Vis–NIR spectroscopy: from leaf dry mass production estimate to the prediction of macro- and micronutrients in soybean crops

Abstract. Our work aimed to evaluate the use of visible–near-infrared (Vis–NIR) spectroscopy for predicting the production of leaf dry mass (LDM), as well as macro- and micronutrients contents of soybean leaves grown after application of limestone-mining coproducts. The treatments were arranged within a triple factorial scheme (6  ×  2  ×  2  +  2) and placed into pots in a greenhouse. We evaluated the following factors: type of input (limestone-mining coproducts), input particle size (filler and powder), and soil class (Arenosol and Ferralsol). After inputs incubation, the soybean was sown. Then, 42 days after sowing, we collected the foliar spectra, as well as leaves, for further analysis of the contents of macro- and micronutrients in leaves and production of LDM. We managed to adjust models at the stage of prediction with R2p  >  0.50 and RPDp  >  1.50 for the variables LDM, P, K, Mg, S, and Zn, with emphasis on the first four, which presented R2p above 0.65. Therefore, we conclude that Vis–NIR spectroscopy has a potential for predicting LDM and the nutrients contents of soybean subjected to the application of limestone-mining coproducts, with advantages such as speed, low cost, and no use of reagents that are toxic to the environment.

hyperspectral sensing using visible-near-infrared (Vis-NIR) sensor could be helpful in determining the leaf nutrients contents in soybean. The Vis-NIR spectroscopy is commonly used as a remote, non-destructive method for rapid analysis of many attributes of fresh leaves. Vis is the region where photosynthetic pigments, such as carotenoids, chlorophylls, and xanthophyll, absorb strongly. However, the reflection in NIR is dominated by structural reflection of turgid plant cells of the mesophyll. 6 The use of Vis-NIR spectroscopy can assist in the evaluation of applications of alternative inputs since it is a clean and non-destructive way of evaluating the structural and morphological characteristics of leaves, among other parts of plants. Furthermore, due to the ease of data acquisition and processing, it allows application in areas with precision agriculture, having seen the large number of samples demanded in this type of agriculture.
Several studies have demonstrated the potential of this tool for predicting attributes such as water contents in the crops, 7 productivity, 8 pigment levels, 9,10 vegetation index, 11 and leaf nutrients contents, 12 such as N, 5,13 P, 11,13 K, 11,13,14 Ca, 13,15 and Mg. 13,15 However, studies on the use of this technique for the prediction of nutrients in soybean (especially micronutrients) are scarce. This is due to the complexity of nutritional functions and the number of plant chemical compounds, which makes the estimation of leaf nutrients by means of reflectance Vis-NIR a procedure with difficulties in autocorrelation and collinearity. 9,16 The difficulty is even greater in the case of micronutrients, which although they have an essential function in the development of plants, they are present in minor quantities.
In case of nutritional deficiency, the production of leaf dry mass (LDM) can be impaired due to loss of structural, enzymatic, and photosynthetic compounds that are essential for forming a vigorous plant. That way, variability in the production can occur at any time due to nutritional factors. Therefore, it becomes fundamental to understand the plant's response to a certain factor and correlate that answer to the spectral response obtained from its leaves. Understanding the patterns of leaves spectral behavior allows creating models of plant productivity estimate. 17 Based on the potential aforementioned, we aim to solve two basic issues of current soybean production: (I) reducing dependency on the use of soluble fertilizers (through the application of limestone-mining coproducts) and (II) estimating LDM and leaf nutrients contents. That being said, the aim of this study is evaluating the potential of Vis-NIR spectroscopy for predicting macro-and micronutrients contents, as well as in the production of LDM of soybean subjected to the application of limestone-mining coproducts.

Chemical and Mineralogical Characterization of the Inputs
The materials used in this study, which are here called inputs or coproducts, were collected from a depth of 0 to 15 m in a limestone-mining area in Tietê, São Paulo, Brazil. These inputs, which are located at different depths, were named ritmite (from Irati formation), bituminous shale (BS) (from Irati formation), and Corumbataí formation (CF). The contents of CaO, MgO, K 2 O, and S of the inputs used in this study are shown in Table 1. Mineralogy of the inputs was determined through x-ray diffraction with Shimadzu XRD-6000 equipment, in a fraction smaller than 0.074 mm, using CuKα radiation in staircase sweep, at an interval of 2θ of 5°to 70°, with angular step of 0.02°in Bragg configuration. 18 Afterward, the results were interpreted by comparison with the reference standards contained in the powder diffraction file. 19

