Espectroscopia Vis-NIR-SWIR combinada com machine learning para predição de cobre e zinco em solos de vinhedos no sul do Brasil

• Ano de defesa: 2022
• Autor(a) principal: Sangoi, Daniely Vaz da Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/26339/001300000p9t9
• 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; Metais pesados; Pré-processamento; Aprendizado de máquina; Modelo de predição global e regional; Hyperspectral remote sensing; Heavy metals; Pre-processing; Machine learning; Global and regional prediction model; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA::CIENCIA DO SOLO
• Link de acesso: http://repositorio.ufsm.br/handle/1/29565

Careful and sustainable management of vineyard soils is fundamental to the success of vineyards. These areas have been faced with high levels of available copper (Cu) and zinc (Zn) in the soil, as a result of the frequent applications of fungicides based on these elements. This can result in high levels in the soil, resulting in toxicity to plants, also runoff and reach riverbeds, and even leach and reach groundwater. Soil collection and analysis are essential tools to determine the Cu and Zn content available in the soil. However, routine laboratory analyzes are characterized by the destruction of samples and generate chemical residues, which harm the environment if not disposed of correctly. Spectroscopy in the visible, near-infrared and shortwave electromagnetic spectrum (Vis-NIR-SWIR) combined with machine learning methods has shown great advantages and important potential as a technique for predicting soil clay and organic carbon contents. In the scientific and technical field, there is a demand for research related to Vis-NIR-SWIR spectroscopy studies related to other soil attributes, such as Cu and Zn. Therefore, this thesis explores the Vis-NIR-SWIR spectroscopy technique combined with machine learning methods to develop spectral prediction models for available Cu and Zn contents in vineyard soils. The literature shows that the predictive capacity of the models is associated with the characteristics of the soil spectral data and the machine learning method used in the predictive modeling. Therefore, it is fundamental to evaluate the characteristics of the spectral curves, pre-processing and machine learning method in data sets, composed of soils with variation in physical, chemical and mineralogical characteristics, as in the case of subtropical soils in the present study. There is also a need to test the effect of the strategy of stratifying general databases into regional sets on the calibration of prediction models. In STUDY 1, initially, 1482 soil samples were collected from different vineyards in three grape producing regions in the state of Rio Grande do Sul (RS), Brazil: Campos de Cima da Serra, Serra Gaúcha and Campanha Gaúcha. These regions have different levels of Cu and Zn in the soil. Samples were analyzed chemically by Mehlich-1 and spectrally by Vis-NIR-SWIR. The machine learning techniques used were Partial Least Square Regression (PLSR), Cubist (CUB), Support Vector Machine (SVM) and Random Forest (RF). The results identified relevant information in the spectral signatures, which were used to spectrally analyze the available Cu and Zn contents in the soil. It was possible to perform the predictions of these elements due to the inferences with spectrally active soil properties. The highest accuracies were obtained by combining RF with pre-processed spectra using the Savitzky-Golay first derivative technique. As for STUDY 2, the total database (n=1482) was separated into regional subsets based on specific characteristics of the regions where the soil samples were collected. The spectral models were calibrated and after external validation was carried out. With this strategy, it was possible to indirectly estimate the available Cu and Zn contents, with greater accuracy of predictions being observed for models calibrated with regional subsets.