest inventory. The productive capacity was evaluated by the site index (IS) based on ANATRO of 28 dominant trees. Covariance analysis was applied on the IS model. Soil samples were collected from 0-20cm (chemical) and 20-40cm (physical). Three UAs per area were randomized for the harvesting and cubic scaling of the trees. BA was performed by the direct method, considering 10 trees Dg per stand. The T22MHA scene was downloaded from 26/07/2016. In Qgis, a 1A product was generated consisting of stack of bands B2 to B12. In 47 UAs, the reflectance of pixel/band was extracted for the calculation of the vegetation index (IV). GLM was applied to model VT. The estimation of BA was done from the FEBmean. The prediction of VT and BA by the sensor considered 50 IVs and 12 bands, where by cross-validation, the most accurate algorithm was defined among the tested ones (Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN)). The pre-selection of the 10 most important variables for the spatialization of VTpredicted and BApredicted was performed by the RF. The analysis was done by RStudio 3.5.2. The precision of the inventory was <10% in CAP2 and PO areas, and in PO2 was 13.50%. For the IS, sample errors occurred of 15%, 9%, and 11% for PO2, CAP2, and PO areas, respectively. The model height-age of Schumacher, =3,4531 , was adjusted for GLM from the Gama-Identity distribution. The sites were divided into high (21 to 25m), medium (19 to 21m), and low (15 to 19m) productivity. The high productivity was registered in 80% of the UAs of CAP2, 50% of PO2, and 8% of PO. From the 36º month-old, different growth rate was verified. The covariance analysis differentiates the sites more (PO2 and CAP2) and less (PO) productive. The topographic characteristics, the presence of more clay and moderate soil acidity were relevant to turn CAP2 more conducive to the productivity of the species. The function for the estimation of VT in PO2 (88.96m³.ha-1 ± 14.50) and CAP2 (152.35 m³.ha-1 ± 16.45) was in Naslund – Gaussian. The function for VT in PO (139.37m³.ha-1 ± 28.41) was Meyer – Gaussian. The stem contributed with 86.54% of BA, branches and leaves participated with 8.28% and 5.18%. The BA registered for PO2, CAP2, and PO were 34,53ton.ha-1 ± 5,63; 56,54ton.ha-1 ± 7,75; and 51,93ton.ha-1 ± 11,95, respectively. The comparisons between VTobserved and VTpredicted, defined by ANN showed similarities in CAP2 and in PO. The comparisons between BAobserved and BApredicited calculated by RF showed proportionality in CAP2. The most precise estimation of VT and BA occurred to CAP2. The differences in PO2 and PO do not reflect a statistical problem, but rather spectral mixtures.