1. Introduction
Year after year, Brazil breaks soybean production records, and this increase is due to technological advances in several stages of the production process, such as the development of genotypes more adapted to the adverse environmental conditions that tropical and subtropical countries experience [1].
Cultivation species of commercial interest have undergone thousands of selections since the beginning of their cultivation, seeking to select individuals that had the best performance in certain environments. As a result, extensively cultivated plants present more productive responses, adapted to different types of stress or cultivation limitations [2], such as nutritional restrictions, without affecting good plant development.
Due to the economic importance that soybeans have for Brazil, maintaining high productivity to meet demand is of relative importance. Obtaining information quickly and on a large scale has become essential for commodity production [3]. In this scenario, the nutritional monitoring of soybean plants is important for the adequate development of the crop and, mainly, the identification of genotypes that maintain good performance even in restricted environments with good conditions for development.
Soybean nutritional determination analyses are accurate, despite requiring reagents and equipment for the process, in addition to the fact that, in the laboratory approach, there is a need to destroy samples, and that the process of obtaining information is time-consuming [4,5]. High-throughput phenotyping (HTP) allows for the provision of detailed information about the crop throughout the cycle in a quick, efficient manner that is non-destructive of the plants’ leaves [6,7].
Unmanned aerial vehicle (UAV) platforms allow onboard multispectral sensors for the spatiotemporal monitoring of vegetation characteristics to obtain HTP information [8]. Therefore, these HTP technologies allow for the selection of high-performance genotypes with smaller amounts of nutrients, including micronutrients, and the selection of soybean genotypes that have good productive performance in low fertility soils, which is essential in enhancement programs [9], notwithstanding the fact that the maintenance of micronutrients is often forgotten by producers. These technologies are used to classify soybean genotypes based on macronutrient content, a complex challenge due to the lack of a linear relationship between spectral data and nutritional variables. However, machine learning algorithms can overcome this difficulty [9]. In addition, the use of these technologies can be extended to other crops such as corn, relating nitrogen content in leaves with spectral variables [10] or estimating grain yield in corn using a multispectral sensor associated with the Random Forest algorithm, a ML model widely applied in agricultural tasks [11].
Data generated through UAV multispectral images are numerous and need to be associated with crop characteristics, but these associations are not linear, requiring the use of efficient statistical analyses in processing [12]. The use of machine learning techniques is efficient in providing quick and accurate answers about the relationships between agronomic or physiological traits of plants and their spectral responses [13].
Identifying machine learning models that are capable of classifying soybean genotypes according to their micronutrient content using only spectral data as input is relevant and useful for plant breeding programs and agricultural producers. This allows for the identification of genetic materials that use micronutrients more efficiently and for management practices to be adapted. Most studies have used vegetation indices (VIs) as inputs in prediction and classification models, while other studies use only the spectral bands provided by the sensor as input. There is still no agreement on which strategy provides greater classification accuracy involving leaf nutrient content in soybeans.
The objectives of this work are: (I) This study aims to find machine learning models capable of identifying and classifying soybean genotypes according to micronutrient content using multispectral data and (II) to determine which combinations of input variables maximize the accuracy of the models’ classification.
2. Materials and Methods
2.1. Field Experiment
A field experiment was conducted in the 2019/20 harvest with 103 F2 soybean populations in the experimental area of the Federal University of Mato Grosso do Sul, in Chapadão do Sul—MS (coordinates: 18°41′33″ S, 52°40′45″ W) at an altitude of 810 m. The soil of the site is a clayey Red Latosol. The chemical properties of the soil in the 0–0.20 m profile are detailed in Table 1 [14]. According to the Köppen classification, the climate of the region is characterized as Tropical Savanna (Aw). The climatic conditions during the experiment are illustrated in Figure 1.
Soybean genotypes were sown in October 2019 in a soil preparation system with plowing and harrowing. The spacing between rows was 45 cm, with a density of 15 plants per meter. The experiment was conducted in randomized blocks with two replications. Cultural treatments were adjusted according to the specific requirements of the crop at each stage of the development cycle. This included the application of fertilizers and pest and disease control when necessary. Each adjustment was made based on the observed needs of the plant in relation to the growth stage and the prevailing climatic conditions, ensuring that the crop received the appropriate inputs at each stage to optimize development and productivity. Before planting, the seeds were treated with fungicide (Pyraclostrobin + Thiophanate Methyl) and insecticide (Fipronil) at a dose of 200 mL per 100 kg of seeds, in addition to being inoculated with Bradyrhizobium bacteria, using 200 mL of concentrated liquid inoculant per 100 kg of seeds.
