Spectroscopic analysis (UV-VIS-NIR) for predictive modeling of macro and micronutrients in grapevine leaves

Smart Agricultural Technology 10 (2025) 100812 Available online 1 February 2025 2772-3755/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Spectroscopic analysis (UV-VIS-NIR) for predictive modeling of macro and micronutrients in grapevine leaves J.I. Manzano a,* , M. Rodríguez-Febereiro b , M. Fandi˜ no b , M. Vilanova a,c,* , J.J. Cancela b,c a Instituto de Ciencias de la Vid y del Vino (ICVV), Consejo Superior de Investigaciones Científicas-CSIC, Universidad de La Rioja, Gobierno de La Rioja, Finca la Grajera, Carretera de Burgos, Logro˜ no 26080, Spain b GI-1716, Proyectos y Planificaci´ on, Departamento Ingeniería Agroforestal, Escola Polit´ ecnica Superior de Enxe˜ naría, Universidade de Santiago de Compostela, Rúa Benigno Ledo s/n, Lugo 27002, Spain c CropQuality: Crop Stresses and Their Effects on Quality (USC), Associate Unit of Instituto de Ciencias de la Vid y del Vino (ICVV-CSIC), Spain ARTICLE INFO Keywords: Nutritional diagnosis Vine PLS-R Chemometrics Spectroscopy NIR Macronutrients Micronutrients ABSTRACT Assessing nutrient concentrations in grapevines is crucial not only for the overall physiology of the plant but also for the quality of the resulting wine. Accurate determinations are also relevant for enhancing nutrient use efficiency and formulating fertilizer recommendations. Hence, there is a considerable demand for a swift technique to analyze vine organs. Diffuse reflectance spectroscopy coupled with chemometric methods emerges as a potent, cost-effective, and environmentally friendly analytical technique for determining nutrient concentrations in plants. The objective of this study is to ascertain the viability of wide range spectrum (190–2600 nm) spectroscopy in providing precise insights into the nutritional status of vines. Our investigation specifically targets on the determination of C, N, P, K, Ca, Mg, B, Cu, Fe, Mn, Zn, Na, and Al in vine leaves from different wine growing areas, varieties and harvest years. Partial Least Squares Regression (PLS-R) was employed to construct models for the concentrations of these nutrients based on the reflectance measurements of the leaves. The model was trained using 70 % of the samples, while the remaining 30 % constituted the independent validation. Results from the validation set indicated accurate validation for most nutrients, with determination coefficients (r 2 ) of 0.70 for C, 0.72 for N, 0.64 for P, 0.75 for K, 0.84 for Ca, 0.48 for Mg, 0.45 for B, 0.58 for Cu, 0.26 for Fe, 0.82 for Mn, 0.50 for Zn, 0.90 for Na, and 0.69 for Al. The findings revealed that reflectances in the visible (VIS) region of the spectrum played a key role in predicting micronutrients like B, corresponding with photosynthetic pigments (chlorophylls and carotenoids). In contrast, reflectances in the near-infrared region (NIR) had a greater impact on macronutrient prediction, particularly for P and Mg, due to their stronger interaction with organic compounds. The ultraviolet (UV) range played a minor role, highlighting the predominant importance of the VIS-NIR regions in spectroscopic analyses. Finally, the results support the potential of this technique for swiftly and non-invasively predicting both macro and micronutrient levels in grapevine plants, and facilitate the fertilization planning using variety-specific reference levels, or precision viticulture adapted to site-specific demands, including spatial intra-plot variability. 1. Introduction The growth and development of grapevines are heavily influenced by both macro and micronutrients. Deficiencies or imbalances in these nutrients can lead to stunted growth, reduced fruit quality, and increased susceptibility to diseases [11]. Therefore, growers utilize fertilization to modify the nutritional status of the vines, seeking to achieve the right balance between vegetative growth, yield, and crop quality [6]. Hence, understanding the current nutritional status of grapevines is crucial for creating an effective fertilization plan. Reducing excessive fertilizer application not only cuts costs but also enhances quality and reduces the risk of contamination [1]. Moreover, the exact application of a certain nutrient, through a fertilization schedule adapted to the requirements of each phenological stage, according to age, rootstock-cultivar combination, and nutrient application system (fertigation, spraying machine, etc.) requires knowledge of the * Corresponding authors at: Instituto de Ciencias de la Vid y del Vino (ICVV), Consejo Superior de Investigaciones Científicas-CSIC, Universidad de La Rioja, Gobierno de La Rioja, Finca la Grajera, Carretera de Burgos, Logro˜ no 26080, Spain. E-mail addresses: [email protected] (J.I. Manzano), [email protected] (M. Vilanova). Contents lists available at ScienceDirect Smart Agricultural Technology journal homepage: www.journals.elsevier.com/smart-agricultural-technology https://doi.org/10.1016/j.atech.2025.100812 Received 30 October 2024; Received in revised form 30 January 2025; Accepted 31 January 2025 Smart Agricultural Technology 10 (2025) 100812 2 nutritional status quickly and