Can Different Cultivars of Panicum maximum Be Identified Using a VIS/NIR Sensor and Machine Learning?

Abstract

Panicum maximum cultivars have distinct characteristics, especially morphological ones related to the leaf structure and coloration, and there may be differences in the spectral behavior captured by sensors. These differences can be used in classification using machine learning (ML) algorithms to differentiate biodiversity within the same species. The objectives of this study were to identify ML models able to differentiate P. maximum cultivars and determine which is the best spectral input for these algorithms and whether reducing the sample size improves the response of the algorithms. The experiment was carried out at the experimental area of the Forage Sector of the School Farm belonging to the Federal University of Mato Grosso do Sul (UFMS). The leaf samples of the cultivars Massai, Mombaça, Tamani, Quênia, and Zuri were collected from experimental plots in the field. Analysis was carried out on 120 leaf samples from the P. maximum cultivars using a VIS/NIR hyperspectral sensor. After obtaining the spectral data and separating them into bands, the data were submitted for ML analysis to classify the cultivars based on the spectral variables. The algorithms tested were artificial neural networks (ANNs), REPTree and J48 decision trees, random forest (RF), and support vector machine (SVM). A logistic regression (LR) was used as a traditional classification method. Two input models were evaluated in the algorithms: the entire spectrum band provided by the sensor (ALL) and another input configuration using the calculated bands. The reflectances from the P. maximum cultivars showed different behavior, especially in the green and NIR regions. RL and ANN algorithms using all information in the spectrum are able to accurately classify the cultivars, reaching accuracies above 70 for CC and above 0.6 for kappa and F-score. VIS/NIR leaf reflectance can be a powerful tool for low-cost, non-destructive, and high-performance analysis to distinguish P. maximum cultivars. Here, we achieved better model accuracy using only 40 leaf samples. In the present study, the J48 decision tree model proved to have good classification performance regardless of the sample size used, which makes it a strategic model for forage cultivar classification studies in smaller or larger datasets.

1. Introduction

Panicum maximum is one of the main species of cultivated forage grasses, recognized for its high economic value in the agricultural sector, due to its ability to provide high-quality forage [1]. This species plays a fundamental role in sustaining livestock production in several tropical and subtropical regions, where it is widely cultivated. However, studying the phenotypic characteristics of these plants is challenging, given the complex behavior they exhibit in response to environmental factors. Plants have different behaviors and phenotypic expressions, which are inherently complex to study, especially when there is strong interaction with the environment. The environment influences the growth and development of forage plants, altering morphophysiological features, resulting in productive impacts when the crop is of commercial interest [2].

Over the last two decades, there have been major advances in plant breeding and in the application of technology and innovation in molecular analysis linked to plant genotyping [3]. Recently, innovations in phenotypic analysis have also been made [4,5,6,7,8,9]. Using sensors to study plant phenotypes is an effective alternative to studying plant traits. This is performed using a high degree of objectivity and precision of information with high special and temporal resolution, allowing the mapping of large areas and the selection and identification of individuals of greater interest, whether for farmers or researchers, making it possible to extract relevant information from the plant’s architecture regarding its physiology [4,5,10].

Modern hyperspectral data collection techniques have great potential to provide detailed spectral information across a wide range of wavelengths on a variety of materials. Unlike traditional multispectral imaging methods, hyperspectral imaging captures spectral signatures across hundreds of bands, allowing the detection of diverse morphophysiological information from plants subjected to this type of analysis [11]. Hyperspectral remote sensors provide leaf-level information in the visible and near-infrared (VIS-NIR) region, a sensor suggested as a fast and non-destructive method for analyzing plants [12,13]. The use of this information makes it possible to investigate leaf spectral behaviors related to physiology and to detect genotypic and phenotypic differences in plants [14]. However, due to the size of the dataset obtained and the complex relationships between plant traits and hyperspectral variables, processing data obtained using this approach requires robust analysis capable of processing non-linear data to generate accurate prediction or classification models [8,15,16].

When processing information from remote sensors, machine learning (ML) algorithms are more complex than traditional linear prediction models and can be more efficient for studying non-linear associations between the traits under study [17]. ML analysis not only allows for non-linear tasks to be solved, but is also a technology that can explore rules and patterns in large datasets that cannot be identified using conventional analyses [18]. An example of the application of ML in conjunction with data from sensors is the potential of hyperspectral sensors combined with machine learning in the detection of soil-borne plant diseases, serving as an interesting tool in the development of integrated pest management practices [19]. The use of machine learning together with spectral data can contribute to improving predictions of crop parameters and can be a way to model the estimation of the biochemical and biophysical parameters of crops [20].