2.2. Spectral and Nutritional Data
The null hypothesis (H0) of our work is that the spectral variables, when used in machine learning models, do not present a significant relationship with the classification of genotypes regarding micronutrient use, indicating that the models are not capable of differentiating between genotypes. The hypothesis of the work is that there is an ideal set of spectral variables that allows for the classification of soybean genotypes with different levels of micronutrient use efficiency. Therefore, when the crop was in full bloom, the phenological stage of greatest photosynthetic activity and greatest nutrient absorption, 60 days after emergence (DAE), a flight was carried out with a fixed wing aircraft called Sensefly eBee RTK that works with a remote pilot (RPA). The images were taken at 09:00 a.m., a time of day when the sky was cloudless; the flight height above ground level was 85 m, and the average field spatial resolution in each image was 0.089 m, which was adopted for the survey described. The image overlap was 80% and 65% along and across, respectively. The RTK (Real Time Kinematics) technology on board the aircraft was coupled to the Parrot Sequoia sensor. At the peak flight height, which was 80 m in altitude, 30 scenes were taken with a resolution of 8 cm. Each scene had 5 images. Therefore, the total number of images was 30 × 5, with 80% overlap in the longitudinal and 65% between the passes. For accuracy reasons, the images are taken in 3 passes so that the orthorectification is adequate; the other flight instructions and image processing were followed according to [15]. Eighteen ground control point (GCP) markers were placed on the ground around the study area and surveyed with a pair of the Emlid Reach real-time kinematic (RTK) GNSS models. The average flight speed of the RPA was 12.5 m s−1. The Parrot Sequoia has a light sensor that allows for the calibration of the values obtained. Radiometric calibration was conducted for the entire scene using a calibrated reflective target provided by the manufacturer, which ensures the accuracy of the reflectance captured by the sensor. The Sequoia multispectral sensor incorporates an ambient light sensor, allowing the recorded reflectance values to be adjusted in real time for each image, compensating for variations in lighting conditions. This field calibration procedure is performed immediately before flight, ensuring that environmental conditions are taken into account at the time of data capture. The field calibration procedure uses a reference image, which is captured and processed by e-Motion software, version 3.2 configured to automatically perform radiometric adjustment based on the calibration target. The Sequoia sensor used has a horizontal field of view (HFOV) of 61.9°, a vertical field of view (VFOV) of 48.5°, and a diagonal field of view (DFOV) of 73.7°, which allows for wide coverage and facilitates the acquisition of accurate and consistent images during flights. The spectral bands (SBs) obtained by the sensor were: 550 nm, 660 nm, 735 nm, and 790 nm, corresponding to green, red, red edge, and NIR. The images were processed by mosaicking and orthorectifying using the Pix4Dmapper computer program.
After collecting spectral data with the UAV, several vegetation indices (VIs) were calculated, including SAVI, NDRE, EVI, NDVI, MSAVI, GNDVI, MCARI, and SCCCI. On the same day, six trifoliates (the soybean plant in which the main stem presents six sets of trifoliate leaves, that is, leaves composed of three leaflets each) from each plot were collected, washed in different solutions, dried in an oven, and ground in a Wiley mill. Subsequently, the contents of micronutrients, such as boron (B), iron (Fe), copper (Cu), zinc (Zn), and manganese (Mn), were determined following the Bataglia methodology [16].
2.3. Data Analysis
A Pearson correlation network was constructed to graphically demonstrate the relationship between spectral data and the micronutrient content of the genetic materials in question. The soybean genotypes were divided into two clusters using the K-means algorithm, which groups treatments whose centroids are closer together until there is no significant variation in the minimum distance from each observation to each of the centroids. The intention was to divide the soybean genotypes into two groups based on nutritional characteristics, generating useful categories for subsequent supervised classification steps. This preliminary grouping allowed for the creation of a framework to evaluate the differences between groups, identifying an internal pattern in the data.
To represent the cluster divisions, a biplot was constructed with principal component analysis (PCA) from the R software version 4.1.0 [17,18]. Machine learning (ML) models were applied to classify genotypes according to clusters by K-means algorithm. The generated clusters were used as output variables for the classification models, while the input variables were the spectral data of the plants. The K-means algorithm was used for multivariate clustering purposes, not as a machine learning technique for training predictive models. The intention was to divide the soybean genotypes into two groups based on the multivariate characteristics of their nutritional characteristics, generating useful categories for subsequent supervised classification steps. This preliminary clustering allowed for the creation of a framework for evaluating differences between groups without the need for a standard, since the initial intention was not to predict categories, but rather to identify internal patterns in the data. K-means, therefore, served as an exploratory tool, and not as a high-precision machine learning technique. This initial clustering served to structure the data, facilitating the subsequent classification step with the supervised models.
The spectral data for machine learning analysis were organized into three distinct sets: (1) only SBs, (2) only VIs, and (3) a combination of SBs and VIs, in order to identify which set offered greater accuracy. The classification of the clusters was conducted using stratified cross-validation with K-fold of 10, repeated ten times to increase the reliability of the results. All parameters of the classification models were kept in the default settings of the Weka software, version 3.8.5, ensuring the replicability of the tests and the standardization of the results. The models tested were J48 Decision Tree (J48), Random Forest (RF), Support Vector Machine (SVM), Perceptron Multilayer Neural Network (ANN), Logistic Regression (LR), and REPTree (DT); such algorithms were used and compared to their efficiency in work as: [19,20,21]. The algorithms used in this study were chosen because they are the main shallow machine learning algorithms, extensively tested and reported to be effective in solving classification and regression problems in agricultural sciences (Figure 2). The performance of the models was evaluated by accuracy metrics: correct classifications (CC), F-score, and Kappa coefficient.