efficiently, in order to implement precision viticulture concepts [37]. The essential nutrients are sixteen and encompass carbon (C), oxygen (O), hydrogen (H), nitrogen (N), phosphorus (P), potassium (K), calcium (Ca), magnesium (Mg), sulfur (S), iron (Fe), manganese (Mn), zinc (Zn), copper (Cu), boron (B), molybdenum (Mo), chlorine (Cl), and nickel (Ni). C, N, P, K, Ca, and Mg are classified as macronutrients, essential in larger quantities, while Fe, Zn, Cu, B, Mn, Sodium (Na), Aluminium (Al) and other elements are considered micronutrients, needed in smaller amounts, for the optimal growth and development of crop plants [19]. C is not typically categorized as a traditional nutrient for plants in the same sense as others like nitrogen, phosphorus, and potassium, but plays a fundamental role in their growth and development. Plants absorb carbon dioxide (CO 2 ) from the atmosphere during photosynthesis, utilizing it to produce sugars and other organic compounds necessary for their metabolism and structure. These organic compounds serve as energy sources for various physiological processes and as building cellular components such as cellulose, proteins, and lipids [8]. N stands out as the primary nutrient significantly influencing the vegetative development of plants; it serves as a pivotal element for various physiological processes. Deficiency in N manifests as reduced leaf size and evident yellowing [4], where a sustainable nitrogen fertilization plan is required depending on production objectives: yield or quality [37]. P is indispensable for the formation of crucial macromolecules like nucleic acids, phospholipids, and sugar phosphates utilized by plants for the development of different organs [28]. K is relevant in agronomic production, being essential for enzymatic reactions, maintenance of osmotic potential, and facilitation of water uptake during plant maturation. At the vineyard level, potassium plays a crucial role in regulating key physiological mechanisms, particularly the transpiration process and the opening and closing of stomata [33]. In wine, however, excessive potassium levels are common and result in decreased acidity and shorter shelf life, leading to accelerated oxidation and unstable coloration [13]. Ca is mainly related with root growth and development, facilitating nutrient and water absorption from the soil. Initial signs of Ca deficiency appear on young leaves, exhibiting distinct deformations and chlorosis. Mg primarily serves as a constituent of chlorophyll, cell walls, and play a vital role in P translocation and N assimilation. Mg deficiency is typified by chlorosis along leaf margins [1]. Regarding micronutrients, Fe is essential for plant playing a crucial role in various metabolic processes, particularly in chlorophyll synthesis, photosynthesis, and nitrogen fixation. Iron is a component of several enzymes involved in these processes, including catalase, peroxidase, and ferredoxin. Mn contribute to perform photosynthesis and in defense against oxidative stress. B is involved in cell wall formation, carbohydrate metabolism, nucleic acid synthesis and hormone regulation. The deficiency can lead to symptoms such as stunted growth, distorted leaves, and poor flower and fruit development, while excess of B can result in toxicity symptoms like leaf burn and necrosis. Zn is part of chlorophyll molecule and in consequence is related with the photosynthetic process. Zn deficiency can lead to symptoms such as stunted growth, interveinal chlorosis, and reduced yields. Cu is involved in carbohydrate metabolism and photosynthesis [39]. Na is not considered an essential nutrient for most plants, but many plants are sensible to high concentration of this compound. The excessive levels of sodium disrupts the osmotic balance within plant cells. This disruption can inhibit water uptake by plant roots and lead to dehydration and wilting. Additionally, high sodium levels can interfere with the uptake of essential nutrients by plants, further exacerbating nutrient deficiencies [35]. Al is a complex element with both positive and negative impacts on plant growth and development. Its effects depend on factors such as soil pH, aluminium concentration, and the specific plant species involved. In agriculture, managing soil pH and aluminium levels through practices such as liming and soil amendment can help mitigate aluminium toxicity and optimize plant growth [24]. Utilizing ultraviolet (UV), visible (VIS), and near-infrared (NIR) spectroscopy together with chemometric methods offers a promising solution for rapid and reliable determinations of diverse physicochemical parameters. Spectroscopy requires minimal sample preparation, typically only necessitating drying and grinding to mitigate water’s influence on spectral absorbance. NIR absorbance spectra offer insights into the organic matrix through the detection of C–H, O–H, and N–H vibrational modes. While macroand microminerals lack these specific vibrational modes, they are embedded within the organic matrix and interact with one of the mentioned organic functional groups. Consequently, their concentrations can be indirectly assessed through NIR spectra analysis [9]. The