Using Machine Learning Algorithms to Investigate the Impact of Temperature Treatment and Salt Stress on Four Forage Peas (Pisum sativum var. arvense L.)

Abstract

The combination of high or low temperatures and high salt may cause significant harm to the yield, quality, and overall productivity of forage pea crops. The germination process, a crucial phase in the life cycle of forage peas, may be greatly influenced by varying temperature and salinity conditions. To comprehend the influence of these elements on the germination of forage peas, one must use many tactics, including the choice of resilient forage pea cultivars. The experiment aimed to evaluate the response of four forage pea cultivars (Arda, Ozkaynak, Taskent, and Tore) caused by various temperature (10 °C, 15 °C, and 20 °C) and salt (0, 5, 10, 15, and 20 dS m⁻¹) conditions at the germination stage using multivariate analysis and machine learning methods. An observation of statistical significance (p < 0.01) was made regarding the variations between genotypes, temperature–salt levels, and the interaction of the observed factors: germination percentage (GP), shoot length (SL), root length (RL), fresh weight (FW), and dry weight (DW). The cultivar Tore had the best values for SL (1.63 cm), RL (5.38 cm), FW (1.10 g), and DW (0.13 g) among all the cultivars. On the other hand, the Ozkaynak cultivar had the highest value for GP (89.13%). The values of all of the parameters that were investigated decreased as the salt level rose, whereas the values increased when the temperature level increased. As a result, the Tore cultivar exhibited the highest values for shoot length, root length, fresh weight, and dry weight variables when exposed to a maximum temperature of 20 °C and a saline level of 0 dS m⁻¹. It was determined that temperature treatment of fodder peas can reduce salt stress if kept at optimum levels. The effects of temperature and salt treatments on the germination data of several fodder pea cultivars were analyzed and predicted. Three distinct machine learning algorithms were used to create predictions. Based on R² (0.899), MSE (5.344), MAPE (6.953), and MAD (4.125) measures, the MARS model predicted germination power (GP) better. The GPC model performed better in predicting shoot length (R² = 0.922, MSE = 0.602, MAPE = 11.850, and MAD = 0.326) and root length (R² = 0.900, MSE = 0.719, MAPE = 12.673, and MAD = 0.554), whereas the Xgboost model performed better in estimating fresh weight (R² = 0.966, MSE = 0.130, MAPE = 11.635, and MAD = 0.090) and dry weight (R² = 0.895, MSE = 0.021, MAPE = 12.395, and MAD = 0.013). The results of the research show that the techniques and analyses used can estimate stress tolerance, susceptibility levels, and other plant parameters, making it a cost-effective and reliable way to quickly and accurately study forage peas and related species.

1. Introduction

Abiotic stress is a pervasive issue that significantly constrains agricultural output on a global scale [1]. Simultaneously, it has a detrimental impact on the growth and development of plants, thereby impairing their output [2]. Plants are susceptible to salinity at any stage of their growth cycle, beginning from the point when the seeds are deposited into the soil [3,4]. Nevertheless, plants exhibit heightened sensitivity, particularly throughout the stages of germination and seedling growth [5,6].

In the process of seed imbibition, salinity reduces the osmotic potential of the environment in comparison to the osmotic potential of the seed itself, which stops the seed from absorbing water [7]. Consequently, the rate of seed germination is decreased, and the time of seed germination is postponed. Post-germination, salinity may harm the viability of the embryo by causing an excessive buildup of Na⁺ and Cl⁻ ions [8,9]. In addition, salt stress induces the production of reactive oxygen species (ROS) and oxidative damage, leading to the disruption of many macromolecules. Hence, a decline in α-amylase activity leads to a notable drop in the transportation of sugars, thereby limiting the growth and development of the embryo [10].