It is important to highlight that, by using cross-validation, we ensure that the results presented are predictions of the validation and not of model calibration. The ML model is trained on all but one-fold (k-1), and it evaluates the model on the dataset that was not used for training. Here, a random cross-validation sampling strategy with K-fold = 10 and 10 repetitions (total of 100 runs for each model) was applied. As this process was repeated, 100 random test sets were constructed. In summary, this type of validation is repeated sequentially, constantly changing the folder used to validate the algorithm. In this way, the algorithm is always validated with data not used in its training phase. In this experiment, the entire procedure was also repeated 100 times, which means that the models were built from scratch in each repetition.
Once the accuracies of each model and input were obtained, these results were subjected to analysis of variance to evaluate the significance effects of ML models and the inputs tested and the interaction between them. If there was significance for ML, input, or the interaction between them, boxplots were constructed based on the averages of the accuracy metrics, comparing them using the Scott-Knott test at 5% significance [22]. These steps were based on the ‘ggplot2’ and ‘ExpDes.pt’ packages of the R software version 4.1.0.
3. Results
The relationships between spectral variables and micronutrients can be seen by the Pearson correlation network (Figure 3). Red lines indicate negative correlations, green lines indicate positive correlations and the thickness of the line indicates the magnitude of the correlation, becoming thicker when the correlation is above 0.6. The purpose of the figure is to demonstrate that conventional statistical analyses such as Pearson’s correlation do not allow for the establishment of a linear relationship between micronutrient levels and spectral information. There was a high correlation, both positive and negative, between the spectral bands (SBs) and the vegetation indices (VIs) studied, which is expected behavior, since the bands capture data in spectral ranges that partially overlap, and these same data are used to calculate the vegetation indices. Positive correlations can be observed between nutritional variables, with higher magnitudes between Cu and Fe and lower magnitudes between the other micronutrients. Low magnitude or nonexistent correlations are observed between micronutrients and spectral variables, demonstrating the complexity of the relationship between these data.
Principal component (PC) analysis combined with the K-means algorithm allowed us to distinguish soybean genotypes into two groups, cluster 1 and cluster 2 (Figure 4). Clusters were grouped according to the micronutrient content exhibited by the genotypes, showing similarity between the genotypes belonging to the same cluster and distinctions by comparing to the other cluster.
After dividing the 32 soybean genotypes into two groups, boxplots were constructed based on the average nutrients for each cluster formed (Figure 5), seeking to find differences between the groups. The genotypes in cluster 1 presented subtly higher B values compared to those in cluster 2. Similarly, the genotypes in cluster 2 presented the highest levels of Mn and Zn, allowing us to affirm that they are more efficient in the use of these nutrients. Regarding Cu and Fe content, there was no difference between the clusters formed, indicating that there is no difference between the contents of these nutrients in the leaves of the different genotypes evaluated.
We proceeded by reducing the 32 genotypes to two clusters, based on the content of foliar micronutrients, dividing them into groups with high and low micronutrient contents, and differentiating the genotypes based on their efficiency in the use of micronutrients. This approach is relevant and useful for genetic improvement programs and agricultural producers, as it allows for the identification of genetic materials that use micronutrients more efficiently. Given that all genotypes were grown in the same location and under the same management, the variations observed between the clusters reflect inherent differences between the genotypes, allowing management practices to be adapted or promising materials to be selected for nutritional efficiency.
The accuracy metrics used to evaluate the models were analyzed using analysis of variance to investigate the relationship between ML models and inputs. For the three metrics evaluated, there were interactions between ML models and inputs (Table 2), thus leading to the breakdown for each precision, expressed in boxplot graphs.
In CC, the tested inputs did not show a significant difference for the ANN, LR, RF, and SVM algorithms (Figure 6), which presented with similar performances. The DT, J48, and SVM algorithms did not differ from each other when using the VIs and VIs +SBs inputs. The ANN, LR, and RF models performed the same regardless of the inputs used, while DT, J48, and SVM performed better when VIs information was inserted into the input.
When using the F-score accuracy variable, the tested inputs did not differ from each other when using the ANN, LR, and RF algorithms (Figure 6). The DT, J48, and SVM algorithms performed better with inputs that contained vegetation indices (VIs and VIs+SBs). Evaluating the performance of each input for the algorithms separately, it is noted that for SBs, the ANN, LR, and RF algorithms presented better performance. When using VIs, all algorithms achieved good results, with the exception of LR and SVM. Using the SBs+ VIs input, every applied algorithm reached the same performance.
From the Kappa coefficient, it can be seen that ANN, LR, and RF had similar performance, regardless of the input. DT, J48, and SVM had better performance when VIs or VIs+SBs were taken as inputs (Figure 6). With VIs as input, it is noted that DT, J48, and RF presented higher performance. Every algorithm using VIs+SBs as all inputs obtained similar accuracy responses. With only SBs as inputs, the ANN, LR, and RF algorithms had the best performances.
The RF algorithm performed well regardless of the input used, demonstrating that it is a robust model with good data generalization capacity. The DT and J48 algorithms performed well when using VIs or VIs+SBs inputs. The SVM algorithm performed well with VIs+SBs as input. In general, the inputs that contained information about VIs provided better results for the models in the classification of soybean genotypes.
4. Discussion
The high correlation between vegetation indices (VIs) is due to the approach used with spectral bands (Figure 1). VIs are mathematical models characterized as reliable algorithms for evaluating vegetation cover, growth dynamics, and the nutritional status of plants [17]. Its application at different wavelengths is effective in the accurate and timely monitoring of crops [15].