calibration models exhibited greater accuracy for macronutrients, attributed to their higher concentrations. This outcome might be supported by the fact that elevated nutrient levels in plant tissue typically result in a more pronounced signal-to-noise ratio, thereby facilitating better interpretation of the response through chemometric tools. Additionally, stronger correlations with organic compounds could contribute to this phenomenon [16]. Multiple research endeavours have validated spectroscopy’s capability to determine nutrient concentrations across various crops. Nevertheless, there are still relatively few studies specifically centred on vine leaves or stems. In this way, the level of different nutrients like P, K, Ca, Mg, Mn, Fe, Cu, Zn, and B were described in different vine organs (leaf blades, petioles and berries) using predictive models based in NIR [11]. Thus, estimation of soil nutrients (N), organic matter, and clay in vineyard performing Partial Least Squares Regression (PLS-R) and random forest models with different spectrum ranges and different pretreatments was previously described [29]. Regarding wine and must, there are numerous studies on the use of NIR spectroscopy for the analysis of various quality-related oenological parameters, such as alcohol content, pH, volatile acids, organic acids and reducing sugars in grapes and wine [9,18,27,36]. In other fruit trees, complete nutritional diagnosis in citrus leaves was described in [1] stablishing accurate models for the determination of P, K and B, although the list of nutrients studied included N, Ca, Mg, S, Fe, Cu, Mn, and Zn. Similar studies were realized in different species of citrus leaves determining N, K, Ca, Mg, B, Fe, Cu, Mn, and Zn [2,14,23]. Furthermore, it has been described in [16] that integrating NIR and mid-infrared (MIR) spectra exhibited promising results for assessing both macronutrient (N, P, K, Ca, Mg, and S) and micronutrient (Na, Fe, Mn, B, Cu, Mo, and Zn) concentrations in rice plants. Finally, and conversely to the theory of the lack of specific spectral signatures for micronutrient determinations, some studies have demonstrated reliable prediction accuracy in estimating specific micronutrients (Na, B, Al, Mn, Cu, and Zn) in certain plant species using NIR spectroscopy [7,30]. The European Union (EU) is dedicated to advancing sustainable agriculture and aims to reduce fertilizer usage by 2030, a key objective outlined in its European Green Deal and Farm to Fork strategy (F2F, 2020. Farm to Fork Strategy. Web address: https://ec.europa.eu/food/far m2fork_en). To accomplish this objective, it is crucial to develop tools and reliable predictive models for nutrient assessment, which are essential for formulating informed and sustainable fertilization strategies. In this study, we successfully described the use of PLS-R generated models employing data obtained from spectroscopy over a wide spectrum range, determining the nutritional status of grapevine plants from leaf blades samples. 2. Materials and methods 2.1. Experimental design The plots used for this study were located in different wine growing areas in Spain Galicia: Rías Baixas (Val do Ulla; Val do Saln´ es; O Rosal), Ribeira Sacra (San Victorio; O Verxel; Alais; A Ermida; Santa Cubicia; Quiroga), Ribeiro (Leiro), b) Castilla La Mancha: Albacete (ITAP), c) Extremadura: Badajoz (CICYTEX), and d) Castilla-Le´ on: Valladolid (ITACYL), during the seasons of 2020, 2021 and 2022 (Fig. S1). The J.I. Manzano et al. Smart Agricultural Technology 10 (2025) 100812 3 study was carried out on 339 leaf samples from grapevines of different varieties: Albari˜ no, Treixadura, Loureira, Godello, Branco lexítimo, Mencía, Casta˜ nal, Sous´ on, Syrah, Garnacha and Tempranillo, during the veraison stage to each variety and season (BBCH: 83–85). Specifically, the dataset were split into a calibration set (70 %, corresponding to 234 samples) and a validation set (30 %, corresponding to 105 samples). The distribution of white and red varieties was representative in both sets. The calibration set included 91 white samples and 153 red samples, while the validation set included 40 white samples and 65 red samples. This proportional distribution reflects the overall composition of the dataset, ensuring that the division is both representative and unbiased. 20 leaves were randomly taken from each sampling point in each plot under study. Leaves were selected from fruiting shoots of medium vigor, located on the opposite side of the second cluster, on dates corresponding to the veraison stage of the vineyard. Each sample was process in first instance in the nearest research facilities, and transport them to the USC laboratory to chemical determinations. 