Chemical and Particle-Size Characterization of the Soils
To carry out the experiment, we used soil samples from the surface layer (0 to 20 cm) of a Ferralsol and an Arenosol. 20 They were collected from a native forest area in Mandaguari and Paranavaí, PR, Brazil, respectively. After collection, the samples were ground, air-dried, and sieved with a 1-cm sieve. For chemical and clay content characterization, the samples were dried in an oven with forced-air circulation at 50°C until we obtained the constant mass. Then, the samples were ground and sieved with a 2-mm sieve. Determination of P and K þ (Mehlich-1), Ca 2þ , Mg 2þ , SO 4 2− , pH (water 2:1), and C organic followed the methodologies proposed by Teixeira et al., 21 whereas Si (soluble) followed the methodology described by Korndörfer. 22 The contents of H + Al were evaluated by the pH exchange of the sample against the SMP buffer. 23 After determining the elements, we calculated the sum of bases (SB ¼ Ca 2þ þ Mg 2þ þ K þ ) and the potential capacity of cations exchange (CEC ¼ SB þ H þ þ Al 3þ ), as described in Table 2.

Delineation and Experimental Procedure
The study was conducted in a greenhouse at the State University of Maringá. The treatments were arranged within triple factorial scheme (6 × 2 × 2 þ 2), randomly designed with five repetitions, and placed into pots (n ¼ 130) in a greenhouse. The factors taken into account were type of input, input particle size, and soil class. The types of input used were coproduct 1 (50% CF + 50% ritmite), coproduct 2 (100% CF), coproduct 3 (100% ritmite), coproduct 4 (100% BS), coproduct 5 (30% CF + 15% BS + 55% ritmite), and coproduct 6 (55% CF + 15% BS + 30% ritmite). The two additional controls consisted of treatments in which there was no input application, one for each soil class evaluated. The doses of coproducts 1, 3, and 5 were calculated in accordance with the Soil Chemistry and Fertility Commission of Rio Grande do Sul and Santa Catarina (SCFC-RS/SC), 24 according to the contents of CaO of the inputs and the Ca 2þ of both soil classes since they are mainly characterized as sources of such nutrient. As for coproducts 2, 4, and 6, the doses were calculated in accordance with SCFC-RS/SC 24 based on the contents of K 2 O of these inputs and the K þ of the two soil classes, for they are a material with potential to be used as sources of that nutrient for the soil. The input particle sizes evaluated were filler (100% smaller than 0.30 mm) and powder (100% < 2.00 mm, 70% < 0.84 mm, 50% < 0.30 mm). That gives them a theoretical reactivity of 100% and 68%, respectively. The soil classes used were Arenosol and Ferralsol.
The treatments with the applied amounts of K 2 O, CaO, MgO, and S are described in Table 3. The coproducts doses were applied into pots containing 3.0 kg of dry soil (Ferralsol and Arenosol). Humidity control aimed at keeping the values close to the field capacity of both soil classes used. The pots were placed inside a greenhouse for incubation during 140 days. After incubation, we did the soil sampling with an auger for determining the macronutrients K þ and P (Mehlich-1). Afterward, fertilization recommendations were followed in order to increase the contents of P and K þ of the soil to the critical content of soybean (Glycine max L.) in accordance with what is recommended by SCFC-RS/SC. 24 For correcting P and K þ , we used triple superphosphate and coproduct 2 (CF), respectively, both applied during sowing. We also inoculated soybean with Bradyrhizobium spp. for nitrogen fixation. Right after application of the inputs that are sources of P and K, and after inoculation, we sowed 10 soybean seeds (NIDERA NA5909) per vase, and when the plant reached the V1 stage, we did the thinning, which left 4 plants per pot. A few moments before collecting the soybean leaves, we acquired the radiometric data (reflectance) from the adaxial part of the leaves (n ¼ 130) with the aid of an ASD leaf clip (next item). Leaves were collected 42 days after sowing, at the V6 stage. After that, the leaves were dried in an oven with forced-air circulation at 60°C, until constant mass. Then, we evaluated the production of LDM, as well as the contents of N, P, K, Ca, Mg, S, Cu, Fe, Mn, and Zn. For N, at first, the tissue samples were subjected to sulfuric digestion and, after that, the N was determined through distillation in a semi-micro Kjeldahl device and titration with sulphuric acid (0.025 mol l −1 ). As for S, first came nitricperchloric digestion, followed by extract precipitation with BaCl 2 and determination through a UV-Vis spectrophotometer in transmittance form, in accordance with Da Silva. 25 The other nutrients were determined using an inductively coupled plasma atomic emission spectrometer.