Panicum maximum grass cultivars have different morphological behavior, especially related to the structure and color of the foliage, and there may be a difference in the spectral reflectance signature captured by sensors. This different behavior can be used in classification tasks using ML algorithms to differentiate biodiversity within the same species. Despite recent advances in plant phenotyping and data processing, studies on the discrimination of tropical grass cultivars using hyperspectral sensing and ML are still scarce. According to the bibliographic survey carried out, this study is the first to investigate the combined use of hyperspectral sensor and machine learning (ML) for the discrimination of Panicum maximum cultivars. As there is no research on the joint use of the techniques, this justifies the relevance and originality of the research. Therefore, the objectives here were to identify ML models able to differentiate P. maximum cultivars and to determine which is the best input spectral information for these algorithms and whether reducing the sample size improves the response of the algorithms.

2. Materials and Methods

2.1. Study Area

The field experiment was carried out at the Forage Sector of the Fazenda Escola at the Federal University of Mato Grosso do Sul (UFMS), located in the municipality of Terenos (20°26′34.31″ south and 54°50′27.86″ west), State of Mato Grosso do Sul, Brazil, at an altitude of 530.7 m. The region’s climate is classified as tropical rainy savannah, subtype Aw, with seasonal rainfall distribution [21] and the soil in the experimental area was classified as Red Dystrophic Latosolo [22], characterized by a clay texture, acid pH, low base saturation, and high aluminum concentration.

Table 1. Physical and chemical characterization of the assessment area.

Profundity	pH	P mg dm−³	MO g kg−1	K	Ca	Mg	Al	H + Al	S	T
				Cmol dm−³
0–10	6.2	12.28	32.2	0.28	6.26	3.78	0	5.02	10.32	15.34
10–20	6.2	6.9	30.52	0.21	6.04	3.32	0	5.14	9.57	14.71
20–40	5.86	2.2	22.32	0.08	5.36	1.24	0.14	5.9	6.64	12.54
40–60	5.5	0.9	14.44	0.04	2.9	0.78	1.08	6.84	3.75	10.59
60–80	5.5	0.7	10.02	0.03	2.76	0.48	1.26	6.86	3.29	10.15
80–100	5.58	0.9	8.28	0.04	3.3	0.52	1.26	6.82	3.83	10.65
Profundity			Sand (g kg−1)			Clay (g kg−1)			Silt (g kg−1)
0–10			140			716.67			143.33
10–20			120			740			140
20–40			125			755			120
40–60			116			782			102
60–80			116			788			96
80–100			92.5			790			117.5

pH in water 1:2.5; P and K = phosphorus and potassium, respectively, Mehlich extractor; S = sum of bases (Ca + Mg + K); T = CEC at pH 7.0 [S + (H + Al)]; V = % Sat. bases = 100 S/T; m = % Sat. aluminum = 100 Al (S + Al)⁻¹.

shows the physical and chemical characterization of the assessment area.

Figure 1 shows the rainfall graph of maximum and minimum temperatures in the month in which the leaves were collected. It is worth noting that, in the region, May marks the beginning of winter, when rainfall becomes scarcer.

2.2. Hyperspectral Data

Leaf samples of the cultivars Massai, Mombaça, Tamani, Kenya, and Zuri were collected from experimental plots in the field. The plots were 4 m² in size. Spectral analysis was carried out on 120 leaf samples of each cv, randomly collected throughout the experimental plot, selecting well-developed and healthy leaves, seeking to characterize the leaves of each species in the best possible way. The collections were carried out in the month of May, depending on the availability of transporting the sensor to the city where the evaluations took place, since the sensor belongs to the spectroscopy laboratory (Laspec) of the Federal University of Mato Grosso do Sul, Chapadão do Sul campus. The hyperspectral sensor Ocean Optics, model STS-VIS-L-50-400-SMA, capable of capturing reflectance in the wavelength range between 338 and 824 nm, was used to collect the spectral data. Before obtaining the reflectances of the leaf samples, the sensor was properly calibrated using a barium sulfate plate for the white calibration, as well as isolating the sensor for dark calibration. Reflectances were acquired in a laboratory with an environment controlled using artificial lighting with two 120 W halogen lamps directed at the reading point.