2.2. Chemical determinations on the leaves Leaf blades and petioles were separated, washed with tap water and rinsed with distilled water, and finally oven-dried (Dry Big, J.P. Selecta, Barcelona, Spain) at 70 ◦C for 48 h or ambient drying, to constant weight. Then they were ground with a disc mill in order to pass through a 2-mm mesh and finally stored at room temperature to be analyzed. N concentration was determined by oxidizing the sample and then quantifying the gas produced during this combustion using a thermal conductivity detector (TruSpec CHNS, Leco, St. Joseph, MI, USA) [17]. For chemical analysis of the other nutrients, 1 g of sample was calcined at 500 ◦C for 8 h and subsequently wet-digested with 1 mL of deionized water and 5 mL 2 M HCl. N, C, Al, P, K, Ca, Na, Mg, Fe, Mn, Cu, Zn, and B concentrations were determined with inductively coupled plasma-optical emission spectroscopy (ICP-OES) (Optima 4300DV, PerkinElmer, Norwalk, CT, USA), for evaluate macroelements ICP methodology present a standard error between 0.5 % and 5 % of the mean value, meanwhile, for trace levels of microelements, the standard error ranges between 1 % and 10 % depending on the concentration and matrix interferences. Deionized water was used for all dilutions. Concentrations were expressed in terms of dry weight (mg/kg) and/or percentage (%). 2.3. Chemometric analysis The leaf samples were scanned in reflectance mode using a spectrometer (JASCO V-770, JASCO Corporation, Tokyo, Japan) with a coupled sphere unit (model ISN-923), and two light sources (deuterium and halogen). The spectra were collected in continuous using Spectra Manager software of JASCO. The spectral capture covered the range from 190 to 2600 nm, encompassing the ultraviolet (UV, 190–380 nm), visible (VIS, 381–780 nm), and near-infrared (NIR, 781–2600 nm) regions of the electromagnetic spectrum. The scanning speed was 1000 nm/min. 2.4. Spectral pre-treatment analysis The three spectra obtained from each foliar sample were averaged to obtain a single value and correlated with the nutrient concentrations determined by ICP-OES analysis. Different pre-treatments were applied to the spectra: (i) Standard Normal Variate (SNV), for reducing the dispersion [3], (ii) Mean center (MC) involves centering each variable by subtracting the mean of all its elements from each element in the variable, (iii) Savitzky-Golay first derivative (FD) was utilized to remove constant baseline offsets and offsets that exhibit linear variation with wavelength [1], (iv) Spectroscopic (SP) converting the reflectance units to logarithm of 1/reflectance, and (v) Detrending involves removing unwanted trends, such as baseline shifts and linear variations, to highlight relevant spectral features and enhance the accuracy of chemical or qualitative analysis [3]. After an exploratory statistical analysis, outliers were detected and eliminated of the study (4.98 % of total samples; 12 samples in calibration set, and 5 samples in validation set). Different models, using either a single pre-treatment or a combination of several, were trained and tested. Transformations were calculated using the Fig. 1. Flowchart of the methodology used. J.I. Manzano et al. Smart Agricultural Technology 10 (2025) 100812 4 software Unscrambler® XI (CAMO Software Inc., Woodbridge, Norway) [5]. 2.5. Statistical analyses For the regression analysis, we performed a PLS-regression (PLS-R) for each generated model. PLS-R is a common technique to establish a correlation between sample spectra and the variables of interest [29]. In PLS-R, one of the objectives is to determine the number of components (called latent variables in PLS) that explain the maximum variation in the data [40]. After eliminating outliers, 217 samples were included in the calibration set and 100 samples in the validation set for developing the PLSR models. Additionally, a Leave-One-Out cross-validation (CV) was performed with the calibration set in order to determine the accurate number of components that balances the predictive model, among other parameters. The model with the lowest root-mean-square error (RMSE) and the highest coefficient of determination (r 2 ) in the test of the model was selected. Additionally, we evaluate bias and slope of the PLS-R models. Bias refers to the systematic difference between predicted and actual values, where a bias near zero suggests that predictions are, on average, neither overestimated nor underestimated. Slope describes the relationship between predicted and actual values, with a slope close to 1 indicating strong agreement between them. Furthermore, the weighted regression coefficients (BW) of the PLS-R model were employed to identify the wavelengths crucial for predicting nutrient levels. This approach evaluates the relationship between each wavelength and the concentration of the element being analyzed, highlighting wavelengths with higher absolute BW coefficient values as the most influential in the model [22]. Flowchart of the methodology used is represented in Fig. 1. 