The correlation among nutritional variables may be related to photosynthesis and chlorophyll content in the vegetation canopy, due to its relationship with nutrient content and, consequently, crop productivity [23]. Copper is an essential element of lignin and cuticle formation, as well as contributing to the constitution of enzymes found in chloroplasts [24]. Iron, in turn, is related to the constitution of proteins that act in the photosynthetic process, in addition to being a catalyst for the biosynthesis of chlorophyll [3]. Boron plays a key role in plant metabolism, acting in cell division and elongation, from root formation to grain production [25]. Manganese participates in the activation of enzymes and the functioning of photosystem II [5]. Zinc is a vital enzymatic cofactor for the development of chloroplasts [26]. Thus, a correlation is observed between Cu and Fe, and between both and B, as well as between B, Mn, and Zn, which can be related to the fact that both participate in some direct or indirect part of the photosynthetic process.
The negative correlation found between green and red and vegetation indices can be attributed to the reflectance of green by chloroplasts and the absorption of other plant pigments by plants, such as red by shaded parts of plants [27]. The correlation of green and red with Cu and Fe highlights this condition, since these nutrients are directly related to chloroplasts and chlorophyll. The negative correlation with vegetation indices may be the result of small biochemical changes in plant nutrition, resulting in variations in photosynthetic activity, cellular structure, and, consequently, promoting changes in the reflectance of plants captured by VIs [24].
The characteristic spectral reflectance of vegetation occurs through the expression of three spectra of electromagnetic radiation, namely visible, near-infrared, and short-wave infrared. In the visible spectrum, the absorption of light by plant pigments causes low reflectance, except for a small peak in the green region. In the near-infrared, the internal scattering of light by mesophylls also reduces reflectance, while in the short-wave infrared, reflectance is influenced by cellulose and lignin [28,29]. Thus, it can be deduced that vegetation indices are negatively correlated to green and red bands, due to light absorption conditions, i.e., by photosynthetic organelles that depend on the presence of micronutrients Cu, Fe, B, Mn, and Zn.
The micronutrient contents in soybean genotypes differ from each other, as can be seen in both clusters (Figure 2). Clustering allows for the grouping of genotypes with similarity to each other and distinction from others, such as the occurrence of higher levels of Mn and Zn in the genotypes present in cluster 2, higher levels of B in cluster 1, and the absence of differences in Cu and Fe between clusters (Figure 3).
Soy is a powerful candidate for food biofortification due to its constitution of soluble and fat-soluble vitamins [30]. Therefore, soybean genetic improvement programs focus on the occurrence and increments of nutrients, due to their composition of essential minerals that will make up the grains, such as copper, iron, manganese, zinc, and boron [31]. Furthermore, the compounds present in plant leaves, such as chlorophyll molecules, carotenoids, carbohydrates, and proteins, depend on the presence of micronutrients in their constitution and these contribute to the formation of characteristic reflectance curves [3].
The combination of ML and remote sensing techniques makes it possible to obtain reliable information on various study targets [32]. In the present methods and data, the ANN, LR, and RF algorithms performed well in relation to ML, regardless of the inputs used in the CC parameter (Figure 4). The DT, J48, and SVM algorithms showed better performance when using VIs and VIs+SBs inputs, in both the f-score parameter (Figure 5) and kappa coefficient (Figure 6). Successful application of ANNs requires careful selection of their parameters, which can vary significantly depending on the specific task [33]. In this work, it was observed that the best algorithms varied according to the parameter used but displayed similarity in relation to the inputs used and the superiority of the algorithms.
The ANN and RF algorithms showed good performance regardless of the input used in the f-score parameter, while in the kappa coefficient, the RF algorithm showed the best performance regardless of the input. The superiority of the RF algorithm can be attributed to the fact that it is a classifier that uses samples and training variables to produce several decision trees, making predictions after aggregating results from individual trees [32]. The RF can be defined as a collection of decision tree classifiers, created to increase classification accuracy, while ANN is a supervised trainer that groups data and estimates the desired function [34].
RF has the ability to process high-dimensional data and to capture complex patterns and interactions between variables, such as capturing both linear and nonlinear relationships and handling multicollinearity [35,36], which is particularly relevant in image-based data. The superiority of RF has been extensively reported in prediction and classification studies using spectral data [10,11,37]. Ref. [38] attributes the outstanding ability of Random Forest in classification tasks using spectral variables to its characteristic of randomly selecting features for training, which diminishes the impact of high dimensionality in real spectroscopy data. Outstanding RF performance in spectral information processing has also been attributed to its capacity to extract relevant spectral information and decrease redundant data [39]. RF can successfully select and rank variables with the greatest ability to discriminate between the target classes, which is an important advantage, since the high dimensionality of remotely sensed data makes the selection of the most relevant variables a time-consuming and error-prone task [36].
The use of SBs, Vis, or VIs+SBs differed in relation to the algorithms, with the RF algorithm being superior to the others because, regardless of the input, it stood out. The use of spectral bands for data processing is the fastest way to process information, since the VIs depend on the SBs to compile the data and carry out the calculations, requiring more processing time, and the use of SBs is therefore recommended for this purpose. This result corroborates conditions seen in the discrimination of soybean cultivars [6] and in genotype classification related to the industrial characteristics of grains [40].
