Online near-infrared spectroscopy for soil attributes prediction

• Autor(a) principal: Canal Filho, Ricardo
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/11/11152/tde-21032023-182112/
• topic: Agricultura de precisão; Aprendizado de máquina; Diffuse reflectance spectroscopy; Espectroscopia de reflectância difusa; Machine learning; Precision agriculture; Proximal soil sensing; Sensoriamento proximal do solo; Soil variability; Variabilidade do solo

Precision agriculture (PA) is based on the identification of spatial and temporal variability of the attributes that influence agricultural production. In this sense, techniques that allow monitoring soil and crops in high spatial density have been studied by the PA community. Diffuse reflectance spectroscopy (DRS) is a technique that allows, especially in the near-infrared (NIR) region, to acquire online soil spectra, embedding sensors in agricultural machines. The use of this technique allows data acquisition in high spatial density, which, together with machine learning (ML), are transformed into quali-quantitative data of soil attributes. However, in tropical soils, especially in Brazil, this research area is still poorly developed compared to studies from Australia, the United States of America and Europe. The research project of this dissertation was proposed to expand the development of the technique in Brazilian tropical soils. An experimental area of the University of São Paulo, in Piracicaba-SP, was used to acquire online soil NIR spectra. Different statistical models were tested to predict soil chemical and physical attributes. Calibration and use protocols of DRS in the field were evaluated. The main findings of this dissertation were organized into three chapters. The first one addresses calibration protocols regarding the use of spectrum preprocessing techniques and different statistical models. The results suggest that the use of raw data combined with dimensionality reduction statistical models offer the most efficient strategy for calibration of predictive models. The second chapter addressed the insertion of samples from different areas in the calibration of ML models. The results showed more robust predictions when models were calibrated only with samples from the experimental area itself, denoting the importance of local calibration for the use of DRS NIR in online acquisition. In the third and last chapter, the area was revisited on a second day of spectral acquisition, three weeks after the first one, following the same experimental and instrumental criteria. The ML models calibrated on the first day were tested for prediction of soil attributes with spectra from the second day of acquisition. Low predictive performance of the models was reported in this scenario, indicating the need for local calibrations not only in space, but also in time, for the technique to perform properly. The results reported in this dissertation prove the potential of the technique for agriculture, as they show that it is possible to predict soil attributes with online NIR spectra. Furthermore, this work can help in the development of PA practices, and offer guidelines for future research that seek the development of DRS for prediction of soil attributes in the field, to establish its large-scale use in agriculture.