3. Results and discussion 3.1. Descriptive statistics of the foliar nutrient concentrations Table 1 presents the descriptive statistical data regarding the concentrations of foliar nutrients in the leaves collected from vineyards, as determined through classical procedures. This table includes the values observed across all samples, as well as those for the calibration and validation sets separately. The minimum and maximum values of the validation set samples lie within the range of the calibration set’s extreme values. This ensures the model can be verified under appropriate conditions. The samples collected in the 2021 season from the Leiro region, corresponding to the Sous´ on, Mencía, and Treixadura varieties (data not shown), exhibited reflectance spectra with a percentage of reflectance notably different (higher or lower) compared to the mean of the remaining samples. This difference could be attributed to significant variations in the concentrations of Cu (47.93 mg/kg in the overall dataset vs. 178.50 mg/kg in these samples), Mn (146.07 mg/kg vs. 78.09 mg/kg), Zn (18.47 mg/kg vs. 48.85 mg/kg), and Na (129.78 mg/kg vs. 11.24 mg/kg). These elements, which are present in notably different concentrations in these samples, are likely responsible for the observed spectral differences. 3.2. Principal component analysis: exploratory analysis Fig. 2 shows the projection of the spectral data from the 341 leaf samples into the first two principal components, which explain 57 % and 26 %, respectively, of the total variability between samples. The distribution of the samples chosen to train the model (calibration set) and those selected for validation (validation set) was similar, as can be observed in Fig. 2. This is important when constructing a model that addresses as many observations as possible. Table 1 Basic statistics of foliar nutrient concentrations obtained through chemometric analysis. Macronutrients (C, N, P, K, Ca, and Mg) are expressed as percentages (%), while micronutrients (B, Cu, Fe, Mn, Zn, Na, and Al) are expressed in milligrams per kilogram (mg/kg). Nutrients Total sample set (n ¼339) Calibration set (n ¼234) Validation set (n ¼105) Mean Max Min SD Median Mean Max Min SD Median Mean Max Min SD Median Carbon (C) 45.42 49.92 39.14 2.04 45.87 45.25 49.28 39.14 2.08 45.99 45.79 49.92 40.73 1.96 45.83 Nitrogen (N) 1.98 3.14 0.57 0.44 1.96 1.95 3.05 0.57 0.37 1.95 2.04 3.13 0.58 0.57 1.98 Phosphorus (P) 0.17 0.45 0.01 0.06 0.16 0.17 0.45 0.01 0.06 0.16 0.17 0.34 0.02 0.05 0.17 Potassium (K) 0.97 4.24 0.22 0.57 0.85 0.89 4.24 0.22 0.55 0.80 1.17 4.11 0.53 0.57 1.02 Calcium (Ca) 1.88 4.65 0.45 0.70 1.88 1.97 4.65 0.45 0.72 1.92 1.71 3.07 0.57 0.64 1.73 Magnesium (Mg) 0.26 1.44 0.05 0.13 0.25 0.28 1.44 0.05 0.14 0.26 0.23 0.43 0.08 0.07 0.24 Boron (B) 15.15 51.60 2.08 7.69 13.65 14.98 51.6 2.08 8.28 13.08 15.45 34.62 6.52 6.43 14.24 Copper (Cu) 47.93 369.54 1.16 50.18 43.59 46.74 272.66 1.16 40.16 48.04 48.94 369.54 1.64 68.87 12.83 Iron (Fe) 47.45 311.20 1.27 30.74 41.47 46.29 311.2 1.27 31.26 40.98 50.56 192.16 2.54 30.46 42.60 Manganese (Mn) 146.07 2863.59 4.36 323.51 79.64 119.15 2508.50 4.36 285.74 72.70 215.57 2863.59 12.88 398.81 91.40 Zinc (Zn) 18.47 92.21 0.96 17.29 9.64 18.34 92.21 0.96 15.99 10.92 16.90 88.07 2.18 19.05 6.32 Sodium (Na) 129.78 367.94 3.52 125.95 89.66 131.34 357.20 3.52 125.74 101.20 118.15 367.94 5.53 126.52 72.30 Aluminum (Al) 97.27 4848.61 0.08 374.68 48.08 100.01 4848.61 0.08 436.00 45.30 92.15 4617.90 4.80 190.26 57.98 SD: standard deviation. J.I. Manzano et al. Smart Agricultural Technology 10 (2025) 100812 5 3.3. PLS-R analysis Fig. 3 shows the raw reflectance spectra of the 341 leaf samples. The different pre-treatments studied were applied to these spectra and subsequently used to generate predictive models using PLS-R. Table 2 presents the predictive results for each element using PLS-R with the optimal spectra pre-treatment, while the relationship between the predicted and reference data to validation set with an r 2 >0.60 is presented in Fig. 4. The r 2 values for the macronutrients (C, N, P, K, Ca and Mg) prediction model ranged from 0.48 to 0.84, with Ca having the highest r 2 = 0.84 in the validation set. For the micronutrients (B, Cu, Fe, Mn, Zn, Na, and Al), the r 2 values ranged from 0.45 to 0.90, with Na having the best model with a r 2 =0.90). Different pre-treatments were applied to the spectrum, as described above. There was not an ideal pre-treatment for the determination of all nutrients studied. Thus, the best pre-treatment turned out to be the use of the first derivative, either alone or in combination with MC or SNV. All PLS-R models generated are presented in Supplementary Material. In vineyard management, models with r² ≥0.7 provide predictions with sufficient reliability to guide nutrient management decisions, such as adjusting fertilization strategies or addressing deficiencies. Fig. 2. PCA of raw spectra for the first two principal components from the leaf blades samples. Blue points correspond to the samples from the calibration set, while the red points represent the samples from the validation set. Fig. 3. Reflectance spectra of all the vine leaves analyzed. The samples corresponding to red grape varieties are shown in purple, while those for white grape varieties are shown in blue. J.I. Manzano et al. Smart Agricultural Technology 10 (2025) 100812 6 In this work, the r 2 value for the determination of C was 0.70 (0.83 for calibration set). For C, previous studies [12] achieved an r 2 =0.44 in vine leaves. In a similar way, it has been described in soil nutrients of the vineyard an r 2 =0.84 for the organic matter determination [29]. N is essential for vegetative growth and chlorophyll synthesis, while P supports root development and energy transfer [21]. The PLS-R