The development of this method for classifying soybean genotypes based on spectral data to differentiate micronutrient levels has an impact and importance especially applied to genetic improvement and agricultural management. It allows genetic improvement programs to more quickly and accurately identify genetic materials that are more efficient in the use of micronutrients, helping to select plants that can offer high productivity with less need for inputs. For farmers, this translates into the possibility of optimizing the use of fertilizers and reducing costs, in addition to promoting more sustainable agriculture with less of an environmental impact. In addition, this methodology based on spectral data contributes to the advancement of precision agriculture techniques.
The evidence found in this work suggests that the use of the green and red bands correlates with micronutrients, due to their relationship with enzymes and photosynthetic processes. The PC analysis using the K-means algorithm was efficient in separating the genotypes with the highest concentration of nutrients according to their similarity and divergence with the others, especially when observing the Mn and Zn contents in the genotypes of cluster 2 and B in the genotypes of cluster 1. Regarding the ML technique, the superiority of the RF algorithm was noted, regardless of the input used. The choice of input in relation to SBs, Vis, and VIs+SBs must take into account the ease and speed of obtaining information, and the use of SBs is therefore recommended, since VIs use SBs in formulating calculations and obtaining results. This allows for the identification of genetic materials that use micronutrients more efficiently and management practices to be adapted. In addition, the decision to be made can be made quickly, without the need for chemical evaluation in the laboratory. Future scientific efforts on the subject should consider the investigation of other nutrients in soybean genotypes and the use of SBs for separation and identification of results, since the selection of soybean genotypes depends on the purpose of the selection, and thus providing information on the increase in nutrient content in the leaves, and, consequently, in grains.
5. Conclusions
The motivation for this study was to increase the accuracy and efficiency of soybean genotype classification based on their micronutrient content. Traditional methods of analyzing micronutrient levels are often time-consuming and laborious, creating the need for faster and non-destructive techniques, especially those that use fewer reagents that are toxic to humans and the environment. This study sought to explore the potential of spectral reflectance data and vegetation indices, in combination with machine learning algorithms, providing a more simplified and accurate approach. The purpose of proposing the methodology studied in the work is to improve the ability to identify and differentiate soybean genotypes based on their nutritional profile, contributing to more efficient breeding programs and better crop-management practices. Soybean genotypes can be effectively classified according to their Cu, Fe, Zn, Mn, and B content using spectral reflectance data and vegetation indices. The application of the Random Forest algorithm allows for this classification in a robust manner, regardless of the input used. The use of spectral bands is desirable in relation to vegetation indices, as it speeds up the collection, identification, and separation of information, making the analysis more efficient and accurate. This allows for a faster and more detailed characterization of genotypes in relation to their nutritional profile.
Conceptualization, S.B.C. and D.C.S.; methodology, D.C.S., I.C.d.O., J.L.D.-S. and P.E.T.; software, D.C.S. and P.E.T.; validation, M.C.M.T.F., R.d.C.F.A., F.H.R.B., R.G. and L.P.R.T.; formal analysis, D.C.S.; investigation, P.E.T.; resources, M.C.M.T.F. and C.A.d.S.J.; data curation, P.E.T.; writing—original draft preparation, D.C.S. and J.L.D.-S.; writing—review and editing, S.B.C. and D.C.S.; visualization, L.P.R.T.; supervision, P.E.T.; project administration, P.E.T.; funding acquisition, P.E.T. All authors have read and agreed to the published version of the manuscript.
The raw data supporting the conclusions of this article will be made available by the authors on request.
The authors declare no conflict of interest.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Enlarge this image.Figure 3. Pearson correlation network between nutritional and spectral variables. Green lines indicate positive correlations, while red lines indicate negative correlations.
Enlarge this image.Figure 4. Principal component (PC) analysis and K-means algorithm for clustering samples definition in view of the genotypes’ micronutrient content.
Enlarge this image.Figure 6. Boxplot of correct classification (%), F-score, and Kappa for machine learning models. The means followed by the same capital letters do not differ from each other for the inputs; the means followed by the same lowercase letters do not differ from each other for the algorithms.
Soil chemical properties in the 0–0.20 m profile.
| pH | Al | Ca + Mg | K | S.B. | CTC | P | M.O. | V | m |
| H2O | cmolc dm−3 | mg dm−3 | g dm−3 | % | |||||
| 6.2 | 0.0 | 4.31 | 0.2 | 2.3 | 5.1 | 41.3 | 19.74 | 45 | 0.0 |
Al: exchangeable aluminum; S.B.: Sum of Bases; M.O.: organic matter.
Summary of the analysis of variance for the accuracy metrics used in evaluating the behavior of machine learning models.
| SV | DF | CC | F-Score | Kappa |
|---|---|---|---|---|
| Inputs | 2 | 13.985 * | 0.011245 * | 0.006495 * |
| ML | 5 | 7.346 * | 0.007084 * | 0.0044974 * |
| inputs * ML | 10 | 6.125 * | 0.00435 * | 0.0023323 * |
| Residual | 162 | 3.12 | 0.000769 | 0.0016 |
| CV | 1.89 | 4.81 | 3.12 |
* Significant at 5% probability by the F test. SV: source of variation; DF: degrees of freedom; CV: coefficient of variation; CC: correct classification.
