Please use this identifier to cite or link to this item: http://repositorio.ufla.br/jspui/handle/1/58294
Title: Innovation in foliar analysis using portable X-ray fluorescence spectrometry with aid of machine learning
Other Titles: Inovação na análise foliar usando espectrometria de fluorescência de raios X portátil com auxílio de aprendizado de máquina
Authors: Ribeiro, Bruno Teixeira
Lopes, Guilherme
Costa, Enio Tarso de Souza
Keywords: Nutrição de plantas
Café - Análise foliar
Sensores proximais
Aprendizado de máquina
Plant nutrition
Coffee - Leaf analysis
Proximal sensors
Machine learning
Issue Date: 25-Aug-2023
Publisher: Universidade Federal de Lavras
Citation: COSTA, M. V. da. Innovation in foliar analysis using portable X-ray fluorescence spectrometry with aid of machine learning. 2023. 91 p. Dissertação (Mestrado em Ciência do Solo)–Universidade Federal de Lavras, Lavras, 2023.
Abstract: The assessment of nutritional status of plants is very important for adequate management and guarantee high productivity. Conventionally, foliar analysis has been performed using laboratory-based methods which require sampling, preparation, digestion and determination. This is a complete non environmentally-friendly analysis and time-consuming. Thus, the portable X-ray fluorescence spectrometry (pXRF) can be a promising method for fast, easy, non-destructive and environmentally-friendly foliar analysis. This work aimed to investigate methodological approaches influencing the pXRF performance for plant leaf analysis. In chapter #1, the performance of pXRF for direct analysis of intact leaves from different crops with contrasting water content and anatomy was assessed. Bean, soybean, castor bean, coffee, eucalyptus, mango, guava and corn leaves were analyzed via pXRF at the following conditions: i) intact and wet (fresh) conditions (Cwet); ii) intact and dry conditions (Cdry). The objective was to assess the effect of analyzed leaf surface (adaxial or abaxial) and water content. The pXRF results were used to predict the real concentrations of macro (N, P, K, Ca, Mg, and S) and micronutrients (B, Fe, Cu, Zn, and Mn) using machine learning techniques (Random Forest – RF). The analyzed surface did not influence the pXRF results. Yet the water content has a significant effect on pXRF results underestimating the concentrations, mainly for P and S. The pXRF results obtained directly on intact leaves and modeled by RF very accurately predicted the concentration of both macro- and micronutrients, even those that are not detected by pXRF (N and B). This achievement was attributed to antagonism and synergism uptake relationships. The most important variable for modelling reflected reasonably these uptake relationships. In Chapter #2, the performance of pXRF at different operational conditions was evaluated to access the elemental composition of dried and ground coffee leaves. Coffee leaves from a hydroponic experiment (Hoagland & Arnon) with different concentrations of macronutrients were used. The following operational conditions were assessed: dwell time, calibration (from manufacturer or user), current (A), and voltage (V). The results obtained by pXRF in the best operational conditions were submitted to simple linear regression (LR), multiple linear regression (MLR), and machine learning algorithms (random forest – RF; and support vector machine – SVM) in order to predict the real concentrations of macro- and micronutrients obtained via conventional method, as well as predict categorically the nutritional status as low, adequate, and high. Operational conditions influenced significantly the pXRF performance and must be strongly considered. By parsimony, the elements K, Ca, Cu, and Mn were satisfactorily predicted using LR and MLR. The prediction of all the evaluated nutrients were considerably improved by using machine learning, mainly RF. From this work, it is possible to conclude that pXRF is a really promising method for fast, economical and eco-friendly analysis of plants under laboratory conditions. Furthermore, the direct analysis of plant leaves in the field added to machine learning techniques can substantially facilitate the assessment of nutritional status of plants, contributing for smart farming agriculture and artificial intelligence approaches.
URI: http://repositorio.ufla.br/jspui/handle/1/58294
Appears in Collections:Ciência do Solo - Mestrado (Dissertações)



This item is licensed under a Creative Commons License Creative Commons