The treatment of seeds with temperature may have a considerable impact on the germination and early development phases of seeds that have been subjected to salt stress [11,12]. Elevated temperatures have the potential to compound the negative consequences of salinity by further reducing the osmotic potential and increasing the evaporation rate, which ultimately results in more severe dehydration stress [13]. On the other hand, it has been discovered that certain temperature regimes may mitigate the negative consequences of salt stress [9]. An example of this would be that appropriate heat conditions have the potential to boost the enzyme activity associated with germination, such as α-amylase, which in turn promotes greater mobilization of stored reserves. Furthermore, temperature treatments have the potential to affect the production and activity of heat shock proteins (HSPs) and antioxidant enzymes. These proteins and enzymes play essential roles in reducing the oxidative damage caused by reactive oxygen species (ROS) when they are subjected to salt stress [14,15]. In order to boost seedling vigor and overall plant resistance to salt, appropriate temperature treatments may be used. These treatments work by stabilizing cellular structures and increasing processes that allow plants to tolerate stress [16].

There is a need for changes to be made to the existing systems in agricultural production because of factors such as population expansion, rising temperatures, unpredictable weather patterns, and erratic rainfall [17]. Traditional feed crops such as alfalfa and common vetch from legumes, as well as traditional forage crops like sainfoin, are farmed in Türkiye. The use of natural rangelands, stubble, and cereal straw are the primary components of sustainable livestock production [18]. As a result of the diverse climatic and soil features that exist throughout the nation, forage crops have the potential to be grown as primary and secondary crops in coastal areas as well as in central transitional regions [19]. Not only are legume forage crops notable for their many benefits, which include the fixation of nitrogen, the enhancement of soil quality, and the favorable contribution they provide to the primary crop, but they are also a source of high-quality roughage [20]. One of the most important cool-season leguminous forage species is the forage pea, also known as Pisum sativum ssp. arvense (L.) Moench. It is grown to produce grain or hay in climate zones that are classified as temperate. Furthermore, the plant has the potential to be grown as an intermediate crop in addition to the primary crop in regions with cold weather [21]. However, its variety-level reaction to temperature and salinity fluctuations, notably during germination, is unclear. Since there is little research on this topic, forage pea varieties’ germination responses to abiotic stressors must be examined. Furthermore, optimization and forecasting are particularly desired in agricultural research, and solid findings may be produced by using modern computer techniques such as machine learning algorithms that employ fewer inputs [22,23]. These approaches are advantageous because they are easily accessible and can be implemented without requiring extensive resources or specialized equipment. However, the usefulness of each model may change depending on the data used and the way the experiment is set up [24,25]. As a result, for plants grown on a large scale, like forage peas, it is crucial to anticipate how the plant will react to its surroundings during the germination period [26].

In light of this knowledge, the primary objective of this study was to analyze the growth of various forage pea cultivars under salt stress during the germination stage, with temperature treatment applied. Germination is a complex yet crucial process for plant growth and development. The secondary objective was to utilize machine learning algorithms to predict and optimize the germination parameters of fodder peas exposed to salt stress and temperature treatment.

2. Materials and Methods

2.1. Material

Arda, Ozkaynak, Taskent, and Tore forage pea cultivars were used as materials in the experiments. The seeds were obtained from the Erciyes University Faculty of Agriculture, and the experiment was set up in its laboratory. The seeds were stored dry in cloth bags at room temperature until use. The seeds were first sterilized with 5% commercial bleach (ACE) for five minutes, then rinsed with distilled water three times. NaCl (Merck) concentrations at electrical conductivities (EC) of 5, 10, 15, and 20 dS m⁻¹ were adjusted before the start of the experiment. Distilled water served as a control (0 dS m⁻¹). A total of 25 seeds for each cultivar were placed into three-layer filter papers (20 × 20 cm) irrigated with 7 mL of the respective solutions for each paper. After filter papers with seeds were rolled, they were placed into a sealed plastic bag to prevent moisture loss with 4 replications. The packages were incubated at 10 °C, 15 °C, and 20 °C in the dark. The seeds were considered germinated with the emergence of the radicle ≥ 2 mm. Root length (RL) (cm), shoot length (SL) (cm), fresh weight (FW) (g), and dry weight (DW) (g) were measured by randomly selecting 10 seedlings per replication. Germination percentage (GP), RL, SL, and FW values were recorded on the 14th day. To calculate the dry weight, the fresh shoots were dried for 72 h at 60 °C.

