Espectrorradiometria e análise multivariada na predição dos atributos químicos e físicos dos solos no noroeste do Paraná

• Ano de defesa: 2013
• Autor(a) principal: Romagnoli, Franciele
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Maringá; Brasil; Departamento de Agronomia; Programa de Pós-Graduação em Agronomia; UEM; Maringá, PR; Centro de Ciências Agrárias
• Palavras-chave em Português: Solos; Espectrorradiometria; Modelo de predição; Análise multivariada; PLSR; Regressão Stepwise; Paraná; Brasil.; Soils; Spectroradiometry; Prediction model; Multivariate analysis; Stepwise regression; Brazil.; Ciências Agrárias; Agronomia
• Link de acesso: http://repositorio.uem.br:8080/jspui/handle/1/1233

This study aimed to evaluate two methods of multivariate analysis: PLSR through the chemometric ParLes program and multiple linear regression with Stepwise variable selection by SAS in the prediction of soil chemical (SiO2, Al2O3, Fe2O3, TiO2 and C.O.) and physical attributes (sand, silt and clay) using reflectance data of a laboratory spectroradiometer. The areas chosen for study are located in the northwest of Paraná in lithosequence located in the transition range between Basalt and Sandstone. Soil samples were collected characterizing the A and B horizons and subjected to physical, chemical and mineralogical analyzes at the Universidade Estadual de Maringá. The spectral readings were taken in a spectroradiometer within the range of 350 to 2.500 nm, which allowed the prediction of soil properties through its spectral response. Two methodologies were adopted for the construction of predictive models; the first was the use of bulk samples (HA + HB) for the prediction of attributes. The second method was the separation of the HA samples (0-0,20 m) and HB (0,80 - 1 m) and consequently, obtaining a prediction model for the horizons separated diagnostically. The aggregate samples (A and B horizons), allowed to obtain prediction models with better performance than the models adjusted for the samples belonging to A and B horizons, both for PLSR and for multiple linear regression with Stepwise selection of variables for almost all studied attributes, except only for the C.O., which had better performance for the HA. Silt, C.O. and TiO2, were the attributes that had the worst performances, both in the calibration and validation stages for the two statistical methods. Parles and SAS showed similar performances for the prediction models calibration stage. For the validation stage, with adjusted models from xv the bulk samples, the PLSR technique overcame the Stepwise regression for the attributes Al2O3, TiO2 and C.O. and they were similar for the other attributes. For the HA models, the PLSR technique performed better than the Stepwise regression for the attributes Fe2O3 and silt and worse for the SiO2 and TiO2. Meanwhile for the HB, the Stepwise regression allowed the adjustment of better models for TiO2 and C.O. and similar for the other attributes. Therefore, the conclusion is that the use of bulk samples generated prediction models with better performances, except for C.O.. The Stepwise regression method is the most suitable since it provides the prediction model.