Dissertação

Uso da espectroscopia no infravermelho próximo (FT-NIR) como ferramenta na discriminação de espécies herborizadas de Burseraceae oriundas de diferentes locais da Amazônia Legal

Identification of Amazon forest species is difficult and several new techniques are being developed and tested. Near-infrared spectroscopy (FT-NIRS) what quantifies and characterizes the organic compounds in plant tissue is one method which shows promise in species recognition. We tested different m...

ver descrição completa

Autor principal: Botelho, Izailene Monteiro Saar
Grau: Dissertação
Idioma: por
Publicado em: Instituto Nacional de Pesquisas da Amazônia - INPA 2020
Assuntos:
Acesso em linha: https://repositorio.inpa.gov.br/handle/1/12749
http://lattes.cnpq.br/3608600451755381
Resumo:
Identification of Amazon forest species is difficult and several new techniques are being developed and tested. Near-infrared spectroscopy (FT-NIRS) what quantifies and characterizes the organic compounds in plant tissue is one method which shows promise in species recognition. We tested different methods for obtaining the FT-NIRS spectra of herbarium samples: 1) using the disc supplied with the standard equipment; 2) using a vinyl acetate disc (EVA); 3) using only the mounted leaves on paperboard. We also tested the ability of the technique to recognize species from different geographic regions of the Amazon: 1) samples only from Manaus, 2) samples from other locations, excluding Manaus and 3) samples from the entire region. The data were submitted to principal components analysis (PCA) and linear discriminant analysis (LDA) different cross validation methods. It was observed a high distinction between the exsicate spectra mounted and unmounted in the PCA, the same did not occur between the standard disk and the EVA disk. With the LDA comparing the assembled and unmounted exsicles, the model had recognition of 86.4% and unmounted 93.9%. The percentage of correctness with the EVA disk and the standard disk were similar (88.2% and 90.3%, respectively). The results indicate that, individually, models can recognize species. When these models were compared in a cross validation, the model using mounted material as the base correctly predicted the identity of unmounted material with 71% accuracy, while the inverse accuracy was only 30%. These results indicate that, although the model of samples assembled has a high percentage of correctness, by including the influence of paper in the spectrum, there are variations difficult to control when thinking about a global model, with several herbal and different materials used in the assembly. The evaluation of the influence of geographic variation showed that the model based on samples from the entire Amazon had a prediction accuracy of 90.6%, while the model based on only one locality was 66.5%. In this case, the results indicate that the error may be due to the lack of intraspecific variation of the model and not actually to the error of the identifier. We suggest, therefore, to standardize the spectral collection of exsicatas with materials that do not have as much variation as the use of black EVA; and to capture as much intraspecific variation as possible with samples from different locations to achieve a quality spectral model to identify Amazonian plants.