2.2. Analysis of Variance

Germination experiments were evaluated using analysis of variance (ANOVA) in accordance with a completely randomized trial design comprising a 4 (cultivar) × 3 (temperature) × 5 (EC) factorial arrangement with four replicates. Disparities among applications were detected utilizing the Duncan multiple comparison test, with a significance level set at 5% (Supplementary Tables S1–S4). XLSTAT (version 2023.1.3, Addinsoft, Paris, France) was utilized to assess the data.

2.3. Machine Learning (ML) Analysis

The objective was to estimate the output variables (germination percentage, shoot length, root length, fresh weight, and dry weight) by modeling the input variables (cultivars, temperature, and EC). The study included three machine learning (ML) algorithms: the multivariate adaptive regression spline (MARS) model [27], the extreme gradient boosting (XGBoost) model [28], and the Gaussian processes classifier (GPC) model [29]. These ML algorithms were chosen for their ability to handle complex and nonlinear relationships between variables, as well as their robustness in making accurate predictions based on multiple input factors, which is essential for optimizing the germination parameters under varying conditions of salt stress and temperature treatment.

The statistical notation for the MARS method, which was utilized in this study to predict output variables based on independent factors, is as follows:

$$\hat{y}={\beta }_0+{\displaystyle \sum }_{m=1}^M{\beta }_m{\displaystyle \prod }_{k=1}^{K_m}{h}_{\mathit{km}}\left(X_{v\left(k,m\right)}\right)$$

(1)

The mathematical expression h_km(X_v(k,m)) denotes the basis function, where ${\beta }_0$ represents a constant (intercept) and denotes the coefficient of basis functions. The v(k,m) is an index of the independent variable in the mth component of the kth product. K_m is the parameter that restricts the order of interaction.

Utilizing a gradient boosting framework, the XGBoost model is an algorithm of the gradient boosting decision tree variety that has been optimized for speed and performance. One notable benefit of XGBoost is its ability to learn from mistakes and reduce error rates through iterative processes. The mathematical formula of the XGBoost model is as follows:

$${\widehat{y}}_i={{\displaystyle \sum }}_{k=1}^K{f}_k\left(x_i\right)$$

(2)

where ${\widehat{y}}_i$ represents the value that the model predicts for the ith observation. The K denotes the number of decision trees in total. The $f_k$ is the kth decision tree. The feature vector $x_i$ corresponds to the ith observation.

The random variable distribution is described by the Gaussian probability density function. The statistical representation of the Gaussian process method used in the study is as follows:

$$f\left(x\right)~GP\left(m\right.\left(x\right),k\left(x\right.,x^{\prime }))$$

(3)

The symbol m(x) represents the mean function of the Gaussian process, while k(x,x′) signifies the kernel function. Similarities between data points are identified by the covariance function (kernel). A constant value, such as zero or the mean of the training dataset, is generally assigned to the mean function. The mean of the training dataset was utilized in this investigation.

The evaluation of the effectiveness of the algorithm incorporated the use of four essential metrics: mean absolute deviation (MAD), mean absolute percentage error (MAPE), root mean square error (RMSE), and R-squared (R²). Denoted as R², the coefficient of determination measures the extent to which the model accurately reproduces the empirical data (Equation (4)). RMSE, represented by Equation (5), measures the degree of proximity that exists between the predicted and actual values. One of the beneficial characteristics of MAPE is its independence from scale, which enables the representation of changes in percentage form (Equation (6)). In addition, the root mean absolute deviation (MAD), which is denoted by Equation (7), establishes the comprehensive pattern of errors in predictions [27,30]. The R package caret was utilized to randomly partition the dataset into two sets: 30% for the testing set and 70% for the training set. As shown in Equation (8), the optimal hyperparameters for each machine learning model were determined utilizing the Grid Search Cross-Validation (GCV) technique [31]. Additionally, standard deviations were calculated (Equation (9)). The training/testing sample size in the dataset is denoted by n; the measured real value is $y_i$; the predicted value is $y_{ip}$; and the mean of the measured values is $\overline{y}$. The penalty function for the complexity of the model encompassing λ terms is denoted as M(λ). The s_m is the standard deviation of the model errors, and the s_d is the standard deviation of outputs. The computation of performance metrics and machine learning algorithms was conducted utilizing the R (version 4.3.2) programming language [27,32].