References
1. Cattelan, A.J.; Dall’Agnol, A. The Rapid Soybean Growth in Brazil. OCL; 2018; 25, D102. [DOI: https://dx.doi.org/10.1051/ocl/2017058]
2. O’Rourke, J.A.; McCabe, C.E.; Graham, M.A. Dynamic Gene Expression Changes in Response to Micronutrient, Macronutrient, and Multiple Stress Exposures in Soybean. Funct. Integr. Genom.; 2020; 20, pp. 321-341. [DOI: https://dx.doi.org/10.1007/s10142-019-00709-9] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31655948]
3. dos Santos, G.L.A.A.; Reis, A.S.; Besen, M.R.; Furlanetto, R.H.; Rodrigues, M.; Crusiol, L.G.T.; de Oliveira, K.M.; Falcioni, R.; de Oliveira, R.B.; Batista, M.A. Spectral Method for Macro and Micronutrient Prediction in Soybean Leaves Using Interval Partial Least Squares Regression. Eur. J. Agron.; 2023; 143, 126717. [DOI: https://dx.doi.org/10.1016/j.eja.2022.126717]
4. Furlanetto, R.H.; Rafael Nanni, M.; Guilherme Teixeira Crusiol, L.; Silva, G.F.C.; Junior, A.d.O.; Sibaldelli, R.N.R. Identification and Quantification of Potassium (K+) Deficiency in Maize Plants Using an Unmanned Aerial Vehicle and Visible/near-Infrared Semi-Professional Digital Camera. Int. J. Remote Sens.; 2021; 42, pp. 8783-8804. [DOI: https://dx.doi.org/10.1080/01431161.2020.1871091]
5. Rodrigues, M.; Nanni, M.R.; Cezar, E.; dos Santos, G.L.A.A.; Reis, A.S.; de Oliveira, K.M.; de Oliveira, R.B. Vis–NIR Spectroscopy: From Leaf Dry Mass Production Estimate to the Prediction of Macro-and Micronutrients in Soybean Crops. J. Appl. Remote Sens.; 2020; 14, 44505. [DOI: https://dx.doi.org/10.1117/1.JRS.14.044505]
6. Gava, R.; Santana, D.C.; Cotrim, M.F.; Rossi, F.S.; Teodoro, L.P.R.; da Silva Junior, C.A.; Teodoro, P.E. Soybean Cultivars Identification Using Remotely Sensed Image and Machine Learning Models. Sustainability; 2022; 14, 7125. [DOI: https://dx.doi.org/10.3390/su14127125]
7. Silva Junior, C.A.d.; Teodoro, P.E.; Teodoro, L.P.R.; Della-Silva, J.L.; Shiratsuchi, L.S.; Baio, F.H.R.; Boechat, C.L.; Capristo-Silva, G.F. Is It Possible to Detect Boron Deficiency in Eucalyptus Using Hyper and Multispectral Sensors?. Infrared Phys. Technol.; 2021; 116, 103810. [DOI: https://dx.doi.org/10.1016/j.infrared.2021.103810]
8. Khan, Z.; Rahimi-Eichi, V.; Haefele, S.; Garnett, T.; Miklavcic, S.J. Estimation of Vegetation Indices for High-Throughput Phenotyping of Wheat Using Aerial Imaging. Plant Methods; 2018; 14, 20. [DOI: https://dx.doi.org/10.1186/s13007-018-0287-6]
9. Santana, D.C.; Teixeira Filho, M.C.M.; da Silva, M.R.; Chagas, P.H.M.d.; de Oliveira, J.L.G.; Baio, F.H.R.; Campos, C.N.S.; Teodoro, L.P.R.; da Silva Junior, C.A.; Teodoro, P.E. Machine Learning in the Classification of Soybean Genotypes for Primary Macronutrients’ Content Using UAV–Multispectral Sensor. Remote Sens.; 2023; 15, 1457. [DOI: https://dx.doi.org/10.3390/rs15051457]
10. Osco, L.P.; Junior, J.M.; Ramos, A.P.; Furuya, D.E.; Santana, D.C.; Teodoro, L.P.; Gonçalves, W.N.; Baio, F.H.; Pistori, H.; Junior, C.A. et al. Leaf Nitrogen Concentration and Plant Height Prediction for Maize Using UAV-Based Multispectral Imagery and Machine Learning Techniques. Remote Sens.; 2020; 12, 3237. [DOI: https://dx.doi.org/10.3390/rs12193237]
11. Baio, F.H.R.; Santana, D.C.; Teodoro, L.P.R.; Oliveira, I.C.d.; Gava, R.; de Oliveira, J.L.G.; Silva Junior, C.A.d.; Teodoro, P.E.; Shiratsuchi, L.S. Maize Yield Prediction with Machine Learning, Spectral Variables and Irrigation Management. Remote Sens.; 2023; 15, 79. [DOI: https://dx.doi.org/10.3390/rs15010079]
12. Ling, B.; Goodin, D.G.; Raynor, E.J.; Joern, A. Hyperspectral Analysis of Leaf Pigments and Nutritional Elements in Tallgrass Prairie Vegetation. Front. Plant Sci.; 2019; 10, 142. [DOI: https://dx.doi.org/10.3389/fpls.2019.00142]
13. Barbedo, J.G.A. Detection of Nutrition Deficiencies in Plants Using Proximal Images and Machine Learning: A Review. Comput. Electron. Agric.; 2019; 162, pp. 482-492. [DOI: https://dx.doi.org/10.1016/j.compag.2019.04.035]
14. Teixeira, P.C.; Donagemma, G.K.; Fontana, A.; Teixeira, W.G. Manual de Métodos de Análise de Solo; Embrapa: Brasília, Brazil, 2017; 573p.