model developed to predict N concentration assessed an r 2 =0.83 and lowest root-mean-square error prediction (RMSEP) =0.55. In prior research on grapevine leaves, a model with a high value of r 2 (0.95) was determined, when combining all samples in the same database (leaves blades and petioles, and grape berries) [12]. In this way, in citrus leaves, favorable results were obtained in several studies with VIS-NIR spectroscopy showing an r 2 =0.78 in [2] and 0.91 (RMSE =1.06) after fertilization with various N doses [26]. In other crops, such as olives, it has been reported models with an r 2 =0.91 using only the short-wave infrared part of the spectrum [31]. Finally, similar results to those presented in this study concerning N determination were described in [25,38]. P was predicted with an r 2 =0.64 and an RMSE =0.02. In another study presented by [11] in vine leaves, P was determined with a ratio of performance to deviation (RPD) =1.02, but with an r 2 =0.77 using the raw spectrum of NIR in berries. On citrus leaves, similar r 2 were described ranging 0.75–0.77 with low RMSE [2,26]. K participates in enzyme activation and improves fruit quality and resistance to several diseases. The model generated for K predicted this element with an r 2 for the validation set =0.75 and an RMSE =0.39. Using just the NIR range of the spectrum, an r 2 =0.42 was reported in [11] to leaf blades, and 0.76 and 0.79 in berries and petioles, respectively. In their study, Oliveira and Santana [25] obtained an r² =0.76 and an RMSE of 1.30 while analyzing eucalyptus leaves. The PLS-R model developed to predict Ca showed an r 2 =0.84 and a RMSE of 0.35. Comparable results have been found for leaf blades with and r 2 =0.88, but with higher value of RMSE and lower set size (n = 159) [11]. In this way, different models obtained with sample leaves from different crops such as citrus trees, or eucalyptus, predicted Ca concentration with r 2 ranging from 0.62 to 0.81 [2,25,26]. Regarding micronutrients (B, Cu, Fe, Mg, Mn, and Zn), Mg wasdetermined as well as Fe with an r 2 =0.48 and 0.26, respectively. Other researchers were able to predict the magnesium content of leaf blades with better accuracy in vine (r² =0.60) [11], and in sugarcane, achieving an r² =0.97 and an RMSEP =0.005, using multiple linear regression and PLS-R in the spectral range of 780–2500 nm [41]. Concerning Fe, the predictive models are poor in general in the literature, being unable to predict successfully in the most of the crops analysed. Only in grapevine leaf blades was observed a moderate correlation with an r 2 =0.58 in these organs, which could be used for classification (excess, deficiency,…) [11]. The prediction of B was determined with an r 2 =0.45 and an RMSEP =4.60, surprisingly the r 2 in the calibration set was 0.62. This model performed was better in vine leaves samples than another reported by [11] with a RPD =1.18. The best model for B was presented in eucalyptus with a high level of estimation with an r 2 =0.83 and an RMSEP =8.59 [25]. Moderate prediction model was generated for Cu determination with an r 2 =0.58 and an RMSEP =56.20. Similar results were obtaining by [11] using NIR spectroscopy in grapevine leaves reaching an r 2 =0.61 in leaf blades. The Mn concentration was predicted with an r 2 =0.58 and an RMSEP =56.20. Similar outcomes were achieved by [11]. However, this element was predicted in citrus leaves (r 2 = 0.69, RMSEP =40.75) [2]. The PLS-R model for Zn determination predicted this element with an r 2 =0.70 in the calibration set and an r 2 =0.50 for the validation set with an RMSE =19.46 in the leaf samples. In a previous study, Zn was determined (r 2 =0.82) in grape berries but worse in leaf blades (r 2 =0.29) [11]. Concerning Na, there is not numerous studies of prediction by spectroscopy of this element. Furthermore, it has been described an r 2 =0.83 in the prediction of Na in legume plants [10]. The best PLS-R model presented in this study is precisely the one related to the prediction of sodium (r 2 =0.90; RMSE = 50.74). Finally, the model created for Al predicted this element quite well, achieving a high r² value =0.69. Another study in white pine and red oak using VIS-NIR reflectance predicted this element with an r 2 = 0.82 [15]. The results presented demonstrate good predictive performance for the elements analyzed, with nitrogen (N) and manganese (Mn) showing relatively strong results. However, Fig. 3 reveals some variation in the predictions, especially at lower concentrations, where the points show less continuity along the line. This behavior is also observed for potassium (K) and phosphorus (P), which, despite high coefficients of determination, exhibit some prediction variability, particularly at lower values. These findings suggest that while our models are robust for certain concentration ranges, there is room for improvement, especially at the extremes. Future work will focus on refining these models, exploring alternative modelling techniques, and adjusting the spectral range and pre-treatments to enhance prediction accuracy across all concentration levels. 