The sensor was connected to a computer to record each reading using sensor-specific software called OceanView 2.0, which records the readings taken by the equipment with a 1.5 nm spectral resolution (FWHM). Information corresponding to the range between 338 and 449 nm was excluded from data processing due to the presence of noise that could compromise the quality and accuracy of subsequent analyses. The 780 wavelengths were grouped into representative interval means in 16 bands that are related to important physiological processes for the plant, according to the methodology proposed by da Silva Junior et al. (2018) [7]. The methodology was adapted because the range of the sensor spectrum used by the authors was greater than that used here. These 16 bands were summarized according to their relationship to Xanthophyll, Synthesis of Chlorophyll, b-carotene, Chlorophyll b, and a-carotene. It covers the following wavelengths: B1 (450.14–474.81 nm), B2 (480 nm), B3 (481.35–499.60 nm), B4 (501–529.7 nm), B5 (531.11–539.62 nm), B6 (540 nm), B7 (541.03–648.60 nm), B8 (650 nm), B9 (661.44–669.65 nm), B10 (675 nm), B11 (676.42–683.68 nm), B12 (685.13–688.53 nm), B13 (690.47–699.68 nm), B14 (701.14–708.91 nm), B15 (710 nm), and B16 (711.35-729.85 nm).

2.3. Machine Learning and Statistical Analyses

After obtaining the spectral data and separating it into bands, the data were submitted to machine learning (ML) analysis to classify the cultivars. The algorithms tested were the Multilayer Perceptron artificial neural network (ANN), REPTree (DT) and J48 decision trees, random forest (RF), and support vector machine (SVM). A traditional logistic regression (LR) was used as the control classification model. All of the algorithms were selected according to Santana et al. (2023) [23], Pereira Ribeiro Teodoro et al. (2023) [8], and Gregori et al. (2023) [24].

The parameters used for each algorithm were the Weka 3.8.5 software defaults, except for the ANN, which used a Multilayer Perceptron with 10 neurons in the first layer and 10 neurons in the second layer in order to increase data processing efficiency. Two input models were tested in the algorithms: the entire spectrum band provided by the sensor (ALL) and using only the calculated bands (SB), in an attempt to identify the one that provides the algorithms with the highest classification accuracy. The performance of the models was evaluated according to the accuracy metrics percentage of correct classifications (CC), F-score, and Kappa coefficient, also provided by the Weka 3.8.5 software.

Afterwards, with the best spectral information input, different leaf sample sizes were tested as inputs: 40, 80, and 120 samples, following all of the parameters already mentioned.

The performance of the algorithms after processing the accuracy metrics was subjected to an analysis of variance in an attempt to identify significant differences between the ML models tested. Furthermore, the analysis of variance made it possible to assess the interaction between the inputs and the models, generating boxplots with the means of each accuracy metric using the “qgraph” and “ggplot2” packages of the R 4.1.0 software program [25]

3. Results

P. maximum cultivars showed different spectral behavior (Figure 2). It can be seen that, in the visible region (VIS, 550–700 nm), especially between wavelengths 500 and 565 nm in the green range (B), Mombaça had the highest reflectance, followed by Tamani, Massai, Kenya, and Zuri. In the red range, which is absorbed for photosynthetic processes, Mombaça, Tamani, and Massai behaved very similarly due to the overlapping of the curves. In this same range of the spectrum, there was a higher absorption of the green wavelength by Zuri and Kenya.

In the NIR region, there was a higher range of distinction between the P. maximum cultivars evaluated. Mombaça had the highest reflectance in this range. Zuri, Kenya, and Tamani had similar behavior throughout the NIR region. The lowest reflectance was observed for Massai.

Accuracy measures showed a significant interaction between ML and inputs (p-value < 0.01). The use of WL provided higher mean correct classifications for all ML algorithms (Figure 3). When the input used was SB, the RL and ANN algorithms outperformed the others (>70). When WL was used, the RL and SVM algorithms showed the best performance (76.65 and 76.97, respectively).

Regarding the F-score (Figure 4), the RL and SVM algorithms performed best using WL (0.76 and 0.75, respectively), while ANN and RF performed similarly regardless of the input. When SB was used as an input, RL and ANN performed better than the others (0.70 and 0.68, respectively). WL provided the best overall performance for RL and SVM.

WL input provided the best performance for all algorithms with regard to the Kappa coefficient (Figure 5). When comparing the accuracy of the algorithms using SB, the best results were achieved by RL and ANN (0.65 and 0.63, respectively). When the input used was WL, RL and SVM showed higher accuracy (0.71).

In the spectrum range, the green visible region (VIS) and the near-infrared (NIR) region between 700 and 800 nm were the ones that most influenced the distinction of Panicum maximum cultivars. Thus, when using all of the spectral information provided by the sensor (WL), the overall performance of all algorithms improved, especially RL and SVM, while the SB bands favored RL and ANN. These results indicate that the choice of algorithm and spectral region directly affects the classification accuracy, with WL being the input that provides the best accuracy for most algorithms.