15. da Silva, E.E.; Rojo Baio, F.H.; Ribeiro Teodoro, L.P.; da Silva Junior, C.A.; Borges, R.S.; Teodoro, P.E. UAV-Multispectral and Vegetation Indices in Soybean Grain Yield Prediction Based on in Situ Observation. Remote Sens. Appl.; 2020; 18, 100318. [DOI: https://dx.doi.org/10.1016/j.rsase.2020.100318]
16. Bataglia, O.C.; Furlani, A.M.C.; Teixeira, J.P.F.; Furlani, P.R.; Gallo, J.R. Methods of Chemical Analysis of Plants. Boletim Tecnico Instituto Agronomico; 1983; 78, pp. 1-48.
17. Baio, F.H.R.; Silva, E.E.; Vrech, M.A.; Souza, F.H.Q.; Zanin, A.R.; Teodoro, P.E. Vegetation Indices to Estimate Spray Application Rates of Crop Protection Products in Corn. Agron. J.; 2018; 110, pp. 1254-1259. [DOI: https://dx.doi.org/10.2134/agronj2017.12.0718]
18. Team, R.C. R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing: Vienna, Austria, 2013.
19. Gregori, G.S.d.; de Souza Loureiro, E.; Amorim Pessoa, L.G.; Azevedo, G.B.d.; Azevedo, G.T.d.O.S.; Santana, D.C.; Oliveira, I.C.d.; Oliveira, J.L.G.d.; Teodoro, L.P.R.; Baio, F.H.R. Machine Learning in the Hyperspectral Classification of Glycaspis Brimblecombei (Hemiptera Psyllidae) Attack Severity in Eucalyptus. Remote Sens.; 2023; 15, 5657. [DOI: https://dx.doi.org/10.3390/rs15245657]
20. Santana, D.C.; de Queiroz Otone, J.D.; Baio, F.H.R.; Teodoro, L.P.R.; Alves, M.E.M.; da Silva Junior, C.A.; Teodoro, P.E. Machine Learning in the Classification of Asian Rust Severity in Soybean Using Hyperspectral Sensor. Spectrochim. Acta A Mol. Biomol. Spectrosc.; 2024; 313, 124113. [DOI: https://dx.doi.org/10.1016/j.saa.2024.124113]
21. Pereira Ribeiro Teodoro, L.; Estevão, R.; Santana, D.C.; Oliveira, I.C.d.; Lopes, M.T.G.; Azevedo, G.B.d.; Rojo Baio, F.H.; da Silva Junior, C.A.; Teodoro, P.E. Eucalyptus Species Discrimination Using Hyperspectral Sensor Data and Machine Learning. Forests; 2023; 15, 39. [DOI: https://dx.doi.org/10.3390/f15010039]
22. Scott, A.A.J.; Knott, M. A Cluster Analysis Method for Grouping Means in the Analysis of Variance Published by: International Biometric Society Stable. Biometrics; 1974; 30, pp. 507-512. [DOI: https://dx.doi.org/10.2307/2529204]
23. Mao, Z.-H.; Deng, L.; Duan, F.-Z.; Li, X.-J.; Qiao, D.-Y. Angle Effects of Vegetation Indices and the Influence on Prediction of SPAD Values in Soybean and Maize. Int. J. Appl. Earth Obs. Geoinf.; 2020; 93, 102198. [DOI: https://dx.doi.org/10.1016/j.jag.2020.102198]
24. Gitelson, A.; Solovchenko, A.; Viña, A. Foliar Absorption Coefficient Derived from Reflectance Spectra: A Gauge of the Efficiency of in Situ Light-Capture by Different Pigment Groups. J. Plant Physiol.; 2020; 254, 153277. [DOI: https://dx.doi.org/10.1016/j.jplph.2020.153277] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32979788]
25. Tomicioli, R.M.; Leal, F.T.; Coelho, A.P. Limitação Da Produtividade Pela Deficiência de Boro Nas Culturas Da Soja, Milho, Feijão e Café. South Am. Sci.; 2021; 2, e21100. [DOI: https://dx.doi.org/10.17648/sas.v2i1.100]