Radiometric Data (Reflectance) of the Soybean
The radiometric scanning of the soybean leaves was done with an ASD FieldSpec 3 spectroradiometer operating at 350 to 2500 nm, with spectral resolution of 3 nm at 350 to 1400 nm and 30 nm at 1400 to 2500 nm [short-wave infrared (SWIR)]. Using the leaf clip allowed for collecting the leaf reflectance spectra of living vegetation directly in the greenhouse (Fig. 1). For spectra data collection, initially the spectroradiometer was optimized to adjust the sensitivity of the instrument's detectors according to the specific illumination conditions at the time of measurement with 20 scans per white reference (Spectralon standard). Then, to achieve the relative reflectance measurements, the white reference was collected before scanning the leaves samples until a nice, clean, 100% reference line was obtained according to the specifications provided by the Reflection and Calibration Laboratory. 26 For each plant individual, three leaves were measured in reflectance mode. The mean was then calculated according to recommendations of Nanni and Demattê. 27 Fig. 1 Scheme of conducting the experiment and collecting spectral data on soybean.

Data Processing and Statistical Analysis
The results of both LDM and contents of macro-and micronutrients of the leaves were initially subjected to descriptive statistical analysis with the Excel software-XLSTAT. 28 We used The Unscrambler 10.4 software to smoothing the reflectance (soybean leaves) data through the multiplicative signal correction (MSC) method in order to reduce the additive scattering effect. 29 The rationale behind MSC is to compensate for these additive effects. In addition, the preprocessing technique MSC was selected because it has been shown to improve the performance of spectroscopy models. The results were centered on the mean and subjected to partial least squares regression (PLSR) 30 with the iterative algorithm NIPALS 31 in order to estimate of LDM and macro-and micronutrients contents of soybean leaves.
To generate the prediction models from the reflectance spectrum (Vis-NIR), we first removed the wavelengths of marginal (350 to 379 nm), considered to be of greater noise 32 and only the wavelengths in the range between 380 and 1000 nm (Vis-NIR) were used since this range is often employed to probe the properties of living plant leaves. 6 Subsequently, spectral scanning's of the 130 samples were divided into two groups. The first represented 91 samples used for generating the model (calibration), and the second represented 39 samples used for validating (prediction) the adjusted PLSR model. These samples were randomly selected from the data set, thus avoiding bias that could influence the model quality. The leave-one-out cross-validation (cv) method was used as a preliminary form of predicting the soybean attributes. 33 In the prediction (p) process, parameters such as determination coefficient (R 2 ), root-mean-square error (RMSE), ratio of performance to deviation (RPD), and bias were used to evaluate quality and accuracy of the model (Fig. 2). Rossel et al. 34 and Saeys et al. 35 state that the values of R 2 are classified as: R 2 < 0.50 (models with poor predictions only capable of distinguishing high and low values); R 2 between 0.50 and 0.65 (models with moderate predictions that may be used for evaluation and correlation); R 2 between 0.65 and 0.80 (good prediction models that enable quantitative predictions; R 2 between 0.80 and 0.90 (models of very good quantitative predictions); and, finally, R 2 ≥ 0.90 (excellent prediction models). Chang et al. 36 and Dunn et al. 37 have suggested that models with RPD values lower than 1.5 should be considered insufficient for the majority applications, whereas models with RDP values between 1.5 and 2.0 are considered useful in relation to the accuracy of their predictions, and models with values greater than 2.0 should be considered excellent. Figure 3 shows the peaks of x-ray diffractometry of the coproducts used in the study, in compliance with Rodrigues et al. 30 For ritmite [ Fig. 3(a)], we verified peaks related to quartz, smectite, calcite, dolomite, and pyrite. As for CF [ Fig. 3(b)], the peaks identified referred to smectite, illite, quartz, feldspar, and mica. Regarding BS [ Fig. 3(c)], the peaks occurred due to quartz, albite, mica, and pyrite. These results, in most of the cases, corroborate the available literature on their mineralogy. 38-40 Table 4 shows the results of the descriptive analysis of LDM and the nutrients contents found in the soybean leaves after cultivation in soil incubated with mining coproducts. The coefficient of variation values (CV) ranged from 9.97 % for Ca and 91.02 % for Mn.