Using WL as model inputs provided better accuracy. Different sample sizes were tested for data processing in order to optimize the analyses for classifying the Panicum maximum cultivars. Comparing the three sample sizes, it can be seen that using 40 leaf samples provided the best accuracy for the models, followed by 120 samples for the correct classification accuracy metric (Figure 6). J48 showed the best performance regardless of the sample size used (>80).

In the F-score accuracy metric, using 40 leaf samples also provided better accuracies for the algorithms (Figure 7), followed by 120 leaf samples. By evaluating the performance of each algorithm, DT and J48 had better accuracies when 40 samples were used (>0.85). When 80 samples were used, LR, DT, and J4.8 performed better (>0.75). DT and J48 also showed better results with 120 leaf samples (>0.70).

The behavior of the algorithms for the Kappa coefficient was variable, in which RL, ANN, SVM, and J48 performed better when 40 leaf samples were used (Figure 8). RF showed no difference between the three leaf sample sizes used. DT performed similarly when using 40 and 120 samples (>0.70). Using 40 leaf samples provided better performance for J48 than for the other algorithms (>0.80). This same behavior was observed for J48 when using 80 and 120 leaf samples (>0.75).

Reducing the number of samples for ML model classification can be a strategy to save resources and time in field evaluations and data processing, achieving efficient model responses with a smaller sample size. Here, we achieved better model accuracy using only 40 leaf samples, the smallest sample size tested. However, this reduction must be carried out carefully, as reducing the number of samples can compromise the quality of the model and its ability to generalize, which makes the analysis of different ML models a crucial strategy. By testing different models and sample sizes, it is possible to identify the best strategy that combines the highest classification accuracy and the smallest sample size, or even models that perform better regardless of the sample size used. In the present study, the J48 decision tree model proved to have good classification performance regardless of the sample size used, which makes it a strategic model for forage cultivar classification studies in smaller or larger datasets.

4. Discussion

Since there is different hyperspectral behavior between P. maximum cultivars, it is possible to use this information to distinguish between cultivars from the same species using machine learning (ML) data processing. Six algorithms, two inputs, and three accuracy metrics were tested to check the performance of ML techniques. WL provided better performance among the algorithms than the SB, which can be attributed to the higher spectral information provided by this input, allowing better distinction by the algorithms. The best algorithms for classification using WL were RL and SVM. Using only the bands, RL and ANN outperformed other algorithms.

P. maximum cultivars have different morphological traits and nutritional quality, even though they belong to the same species. Veras et al. 2020 [26], highlight several morphophysiological traits that distinguish P. maximum cultivars, for example, canopy height, leaf/stem ratio, forage production, tiller density, and ability to adapt to climatic adversities. Due to these morphological and physiological differences between the cultivars, they have different biochemical activities, reproducing different reflectances between the cultivars.

The Massai and Mombaça cultivars were the most differentiated in the NIR region (Figure 2D). However, in the VIS region, especially in the spectral range corresponding to blue and red (Figure 2A,C, respectively), their behavior was similar. The photosynthesizing pigments found in plant leaves, such as chlorophylls and carotenoids, significantly absorb wavelengths in the VIS region [7], which is then converted to photosynthetic activities to supply the plant.

There was a higher distinction between the cultivars in the NIR region. Similarly, previous researchers [16] found that the NIR region was more sensitive in differentiating between sorghum hybrids, a fact that corroborates the findings of this study, with the NIR region being the one that was most distinguished between P. maximum cultivars. The NIR region is directly related to biochemical activities and leaf structure, and this band of the spectrum is used in processes such as tissue elongation [27]. The Massai and Tamani cultivars showed higher tiller densities compared to the other cultivars [26] due to the compensatory mechanism and tiller size/tiller. One example is the Massai, which has a greater number of tillers, but they are smaller in size when compared to the Mombaça cultivar [26], in which this morphological structure can be attributed to the lower reflectance in the NIR region, especially when compared to Mombaça.

Besides the NIR region, the green range (Figure 2B) also showed greater distinction between the cultivars. The pigments absorb around 20 to 30% of the green wavelength, reflecting more in this spectrum than absorbing and using what is absorbed in photosynthesis, especially by supplying energy to the leaves in the lower layers of the canopy [28,29,30]. This identification of specific bands within the spectrum that contribute to the distinction of genotypes may allow for the distinction between species to be achieved more effectively, allowing for the efficient use of computational models due to better spectral differentiation [31].