26. Malavolta, E. Manual de Nutrição Mineral de Plantas; Agronômica Ceres: Sao Paulo, Brazil, 2006; Volume 1.
27. Smith, H.L.; McAusland, L.; Murchie, E.H. Don’t Ignore the Green Light: Exploring Diverse Roles in Plant Processes. J. Exp. Bot.; 2017; 68, pp. 2099-2110. [DOI: https://dx.doi.org/10.1093/jxb/erx098]
28. Behmann, J.; Steinrücken, J.; Plümer, L. Detection of Early Plant Stress Responses in Hyperspectral Images. ISPRS J. Photogramm. Remote Sens.; 2014; 93, pp. 98-111. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2014.03.016]
29. Jensen, J.R. Remote Sensing of the Environment: An Earth Resource Perspective 2/e; Pearson Education India: Gujarat, India, 2009; ISBN 8131716805
30. Agyenim-Boateng, K.G.; Zhang, S.; Islam, M.S.; Gu, Y.; Li, B.; Azam, M.; Abdelghany, A.M.; Qi, J.; Ghosh, S.; Shaibu, A.S. et al. Profiling of Naturally Occurring Folates in a Diverse Soybean Germplasm by HPLC-MS/MS. Food Chem.; 2022; 384, 132520. [DOI: https://dx.doi.org/10.1016/j.foodchem.2022.132520]
31. Padalkar, G.; Mandlik, R.; Sudhakaran, S.; Vats, S.; Kumawat, S.; Kumar, V.; Kumar, V.; Rani, A.; Ratnaparkhe, M.B.; Jadhav, P. et al. Necessity and Challenges for Exploration of Nutritional Potential of Staple-Food Grade Soybean. J. Food Compos. Anal.; 2023; 117, 105093. [DOI: https://dx.doi.org/10.1016/j.jfca.2022.105093]
32. Schwalbert, R.A.; Amado, T.; Corassa, G.; Pott, L.P.; Prasad, P.V.V.; Ciampitti, I.A. Satellite-Based Soybean Yield Forecast: Integrating Machine Learning and Weather Data for Improving Crop Yield Prediction in Southern Brazil. Agric. Meteorol.; 2020; 284, 107886. [DOI: https://dx.doi.org/10.1016/j.agrformet.2019.107886]
33. Bashashati, H.; Ward, R.K.; Birch, G.E.; Bashashati, A. Comparing Different Classifiers in Sensory Motor Brain Computer Interfaces. PLoS ONE; 2015; 10, e0129435. [DOI: https://dx.doi.org/10.1371/journal.pone.0129435]
34. Ayoub Shaikh, T.; Rasool, T.; Rasheed Lone, F. Towards Leveraging the Role of Machine Learning and Artificial Intelligence in Precision Agriculture and Smart Farming. Comput. Electron. Agric.; 2022; 198, 107119. [DOI: https://dx.doi.org/10.1016/j.compag.2022.107119]
35. Breiman, L. Random Forests. Mach. Learn.; 2001; 45, pp. 5-32. [DOI: https://dx.doi.org/10.1023/A:1010933404324]
36. Belgiu, M.; Drăguţ, L. Random Forest in Remote Sensing: A Review of Applications and Future Directions. ISPRS J. Photogramm. Remote Sens.; 2016; 114, pp. 24-31. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2016.01.011]
37. Purwanto, Y.A.; Widodo, S.; Iriani, E.S. Rapid Assessment of Vanilla (Vanilla Planifolia) Quality Parameters Using Portable Near-Infrared Spectroscopy Combined with Random Forest. J. Food Compos. Anal.; 2024; 133, 106346.
38. Chen, W.; Yan, B.; Xu, A.; Mu, X.; Zhou, X.; Jiang, M.; Wang, C.; Li, R.; Huang, J.; Dong, J. An Intelligent Matching Method for the Equivalent Circuit of Electrochemical Impedance Spectroscopy Based on Random Forest. J. Mater. Sci. Technol.; 2025; 209, pp. 300-310. [DOI: https://dx.doi.org/10.1016/j.jmst.2024.05.024]
39. Zheng, B.; Xiao, T.; Wang, M.; Tian, S.; Tan, B. Soil Nitrogen Detection Based on Random Forest Algorithm and near Infrared Spectroscopy. Proceedings of the 2022 International Conference on Computation, Big-Data and Engineering (ICCBE); Yunlin, Taiwan, 27–29 May 2022; pp. 286-289.
40. Santana, D.C.; de Oliveira, I.C.; de Oliveira, J.L.G.; Baio, F.H.R.; Teodoro, L.P.R.; da Silva Junior, C.A.; Seron, A.C.C.; Ítavo, L.C.V.; Coradi, P.C.; Teodoro, P.E. High-Throughput Phenotyping Using VIS/NIR Spectroscopy in the Classification of Soybean Genotypes for Grain Yield and Industrial Traits. Spectrochim. Acta A Mol. Biomol. Spectrosc.; 2024; 310, 123963. [DOI: https://dx.doi.org/10.1016/j.saa.2024.123963] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38309004]
© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Dthenifer Cordeiro Santana 1 ; Marcelo Carvalho Minhoto Teixeira Filho 2
; de Oliveira, Izabela Cristina 1 ; Rita de Cássia Félix Alvarez 1 ; João Lucas Della-Silva 3
; Fábio Henrique Rojo Baio 1
; Gava, Ricardo 1
; Larissa Pereira Ribeiro Teodoro 1
; Carlos Antonio da Silva Junior 3
; Teodoro, Paulo Eduardo 1
1 Department of Agronomy, Federal University of Mato Grosso do Sul (UFMS), Chapadão do Sul 79560-000, MS, Brazil;
2 Department of Agronomy, State University of São Paulo (UNESP), Ilha Solteira 15385-000, SP, Brazil;
3 Department of Geography, State University of Mato Grosso (UNEMAT), Sinop 78550-000, MT, Brazil;