3.4. Analysis of significant wavelengths in the construction of predictive models Regression coefficient along the spectra, associated with the most accurate predictive models for all the nutrients studied are represented in the Fig. 5. In this way, the most relevant wavelengths are indicated in Table 3. In leaves, the photosynthetic pigments (chlorophylls and carotenoids) absorb about 90 % or more of the light situated in the visible region (351–700 nm). Conversely, the NIR lacks molecules with strong absorption, leading to plants refracting or transmitting almost all radiation in this range, retaining only about 10 % [32]. These essential photosynthetic pigments in leaves play a critical role in plant photosynthesis, Table 2 Results for calibration, cross-validation, and validation set using PLS-R of the best spectroscopic models. Nutrient Symbol Pre-Treatment Components Calibration Leave-One-Out Cross-Validation Validation Set r 2 RMSE r 2 RMSE r 2 RMSE Bias Slope C MC+FD 7 0.83 0.85 0.66 1.21 0.70 1.54 −0.62 0.68 N SNV 7 0.72 0.30 0.55 0.38 0.83 0.55 −0.16 0.75 P Raw 7 0.47 0.04 0.38 0.04 0.64 0.02 0.02 0.81 K FD 7 0.68 0.31 0.56 0.47 0.75 0.39 0.03 0.60 Ca Raw 7 0.47 0.52 0.41 0.55 0.84 0.35 0.07 0.78 Mg SNV 7 0.35 0.08 0.26 0.08 0.54 0.07 0.04 0.74 B MC+FD 6 0.62 4.99 0.43 6.19 0.45 4.60 0.71 0.45 Cu Raw 6 0.52 27.66 0.43 30.18 0.58 56.20 0.86 0.28 Fe FD 2 0.26 10.08 0.14 11.59 0.41 12.89 −3.36 0.25 Mn FD+SNV 7 0.80 126.86 0.54 229.00 0.82 229.79 0.38 0.69 Zn FD 5 0.70 9.02 0.54 11.16 0.50 19.46 0.90 0.39 Na FD+SNV 7 0.82 53.74 0.51 88.46 0.90 50.74 0.30 0.77 Al SNV 7 0.54 14.72 0.47 15.85 0.69 18.29 0.20 0.50 RMSE: Root mean square error; MC: Mean center; FD: first derivative; SNV: standard normal variate. J.I. Manzano et al. Smart Agricultural Technology 10 (2025) 100812 7 Fig. 4. Relationship between predicted and reference values of the best PLS-R models (r 2 >0.6) in the validation set. J.I. Manzano et al. Smart Agricultural Technology 10 (2025) 100812 8 converting sunlight and CO 2 into sugars. Chlorophylls are divided into chlorophyll a (Chl-a) and chlorophyll b (Chl-b), responsible for the characteristic green color of leaves, with absorption peaks around 450 and 680 nm. Carotenoids are divided into carotene a, carotene b and xanthophylls, and exhibit strong light absorption in the blue region of the spectrum (450–500 nm). Almost all the elements determined in this study are either directly or indirectly related to chlorophyll and the photosynthetic process. In this way, the visible part of the spectrum is relevant regarding the contribution to the generation of the proposed models (190 – 2600 nm) with 61 reflectance peaks providing information to the different models, especially for micronutrients models with (32 main peaks in VIS region) (Fig. 4). In the case of B, the visible spectrum is the only part taken into account for the creation of the PLS-R model, while the rest of elements use at least two-range area (UV, VIS and NIR). On the contrary, within the NIR region of the characteristic plant reflectance spectrum, compounds such as proteins, fatty acids, and starch exhibit prominent absorption features between 800 and 1000 nm [34]. Meanwhile, cellulose and lignin display distinctive absorption peaks around 1250–1800 nm, while nitrate absorption becomes notable at 1600 nm, and sugar content is discernible at 2300 nm [20]. The contribution of NIR range of the spectrum was relevant contributing with 52 selected peaks in the determination of the elements studied, and being most important for the determination of macronutrients with 29 peaks, predictably due to the presence in higher concentrations in plants. In this way, the relationship with the organic compounds is higher than micronutrients, enhancing signal-to-noise ratio, and facilitating the interpretation by PLS-R models. The elements more influenced by the NIR region of the spectrum were P (7 peaks), Mg (6 peaks), Cu (6 peaks), and Ca (5 peaks). The most complex PLS-R model was created for C element with 14 relevant wavelengths (Table 3). This is consistent because it is the element in the highest concentration and the one that generates the greatest interaction with organic matter, and it is Fig. 5. Weight regression coefficients from the PLS-R models for all the nutrient. J.I. Manzano et al. Smart Agricultural Technology 10 (2025) 100812 9 also an element directly involved in the vibration of C – H bonds. Finally, the contribution of the UV range in the determination of nutrients ends up being marginal, with peaks in the prediction of P (2), K (1), Mg (1), and Mn (1). Thus, Mn is the unique micronutrient with a relevant wavelength in the UV region (371 nm). 