Análise discriminante linear em duas dimensões para classificação de dados químicos de segunda ordem

• Ano de defesa: 2017
• Autor(a) principal: Silva, Adenilton Camilo da
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Química; Programa de Pós-Graduação em Química; UFPB
• Palavras-chave em Português: 2D-LDA; Segunda ordem; Classificação; PARAFAC; TUCKER; U-PLS-DA; Second order; Classification; CNPQ::CIENCIAS EXATAS E DA TERRA::QUIMICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/11849

With advances in analytical instrumentation has been increasingly common to obtain second order data by using primarily hyphenated techniques. Despite the advantages obtained by increasing the number of detectors used in sample measurement, the direct data interpretation can be a challenge given the complexity of some matrices. Thus it is important that new chemometric strategies are proposed to support in the interpretation of the data type, such as the twodimensional discriminant linear analysis algorithm (2D-LDA). 2D-LDA was originally proposed in the image processing context for extraction of characteristic vectors with high discriminant power. Despite its promising performance in image processing, the 2D-LDA algorithm has not used in applications involving chemical data. This work investigates the use of 2D-LDA in classification problems involving second order chemical data. Four datasets were used: 2 simulated datasets of excitation / emission matrix fluorescence spectrometry; Auto fluorescence Spectrometry of Parma Ham, Total Synchronous Spectrometry of Edible Vegetable Oil. The results were compared with following algorithms: no feature extraction (NFE); U-PLS-DA (Partial least squares discriminant analysis in unfolded data) and LDA by using TUCKER-3 or PARAFAC scores. In the first simulated data set all models achieved a correct classification rate of 100%. However, in the second simulated data set only NFE model presented classification errors (30%). The Parma ham and vegetable oils data sets obtained the best classification rates by using 2D-LDA and TUCKER-3-LDA (86% and 100%) compared to the models without extraction of characteristics (76% and 77% ), U-PLS-DA (81% and 92%) and PARAFAC-LDA (86% and 92%). In general, the 2D-LDA presented comparable results to the other algorithms and could be considered as a promising strategy in the classification of second order chemical data.