$${\mathrm{R}}^2=1-\left(\frac{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-y_{ip}\right)}^2}{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-\overline{y}\right)}^2}\right)$$

(4)

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(y_i-y_{ip}\right)}^2}$$

(5)

$$\mathrm{MAPE}=\frac{1}{n} {\displaystyle \sum }_{i=1}^n\left|\frac{y_i-y_{ip}}{y_i}\right|\times 100$$

(6)

$$\mathrm{MAD}=\frac{1}{n} {\displaystyle \sum }_{i=1}^n\left|y_i-y_{ip}\right|$$

(7)

$$\mathrm{GCV}\left(\lambda \right)=\frac{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-y_{\mathit{ip}}\right)}^2}{{\left[1-\frac{M\left(\lambda \right)}{n}\right]}^2}$$

(8)

$${\mathrm{SD}}_{\mathrm{ratio}}=\frac{S_m}{S_d}$$

(9)

3. Results

The variance analysis findings for the main factors of the research are shown in

Table 1. Results of variance analysis for the effects of the main factors.

	Germination Percentage (%)	Shoot Length (cm)	Root Length (cm)	Fresh Weight (g)	Dry Weight (g)
Cultivar
Arda	70.13 ± 11.79 b *	1.14 ± 1.40 b	4.89 ± 2.22 bc	1.03 ± 0.73 b	0.12 ± 0.06 b
Ozkaynak	89.13 ± 9.20 a	1.62 ± 2.09 a	4.69 ± 2.39 c	0.83 ± 0.64 d	0.11 ± 0.06 d
Taskent	68.53 ± 12.10 bc	1.25 ± 1.75 b	5.07 ± 2.54 b	0.97 ± 0.66 c	0.12 ± 0.05 c
Tore	67.80 ± 14.86 c	1.63 ± 2.37 a	5.38 ± 2.67 a	1.10 ± 0.87 a	0.13 ± 0.07 a
Temperature (°C)
10	62.95 ± 15.04 c *	0.10 ± 0.17 c	2.93 ± 1.30 c	0.43 ± 0.20 c	0.07 ± 0.02 c
15	76.75 ± 12.01 b	1.38 ± 1.26 b	5.46 ± 2.08 b	1.06 ± 0.56 b	0.14 ± 0.05 b
20	82.00 ± 10.55 a	2.75 ± 2.49 a	6.63 ± 2.24 a	1.45 ± 0.86 a	0.15 ± 0.06 a
EC (dS m−1)
0	82.92 ± 11.57 a *	3.54 ± 3.04 a	7.31 ± 2.51 a	1.84 ± 1.01 a	0.19 ± 0.08 a
5	76.58 ± 12.96 b	1.58 ± 1.16 b	5.90 ± 2.02 b	1.11 ± 0.46 b	0.13 ± 0.04 b
10	74.92 ± 12.61 b	1.27 ± 1.04 c	5.29 ± 1.75 c	0.93 ± 0.41 c	0.12 ± 0.04 c
15	72.25 ± 14.01 c	0.51 ± 0.55 d	4.10 ± 1.39 d	0.64 ± 0.26 d	0.09 ± 0.03 d
20	62.83 ± 16.21 d	0.16 ± 0.31 e	2.43 ± 1.22 e	0.38 ± 0.16 e	0.06 ± 0.02 e
Cultivar	6245.066 ***	3.882 ***	5.149 ***	0.765 ***	0.006 ***
Temperature	7745.400 ***	140.465 ***	287.179 ***	21.107 ***	0.180 ***
EC	2576.733 ***	83.294 ***	163.564 ***	14.668 ***	0.110 ***

* Letters of the same size and notation represent significant findings; ***: significant at p ≤ 0.001.

. The analysis findings indicated statistically significant (p < 0.001) effects of cultivar, temperature, and EC on the germination percentage (GP), shoot length (SL), root length (RL), fresh weight (FW), and dry weight (DW). Among the observed variables, cultivar Tore had higher values for SL, RL, FW, and DW, whereas cultivar Ozkaynak had the highest value for the GP variable. There was a positive correlation between the temperature and the values of the observed variables. As the temperature increased, the values of the variables also increased. The highest values for all variables were seen at the maximum temperature treatment (20 °C). There was a negative correlation between the EC and the values of the observed variables. With the rise in EC, there was a corresponding decrease in the values of the variables. The highest values were determined for all parameters observed in conditions without salt stress.