Due to this distinction across the entire spectrum, especially in the green VIS and NIR bands, it is possible to submit the spectral information to machine learning algorithms so that they can recognize patterns and differentiate between P. maximum cultivars. In all of the accuracy metrics, the use of all of the VIS/NIR (WL) information resulted in the better performance of the algorithms. Meireles et al. (2020) [32] reported that in order to assess and monitor plants and their diversity, it is necessary to obtain as much information from the spectrum as possible, rather than investing only in predefined spectral bands or limited spectral regions. However, the use of SB can be feasible in cases where not all of the spectrum information is available or when there is low data processing capacity, so it can provide satisfactory insights.

RL and ANN performed better than the other algorithms when reduced spectral information (SB) was used, achieving, on average, 70 for CC, 0.70 for F-score, and 0.63 for Kappa. Gregori et al. (2023) [24] found good performance for ANN using SB to classify different levels of Glycaspis brimblecombei attack severity in eucalyptus. For an ANN model to perform more accurately, a large amount of raw data is needed for training, validation, and testing to accurately capture the non-linear relationship between the inputs and outputs used in the algorithm [33]. Khairunniza-Bejo et al. (2021) [15] state that neural networks achieve good accuracy performance even when a small amount of information is used.

Clearly, using all of the spectral information provided to the algorithms improved accuracy: 76% for CC, 0.76 for F-score, and 0.71 for Kappa, on average. The RL and SVM algorithms outperformed the other algorithms in classifying Panicum maximum cultivars. The SVM algorithm is a good classifier in supervised learning and has been used in several classification and prediction tasks in agriculture [34]. Among its applications is the selection of the best spectral range for the early identification of soybean anthracnose [35] and stress identification in plants [36].

RL performed well when both WL and SB were used. Even though SB provides summarized spectral information of the entire spectrum, both inputs highlight that the traditional method is efficient for the proposed task and that use of ML matches the efficiency of this classifier. It should be noted that the combination of SB and SVM achieved lower variation in the accuracy metrics, ensuring higher reliability for the ML algorithm.

When the different sample sizes were tested, the behavior of the algorithms was different. The J48 algorithm showed the best generation capacity because it performed well regardless of the sample size tested. Reducing the sample size for evaluation is important because it demonstrates that it is possible to discriminate eucalyptus species more quickly and accurately using hyperspectral information and machine learning [8]. However, this reduction must be carried out carefully because, with fewer samples, the model may not capture all of the variability and complexity of the task to which it has been subjected and may not generalize well to different scenarios or to samples that deviate from the patterns observed in a reduced training set.

It is expected that the more samples there are for modeling the algorithms, the more accurate, the greater the generalization capacity, and the more robust the model will be [37]. A smaller number of leaves can result in better algorithm performance due to reduced complexity and improved dataset quality. With fewer samples, machine learning models can focus more on the most relevant features, avoiding the risk of overfitting, where the model overfits to the training data and loses the ability to generalize to new data [38]. Additionally, a smaller dataset can simplify the algorithm training process by reducing noise and unnecessary variation that could confuse the model, resulting in more consistent and accurate performance.

The limitations of this study include the restriction of the analysis to only a few cultivars of Panicum maximum, which may limit the generalization of the results to other varieties or environmental conditions. Therefore, the application of these evaluations to other genera and species is essential, providing more information and data variability to train and validate the machine learning models. In terms of perspectives for future work, it is recommended that the research is expanded to include a greater diversity of cultivars and environmental conditions in order to validate and improve the conclusions obtained. The investigation of the relationships between spectral behavior and morphological and physiological characteristics should be deepened. In addition, the application of remote sensing techniques can be explored to facilitate the large-scale mapping of pastures, providing a low-cost and highly efficient tool for the management of forage resources.

5. Conclusions

Panicum maximum cultivars exhibit variations in their spectral behavior, especially in the green and near-infrared (NIR) spectrum regions. Using all available spectral information as input for machine learning algorithms, particularly Support Vector Machine (SVM) and logistic regression (RL), resulted in better discrimination between cultivars, highlighting the importance of exploring the entire spectrum to achieve accurate classification.

In addition, when reducing the number of samples, the J48 algorithm stood out with superior performance, especially when using 40 and 120 leaf samples. This result suggests that, in certain contexts, reducing the number of samples can improve the performance of certain classification algorithms. This may be particularly relevant in scenarios where the collection of large volumes of data is limited or where the aim is to optimize processing time and computational resources without compromising the accuracy of the results.