Assessment of coffee leaves nutritive value via portable X-ray fluorescence spectrometry and machine learning algorithms

Portable X-ray fluorescence (pXRF) spectrometry is a non-destructive technique that has been successfully used to analyze different matrices. Foliar analysis is challenging because some plant nutrients cannot be detected by pXRF. Even so, the uptake interactions among nutrients which reflect different concentrations of macro- and micronutrients can be accessed via pXRF, constituting a basis to obtain predictive models of plant nutrients. The objective of this work was to compare and assess the accuracy of linear regression and non-linear models (support vector machine – SVM; and random forest – RF) to predict the actual concentration of macro- (N, P, K, Ca, Mg, and S) and micronutrients (B, Cu, Fe, Zn, and Mn) in coffee leaves. A greenhouse experiment was conducted using Hoagland and Arnon solution to obtain leaves with contrasting elemental composition. Ground and oven-dried leaf samples were analyzed via pXRF using: i) a manufactured pXRF calibration developed for general earth-materials (Geoexploration mode); ii) the Spectrometer mode with varying voltage and current (15 kV and 25 μA; 10 kV and 10 μA). The same samples were also analyzed via conventional acid digestion and quantified via inductively coupled plasma-optical emission spectroscopy (ICP-OES). The best predictions were obtained using SVM and RF algorithms, with high R2 (0.82 to 0.99) and high residual prediction deviation (RPD) (2.35 to 9.34) values. However, some elements (e.g., K, Ca, Cu, Mn) were successfully predicted using linear models (LR and MLR). Even elements not detected (N and B) by pXRF were accurately predicted using the RF model. The pXRF operational conditions influenced the performance of the models. However, by parsimony, Geoexploration mode provided data for accurate prediction of macro- and micronutrients. This comprehensive study can potentially spark further investigations into examining coffee leaves from plants cultivated under various environmental and management conditions. Additionally, the methodological framework outlined here holds promise for ongoing experimentation across diverse crops, offering a streamlined, non-invasive, eco-friendly, and rapid approach for foliar analysis. [Display omitted] •Concentrations of all coffee leaves nutrients were accurately predicted.•Machine learning models of pXRF data speed foliar analysis and eco-friendliness.•Random forest outperformed other predictive models.•Even undetectable nutrients (N and B) by pXRF were successfully