Uso de bibliotecas espectrais para a predição do carbono orgânico do solo

• Ano de defesa: 2018
• Autor(a) principal: Bueno, Jean Michel Moura
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Agronomia; UFSM; Programa de Pós-Graduação em Ciência do Solo; Centro de Ciências Rurais
• Palavras-chave em Português: Sensoriamento remoto hiperespectral; Predição de carbono orgânico; Processamento espectral; Modelos de aprendizado; Estratificação de dados espectrais; Hyperspectral remote sensing; Prediction of organic carbon; Spectral processing; Learning model; Stratification of spectral dataset; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA
• Link de acesso: http://repositorio.ufsm.br/handle/1/16306

This study demonstrates the use of diffuse reflectance spectroscopy (DRS) applied to soil organic carbon (SOC) prediction. The scope of the study is related to the increase in the demand for soil information to support environmental monitoring, agricultural production and to feed models for simulation of future scenarios. DRS associated to soil spectral libraries (SSL) is an alternative for quantification of SOC content. However, the predictive ability of the models is linked to the characteristics of soil spectral data. Therefore, it becomes essential to evaluate the construction of models when the SSL is composed by samples with high variability in physical, chemical and morphological characteristics, which is the case for soils in the south of Brazil. Considering that the accuracy of the models is defined by the complexity and coverage of pedological features of soils represented in the SSL, associated to techniques of spectral processing and multivariate methods, the objectives of this study were: i) evaluate the effect of preprocessing techniques, multivariate methods and sample stratification of SSL on the accuracy of the models; ii) identify the influence of the spectral complexity of the samples on the quality of predictions; iii) define the minimum features of the SSL that might impact their predictive ability. In STUDY 1, a local SSL composed of 841 samples was used. With this SSL, the influence of preprocessing techniques, multivariate methods and SSL stratification based on spectral variance, soil class, land use and soil sample depth on SOC prediction was assessed. In STUDY 2, a regional-scale SSL composed of 2,599 samples was used. With this SSL, the effect of stratification of the SSL based on regional environmental features, soil texture, land use and spectral class on the accuracy of models was evaluated. Soil spectra presented high spectral variation, due to the pedological and environmental variation of the southern region of Brazil. More accurate models were obtained with the Savitzky-Golay spectral processing technique - 1st derivative associated with the partial least squares regression calibration method, and with the Cubist method with continuous removal processing. The stratification of the SSL based on soil and environmental (physiographic regions) characteristics showed that grouping more homogeneous samples, especially in relation to physiographic regions and land uses, increased the accuracy of the predictions. The reduced number of samples due to stratification negatively affected the performance of the models, especially for groups with high pedological and spectral variation. SOC predictions presented lower accuracy for samples with coarse texture (sand > 15 % and clay < 35 %). The results confirm that the use of SSL for SOC prediction requires a previous study on the data variance.