Estimation of LDM and Nutrients Contents of Soybean Using Vis-NIR Spectroscopy
It can be noted in Fig. 4 that there was a separation between treatments evaluated when compared to their reflectance, which allows to infer about the possibility to use the spectral curves in the differentiation of treatments, as well as in the prediction of plants attributes (LDM and nutrients content). In addition, absorption valleys close to 497 and 690 nm can be noted due to chlorophyll 41 and in the range of 700 to 770 nm, the reflectance sharply increases due to the fact that there is a cavity with a large reflecting surface in the sponge tissue structure in the mesophyll. 42 Table 5 shows the statistical parameters of the PLSR model for the phases of calibration and cross-validation of the LDM and the nutrients contents of soybean that were evaluated (n ¼ 91).

Calibration (c) and cross-validation (cv)
The values of the parameters evaluated (R 2 , RMSE, RPD, and bias) fluctuated depending on the variable studied.
We can see that, at the calibration phase (c) for nutrients N, Ca, and Cu, the models generated were considered poor (R 2 c < 0.32). For Fe and Mn variables, the models constructed were considered moderate (R 2 c 0.55 to 0.60). As for nutrients Mg, S, and Zn, the models were considered good (R 2 c 0.75 to 0.79). Finally, for LDM and nutrients P and K, the models were considered very good (R 2 c 0.85). Regarding the RMSE c , the values ranged from 0.1123 g plant −1 for LDM to 61.3624 mg kg −1 for Mn. As for bias, the values tended to be zero for all the variables.
At the cross-validation (cv) phase, we can see in Table 5    variables, the models were considered poor (R 2 cv < 0.45). For S, Fe, and Zn variables, the models with the R 2 cv (0.55 to 0.60) were considered moderate. For Mg, the model was considered good (R 2 cv 0.70), and for LDM and the nutrients P and K, the models generated were considered very good (R 2 cv 0.80 to 0.84). With regard to the values of RMSE cv , the values ranged from 0.1418 g plant −1 for LDM to 74.8524 mg kg −1 for Mn. As for bias cv , similarly to the calibration phase, the values found were close to zero.
As for the RPD values in the calibration phase, it is noted that the models presented results considered insufficient (<1.32) for the variables N and Ca and useful for Fe (1.56). As for the variables LDM, P, K, Mg, S, Cu, Mn, and Zn, the models found, according to the RPD, were considered excellent (>2.00). In the cross-validation phase, the models generated, considering their respective RPD, were considered insufficient for N, Ca, and Fe (<1.42), useful for Mg and S (1.85 and 1.50, respectively), and excellent for the other variables (>2.03).
The models that were adjusted to predict the nutrient contents of the plant, as well as the best numbers of PLSR factors, were previously tested by cross-validation. Thus, the relationship between the predictor variable (reflectance) and the predicted one (LDM and nutrient contents) was better explained using the models containing 12 factors for S; 11 factors for P and Mg; 10 factors for LDM and K; 9 factors for Zn; 7 factors for Fe; 4 factors for Mn; and 2 factor for N. As for Ca and Cu, the reliability of the model was limited since it presented only 1 factor.

Prediction (p)
To evaluate the capability of the PLSR models adjusted to predict LDM, as well as the nutrients contents in the leaf, these models were tested with a set of spectral data, different from those used for calibration and cross-validation. The scatter plot together with the results of the multivariate statistics is present in Fig. 5 (n ¼ 39).
The R 2 p values found decreased in relation to the previous phases (c) and (cv). For variables N, Ca, Cu, Fe, and Mn, the models generated according to R 2 p were considered poor (<0.24). For variables S and Zn, the models were considered moderate due to the R 2 p found (0.55 and 0.58, respectively). For LDM, P, and Mg, the model was considered good (R 2 p 0.65 -0.70), and for K, the model generated was considered very good (R 2 p 0.80). Regarding the RMSE p , the values obtained ranged from 0.1584 g plant −1 (LDM) to 80.5424 mg kg −1 (Mn). For bias, with the exception of nutrients Fe and Mn, the values found were close to zero. As for the RPD values, it is noted that the models presented results considered insufficient (<1.42) for the variables N, Ca, and Fe. As for the variable LDM and the other nutrients evaluated, the models found, according to the RPD, varied between useful (1.5 to 2.0) and excellent (>2.0).
The following are the correlation graphs between laboratory results (reference) and those obtained via Vis-NIR spectroscopy (predicted) (Fig. 5). As already mentioned, the variables LDM, K, Mg, Zn, S, and P were the ones with less dispersion and, consequently, better prediction capacity. On the other hand, the nutrients N, Ca, Cu, Fe, and Mn were the most dispersed and difficult to estimate. Figure 6 shows the regression coefficients of the PLSR model. We can observe regions of valleys and peaks where the RCs had a higher weight for the model construction. In general, the important wavelengths for the variables were close to 380 to 400 nm, 500 to 530 nm, 600 to 690 nm, and 700 to 750 nm.