4. Conclusions In the realm of nutrient analysis, there’s a growing demand for an efficient and economically viable technique that minimizes sample handling, reduces reagents, and allows real-time assessment on production lines. This study highlights the promising potential of UV–VIS-NIR spectroscopy combined with PLS-R models in predicting macroand micronutrient concentrations in grapevine leaves with considerable accuracy. The models achieved high coefficients of determination for key elements such as calcium (Ca) and sodium (Na), showcasing the robustness of these techniques to support fertilization and nutrient management decisions in vineyards. However, the variability in model performance across different nutrients also reflects the complexity of nutrient interactions in plants. In this way, ionomics, the comprehensive study of the elemental composition of living organisms, offers promising advancements in the field of plant science and has the potential to revolutionize agricultural practices. The ability to predict a wide range of nutrients with r² values ≥0.7 for several key elements ensures these models can serve as reliable tools for vineyard nutrient monitoring and decision-making. Beyond practical utility, the approach exemplifies the potential for spectroscopy to contribute to the broader goals of precision agriculture, including optimized resource use and enhanced crop quality. On the other hand, a key insight from this study is the differential contribution of spectral regions to predictive models. The visible (VIS) region was key for predicting micronutrients like boron (B), linked to photosynthetic pigments. In contrast, the near-infrared (NIR) region was more influential for macronutrients like phosphorus (P) and magnesium (Mg), due to their association with organic compounds and plant structures. The ultraviolet (UV) range had a minor role, emphasizing the importance of VIS-NIR regions in spectroscopic analyses. The findings achieved reinforces the use of UV–VIS-NIR spectroscopy for nutritional diagnosis in vineyards. This method significantly reduces the costs associated with traditional chemical analyses. Additionally, spectroscopy technologies facilitates real-time decisions, allowing vineyard managers to address nutritional deficiencies promptly, optimize fertilizer use, and ultimately improve crop yield and quality. Equally important is the contribution to international databases of the generated models, enriching the scientific community’s knowledge base and encouraging interdisciplinary research. Future works should be focus in the translating this technology to the vineyard in order to improve the nutrient use efficiency and quality of crop plants. Additionally, exploring new non-linear models, such as artificial neural networks, support vector machines, or artificial intelligence, should be considered as alternatives to enhance the obtained results, especially for those elements with worse results employing linear models (Mg, Fe, and B). In response to suggestions regarding practical application, developing a portable device along with a standardized protocol for vineyard growers will also be a key aspect of future work. Finally, future studies will focus on exploring narrower spectral ranges to improve the performance and accuracy of the models for specific nutrients. Funding IRRIVITIS-PID2019–105039RR-C44 (Funding by MCIN / AEI /10.13039/501,100,011,033, Spain) Ethics statement Not applicable: This manuscript does not include human or animal research. If this manuscript involves research on animals or humans, it is imperative to disclose all approval details. If Yes, please provide your text here: Supporting information Comparative statistics and uncertainty indices (r 2 and RMSEP) on reflectance for the validation sets of the model generated by PLS-R for all the nutrients and figure showing the location of study plots in the Iberian Peninsula. CRediT authorship contribution statement J.I. Manzano: Writing – review & editing, Writing – original draft, Visualization, Methodology, Investigation, Formal analysis, Investigation, Conceptualization. M. Rodríguez-Febereiro: Writing – review & editing, Visualization, Methodology, Formal analysis, Conceptualization. M. Fandi˜ no: Writing – review & editing, Visualization, Methodology, Formal analysis, Conceptualization. M. Vilanova: Writing – review & editing, Visualization, Supervision, Investigation, Conceptualization. J.J. Cancela: Writing – review & editing, Visualization, Supervision, Investigation, Funding acquisition, Conceptualization. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments To researchers which collaborate with the collection of leaf samples in the projects: IRRIVITIS-PID2019–105039RR-C4 (Funding by MCIN / AEI /10.13039/501100011033, Spain), AROMAVID (Funding by MCIU / CDTI, Spain), ALBASOUL (Funding by MCIU / CDTI, Spain) and Table 3 Analysis of relevant wavelengths for PLS-R models. Nutrient code PLS-R selected wavelenghts (nm) UV VIS NIR C–450, 478, 512, 544, 601, 629, 651, 702, 733, 769 1389, 1440, 1891, 2032 N–412, 461, 581, 595, 639, 669, 703, 741 1153, 1222, 1290, 2006, 2079 P 267, 303 420, 454, 478, 603 794, 1173, 1401, 1582, 1932, 2108, 2487 K 329 413, 450, 601, 708 1399, 1901 Ca –554, 737 1095, 1273, 1582, 1926, 2104 Mg 302 692 899, 1267, 1423, 1948, 2118, 2402 B–514, 545, 601, 653, 707 – Cu –491, 715 1113, 1292, 1581, 1826, 1932, 2110 Fe –446, 583, 600, 632, 651, 705, 775 1422 Mn 371 411, 534, 599, 649, 683, 745, 770 851, 1145, 1388, 1430, 1875, 2024 Zn –437, 583, 686 793, 1144, 1388, 1900, 2028 Na –514, 545, 600, 699, 776 1418, 1900 Al –545, 603, 694 781, 1388, 1900 The wavelengths of greatest relevance for the generation of the predictive model are highlighted in bold. J.I. Manzano et al.