Interpretation of Regression Coefficients (RC)
For the LDM variable, we can see peaks and valleys, resulting from higher RC, in the regions near 400, 500, and 530 nm. Regarding Cu, Fe, and Mn [Figs. 6(h)-6(j)], despite several wavelengths with higher weights for the model construction, due to the low R 2 values found for these nutrients (Fig. 5), the RC found proved to be of little use. Finally, for Zn [ Fig. 6(k)], higher RC was observed in regions near 400 and 620 nm. Table 6 describes the summary of the most responsive wavelengths according to the coefficient of the PLSR regression (Fig. 6) obtained for the estimation of the variables LDM production and nutrients contents of soybean.

Descriptive Analysis of the LDM and Nutrients Contents of Soybean
The results obtained for the nutrients contents (plant) N, P, K, S, and Mn according to SCFC-RS-SC 24 were considered insufficient for the development of the crop. For Cu, Fe, and Zn, the contents varied between insufficient and sufficient for the crop. Moreover, according to the Pimentel-Gomes 43 classification, the coefficients of variation for the evaluated variables were classified as low for Ca, medium for LDM and N, high for S, and very high for P, K, Mg, Cu, Fe, Mn, and Zn. These results occurred because of the sources of nutrients used, such as mining coproducts. Due to the diversified mineralogy (Fig. 3), these coproducts provided varied input of elements, which contributed to the high range of minimum and maximum values of the nutrient contents. Furthermore, the use of inputs with different input particle size and solubility rates also favoured the heterogeneity in the coefficient of variation values found.

Estimation of LDM and Nutrients Contents of Soybean Using Vis-NIR Spectroscopy
Although the plant spectrum does not present characteristic inflections of the nutrients present in the plants, the peaks and dips caused by chlorophyll (Fig. 4) and the high reflectance of the NIR region typical of plants with the well-structured mesophyll 44 are signs of satisfactory nutritional conditions. 45,46 This allowed the construction of models, such as PLSR, which have the potential to make indirect predictions of nutrients. In addition, the production of LDM can also be directly related to the chlorophyll and the leaves' mesophyll, 6 which allows its estimation, as discussed below.

Calibration (c) and cross-validation (cv)
In the calibration phase of the models for K and P elements, which showed very good R 2 c (0.85), the results are corroborated by other studies using the Vis-NIR spectroscopy that also obtained high R 2 c values for these elements in barley 47 and cucumber 48 plants. Although K and P may not be directly detected when irradiated by Vis-NIR frequencies, organic compounds and other complex plant components containing these elements can be detected and evaluated when inspecting leaf spectra. 49 Thus, the results obtained in this study demonstrate the potential of using Vis-NIR spectroscopy as a tool to help determine the contents of these nutrients in plants.
Regarding the cross-validation phase of the PLSR data, for variables N, Ca, Cu, and Mn, the results obtained in this study were below those achieved by Pandey et al. 6 that when studying the use of Vis-NIR spectroscopy in the prediction of nutrients in maize and soybean leaves, they found R 2 cv for N, Ca, Cu, and Mn, respectively, of 0.88, 0.75, 0.80, and 0.51. For N, this result was not expected since it is one of the elements associated with chlorophyll, 11 which therefore interacts with electromagnetic radiation. 5 For variables Mg, S, Fe, and Zn (cv), the R 2 cv were considered similar to those found by Pandey et al., 6 who obtained values of 0.69, 0.71, 0.73, and 0.64, respectively, for these nutrients. In relation to LDM, the results were similar to those of Silva-Perez et al., 50 who found R 2 cv of 0.89 for this variable in wheat culture. For P and K, the results of R 2 cv were slightly higher than the ones of Pimstein et al., 51 who studied the use of Vis-NIR spectroscopy for predicting nutrients in maize culture and obtained R 2 of P ¼ 0.46 and R 2 of K ¼ 0.77. For S, the results of R 2 cv were lower than those obtained by Pandey et al., 6 who found R 2 cv of 0.71 for that nutrient. In relation to Mg, these results were expected since this nutrient is part of the structure of the chlorophyll molecule, which presents strong interaction with electromagnetic radiation, what allows the creation of models with high R 2 . 52 Other nutrients such as K, Ca, Fe, Mn, Zn, and Cu, all metallic elements that exist mainly as ions in living plant tissues, despite not presenting active spectral absorption characteristics in the Vis-NIR region, have the ability to bind electrostatically or act as binders for larger compounds containing carbon, what allowed relatively good PLSR models for some of them. 6 The RMSE c ; cv values ranged according to the analyzed variable (Table 5). For the nutrients N, K, Ca, S, Cu, Fe, Mn, and Zn, the RMSE cv values found in this study were lower than those of Pandey et al., 6 Another important parameter used to evaluate the calibration models is bias. For that, the values were insignificant, indicating that there was a random point distribution around the regression line, resulting in the absence of bias. 33 As an example, for the variable P, the bias cv of 0.2106 in the cross-validation phase indicates that the measurements for this variable are slightly lower than the reference values. 53

Prediction (p)
In general, in the prediction phase (Fig. 5), the R 2 p values for the evaluated variables were lower than those found in the previous phases (calibration and cross-validation), demonstrating the limitation of the models to capture all data variability. 33 However, the value of the bias p , with the exception of Fe and Mn elements, was low. That shows that there was no bias adjustment in the validation of the prediction models. The attributes that presented the best correlations between the laboratory routine analysis and the PLSR technique were, respectively, K, Mg, P, and LDM since the R 2 p values found were between 0.65 and 0.80 and RPD p above 2.03, characterizing a modelling with good (R 2 p ) and excellent (RPD p ) adjustment. For the nutrients, the results were similar to those of Pandey et al., 6 who found for K, Mg, and P, R 2 p between 0.69 and 0.83, and, in relation to RPD p , values above 1.79. As for nutrients N, Ca, S, Cu, Fe, Mn, and Zn, the same author found higher R 2 p and RPD p values similar to those found in this study. As for S, the results in this study were similar to those found by Cozzolino and Moron 54 who obtained R 2 p of 0.65 working with leguminous crops.
Studies on P and K in other crops showed similar R 2 p when using the Vis-NIR range 32 and lower when using visible, near-infrared, and short-wave infrared (Vis-NIR-SWIR) 11 than the ones found in this study. Some elements (e.g., P) have no absorption in the NIR region, but they can be indirectly detected through their binding with organic complexes, chelates, and pigments, such as chlorophyll. 54 Also, it has been found that potassium is associated with organic acids, especially malate. 55 This makes its identification possible mainly in the wavelengths of the NIR, which have characteristics of electromagnetic radiation absorption associated with organic acids. 56 For the LDM variable, the results of R 2 p were similar to those found of Garriga et al., 57 who found R 2 p of 0.67 for forage production estimate and slightly lower than those of Tzanakakis et al., 58 who found R 2 p of 0.89 for the prediction of biomass production from forest species. According to Pimstein et al., 51 the nutrients present inside plant tissues directly affect the development of biomass and, consequently, affect reflectance, which may have caused good prediction models for the LDM variable.
Regarding the variables N, Ca, Cu, Fe, and Mn, the R 2 p found (<0.50) indicated the existence of models with imprecise adjustments, thus harming the good prediction of plant nutrients. For N, the results of R 2 p obtained in this study were lower than the ones of Tzanakakis et al. 58 studying several crops and Zhang and Li 48 in an experiment with cucumber.
A factor that may have compromised the good prediction of certain nutrients was the type of input used, in this case, mining coproducts. These materials are multi-elementary sources that provide different nutrients to plants in different quantities (Table 4), what makes it difficult to isolate the effect of a certain nutrient. Thus, the plants showed deficiencies for some nutrients such as N, P, K, S, and Mn, which may have presented a condition of "hidden hunger" that interferes with the spectral curve, 59 making it difficult to build the prediction model. Furthermore, the interaction between ions also influences the nutritional condition of the plant 60 and consequently its spectral response.
In relation to micronutrients, several studies have shown that calibrations generally perform less well than those for macronutrients, 61,62 as it was the case in this study. According Liao et al. 62 and Menesatti et al., 13 the NIR and Vis-NIR prediction of micronutrients may fail mainly because of their low concentrations in the leaves. However, for the micronutrient Zn, according to Rossel et al. 34 based on R 2 p , the model shows moderate prediction that may be used for evaluation and correlation. Although Zn is a metallic ion, which does not produce inflections or absorption characteristics in the Vis-NIR region, it can bind electrostatically or as ligands to larger carbon-containing compounds, which enable the quantification by Vis-NIR spectroscopy. 6 In relation to the RMSE p , the variables that presented the smallest proportional values to the units of the evaluated variables were S, Ca, LDM, and N. For these variables, the RMSE p was below 12.5 % in relation to the average of these variables. As for S and LDM, these results corroborate the highest R 2 p found for these variables (>0.50). Other studies report similar RMSE p values for N, P, and K in maize 11 and lower for N, P, K, Ca, Mg, Fe, Mn, and Zn in citrus 13 than those obtained in this study.
As for the RCs (Fig. 6), their interpretation is necessary to avoid possible accidental correlations. 63 Thus, one can better understand how each variable (wavelength) contributes to the significant variation of the LDM estimate model and soybean nutrient contents. For the LDM variable, high RC close to 500 nm may be related to chlorophyll 41 since studies have shown that wavelengths close to this region correlate with chlorophyll and plant nutrient concentration, 64 which allows greater production of LDM.
Thus, the concentration of chlorophyll and nutrient contents are important for leaf mesophyll development, which may have caused high RC near 750 nm. As for the N contents, the RC obtained in this study is supported by other studies that aimed at estimating nitrogen by the Vis-NIR spectroscopy in other crops, such as pepper 65 and sugarcane. 32 Concerning the variables P and K, the results of the RC found in this study are corroborated by the work of Zhang et al., 49 who studied the use of image Vis-NIR spectroscopy to predict P and K in rapeseed crop. Oliveira and Santana, 66 when evaluating the use of Vis-NIR spectroscopy for the prediction of nutrients in Eucalyptus, also found high RC values in the region of 400 to 900 nm for P and N, as in this work. Both nutrients are closely related to photosynthetic traits in plants, 67 which is directly related to high RC in the Vis-NIR reflectance. Pinstein et al. 51 found, in the most part of the spectrum Vis-NIR, a correlation pattern between K and N, suggesting some cross-correlation involving these two elements at the leaf level, which allows the creation of vegetation indices close to these regions. In addition, other studies report the importance of wavelengths of the green 68 and red 69 regions to estimate the concentration of P and K in plants.
The important wavelengths selected for Ca prediction were found in two regions (around 550 and 720 nm), according to previous research studies such as Oliveira and Santana 66 studying the Ca content in Eucalyptus. The reflectance in wavelengths near from 500 nm is associated with active photosynthetic radiation, 49 especially the role of Ca as a cofactor in the photosystem II. 70 The NIR region also presented higher absolute RC values for Ca prediction. Among the functions of Ca in higher plants are cell wall synthesis and cell membrane integrity. 71 In this sense, there is a relationship between the structural components (mesophyll) and the energy in the NIR region. 42 For Mg, the largest RC near 680 nm can be related to the chlorophyll molecule since Mg integrates it, besides being an activator of some enzymes in plants. 71 Regarding S, the high values of RC at near 700 nm are supported by the work by Cozzolino and Moron, 54 who found, near this region, high correlations with S. Furthermore, wavelengths in this region have been used in the management of S in wheat (Triticum aestivum L.). 72 Finally, for Zn, the high values of RC around 400 and 620 nm can be related to chlorophyll, 41 which, in turn, can be correlated with nutrient concentrations. 64

Applications of Vis-NIR Spectroscopy as a Technique for Estimating LDM and Nutrients Contents in Soybean
Soybean plants require significant amounts of mineral nutrients to achieve adequate growth and productivity, and this need depends on the fertility and class of the soil. Although the mineral nutrition of soybeans and the conventional techniques for assessing the nutritional status are much studied, there is limited research on the use of alternative techniques to routine analysis, such as the use of the Vis-NIR spectroscopy in predicting LDM and leaf nutrient contents. Soybean leaves are a complex mixture of chemical compounds such as water, pigments, carbohydrates, proteins, and others. All these compounds contribute substantially to generate a characteristic spectrum of the target (leaves). The nutritional and physical state of the plant (thickness and roughness) also helps to form a differentiated spectrum. In this study, the application of mining coproducts (alternative inputs) in the soil, different in input particle size and solubility, allowed a wide diversity of nutritional conditions, varying between better-nourished plants and others with deficiencies (Table 4). Thus, there was a high CV for the evaluated variables, which affected, in part, the creation of models with better R 2 p and RPD p for certain variables. 73 Despite its limiting conditions, this study demonstrates potential use of the Vis-NIR spectroscopy to predict variables that presented R 2 p and RPD p above 0.50 and 1.50, respectively, as it was the case of variables LDM, P, K, Mg, S, and Zn, with emphasis on the first four, which presented R 2 p above 0.65. The results of this study demonstrated that a system based on a portable equipment, such as the Vis-NIR spectrometer, can provide a better knowledge of the nutritional condition of certain nutrients as well as estimate LDM of soybean grown after application of mining coproducts. Because it is a fast, low-cost, and non-destructive method, Vis-NIR spectroscopy can be applied in areas cultivated with annual crops such as soybean. Due to the ease of acquisition of this equipment and the processing of the data, it can be used in areas with precision agriculture, considering that this type of agriculture requires a large number of samples. 74 Although there are many studies reporting the use of the Vis-NIR-SWIR range as a technique for predicting nutrients in plants, this work demonstrated that the use of the Vis-NIR (380 to 1000 nm) range has the potential to generate satisfactory prediction models for several attributes of soybean plants. This allows for the acquisition of cheaper spectroradiometers, which do not necessarily need to operate in the SWIR range.
Finally, the positive results obtained, in this study, indicate that it may also be feasible to use of the Vis-NIR spectroscopy to evaluate other chemical compounds present in soybean that are important in commercialization.

Conclusion
The application of mining coproducts enabled a wide diversity of nutritional conditions in soybean cultivation, what modified the spectral curves of the plants and contributed to the high range of minimum and maximum values of the nutrient contents (plants).
With the use of PLSR, it was possible to adjust models in the prediction stage with R 2 p and RPD p above 0.50 and 1.50, respectively, for LDM, P, K, Mg, S, and Zn. The contents of N, Ca, Cu, Fe, and Mn showed models with imprecise adjustments (R 2 p < 0.50), compromising the good prediction of these plant nutrients.
In general, the most important wavelengths for the construction of the PLSR model according to the regression coefficients were close to 380 to 400 nm, 500 to 530 nm, 600 to 690 nm, and 700 to 750 nm.
The study confirms the potential of using Vis-NIR spectroscopy to predict LDM and nutrients contents in soybean crops. Thus, Vis-NIR spectroscopy presents itself as a promising alternative to the routine analysis of the nutritional diagnosis of plants for having advantages such as speed, low cost, and no use of reagents that are toxic to the environment; due to the ease of data acquisition and processing, the Vis-NIR spectroscopy allows application in areas with precision agriculture. Further research is necessary to develop more robust models for nutrients such as N, Ca, Cu, Fe, and Mn with control of the variation of individual nutrients.
Although there are studies reporting the use of spectroscopy as a technique for predicting nutrients in plants, the Vis-NIR-SWIR range is used in most studies. Thus, the results found in the present work are promising since only the Vis-NIR range was used and it was possible to build models with good prediction for different nutrients. This allows for the acquisition of cheaper spectroradiometers, which do not necessarily need to operate in the SWIR range.
Marlon Rodrigues received his MSc degree in science from the Federal University of Pelotas, Brazil, in 2017. He is currently a PhD student at State University of Maringá, Brazil. His research interests include surveying and mapping, nutrients, and plant biomass estimation using remote sensing.
Marcos Rafael Nanni is a full professor at the State University of Maringá, Brazil. His research interests include remote sensing, geoprocessing, geographic information system, soil survey and cartography, diagnosis and planning of conservation units, diagnosis and agricultural planning, and precision agriculture. He has authored more than 200 publications, including more than 100 journal papers. He is currently the coordinator of the Graduate Program in Agronomy (concept 6) at UEM. Everson Cezar is currently an adjunct professor in the Department of Agronomy, the State University of Maringá, where he teaches the discipline of soil and fertilization, soil fertility, soil classification and geoinformation applied to agronomy, soil genesis, and morphology and precision agriculture. He has authored more than 100 publications, including more than 40 journal papers.
Glaucio Leboso Alemparte Abrantes dos Santos received his MSc degree in agronomy from the Federal University of Viçosa, Brazil, in 2011. He is currently a PhD student at the State University of Maringá, Brazil, and a professor at the Centro Universitário Ingá, Brazil. His research interests include surveying and mapping, soil attributes estimation using remote sensing.
Amanda Silveira Reis received her MSc degree in agronomy from the State University of Northern Paraná, Brazil, in 2017. She is currently a PhD student at State University of Maringá, Brazil. Her research interests include surveying and mapping, soil organic matter estimation using remote sensing.
Karym Mayara de Oliveira received her MSc degree in agronomy from the State University of Maringá, Brazil, in 2020. She is currently a PhD student at State University of Maringá, Brazil. Her research interests include biomass and plant nutrients prediction using remote sensing applications.
Roney Berti de Oliveira is currently an adjunct professor in the Department of Civil Engineering, the State University of Maringá, where he teaches the discipline of geoprocessing applied to soil survey and spatialization and topography applied to architecture and urbanism. He has authored more than 90 publications, including more than 20 journal papers.