Figure 1 displays the results of the analysis for the interaction between cultivar and temperature, cultivar and EC, and temperature and EC in the investigation. The research revealed a statistically significant (p < 0.001) influence of the interaction between cultivar and temperature on the GP, SL, RL, FW, and DW variables. Among the cultivars that sprouted under the highest temperature (20 °C) conditions, it was found that the Ozkaynak cultivar had the highest percentage of germination. The Tore cultivar exhibited the greatest root length as well as the highest values for both root length and fresh weight. On the other hand, the Arda cultivar had the highest value in terms of dry weight.

The investigation revealed statistically significant effects of cultivar and EC interaction on shoot length (p < 0.001), root length (p < 0.05), fresh weight (p < 0.001), and dry weight (p < 0.001). The Tore cultivar had the greatest values for SL, RL, FW, and DW observed variables. These values were seen in the absence of salt stress.

The findings of the study indicated that the interaction between temperature and EC had a statistically significant impact (p < 0.001) on the GP, SL, RL, FW, and DW variables. The maximum values for all observed variables were obtained when temperature (20 °C) circumstances reached their peak and salinity conditions were at 0 EC.

The findings of the analysis, examining the interaction among cultivar, temperature, and EC, are shown in Figure 2, Figure 3 and Figure 4. The investigation revealed statistically significant effects of cultivar, temperature, and EC interaction on shoot length (p < 0.001), root length (p < 0.01), fresh weight (p < 0.001), and dry weight (p < 0.001). The Tore cultivar had the greatest values for the observed shoot length, root length, fresh weight, and dry weight variables under conditions of maximum temperature (20 °C) and a saline level of 0 EC.

ML Analysis

The effects of varying salinity and temperature conditions on forage pea were predicted by modeling the variables GP, SL, RL, FW, and DW using three distinct machine learning algorithms (MARS, Xgboost, and GPC). The modeling technique included four distinct fodder pea cultivars (Arda, Ozkaynak, Taskent, and Tore), three varying temperature (10 °C, 15 °C, and 20 °C) settings, and five varied salinity (0 EC, 5 EC, 10 EC, 15 EC, and 20 EC) levels as input factors. The observed variables (GP, SL, RL, FW, and DW) were used as output parameters.

Table 2. Algorithm goodness-of-fit criterions for prediction of variables.

Variables	ML Criterion	MARS		XGboost		GPC
		Train	Test	Train	Test	Train	Test
GP	R2	0.863	0.899	0.926	0.823	0.916	0.849
	RMSE	5.530	5.344	3.990	6.560	4.263	6.083
	MAPE	6.515	6.953	4.360	7.541	4.908	7.675
	MAD	4.176	4.125	2.948	4.533	3.308	4.699
	SD ratio	0.370	0.317	0.926	0.420	0.291	0.389
SL	R2	0.891	0.869	0.980	0.897	0.974	0.922
	RMSE	0.641	0.781	0.264	0.690	0.298	0.602
	MAPE	20.896	15.077	13.319	12.970	13.438	11.850
	MAD	0.364	0.427	0.155	0.383	0.167	0.326
	SD ratio	0.330	0.357	0.143	0.319	0.162	0.280
RL	R2	0.854	0.891	0.933	0.875	0.926	0.900
	RMSE	0.939	0.791	0.657	0.805	0.689	0.719
	MAPE	20.276	15.429	12.643	13.333	13.947	12.673
	MAD	0.653	0.615	0.401	0.560	0.433	0.554
	SD ratio	0.382	0.328	0.259	0.353	0.272	0.315
FW	R2	0.880	0.903	0.973	0.966	0.970	0.956
	RMSE	0.255	0.273	0.123	0.130	0.129	0.149
	MAPE	29.992	32.514	9.694	11.635	9.719	12.947
	MAD	0.186	0.196	0.078	0.090	0.082	0.107
	SD ratio	0.346	0.306	0.165	0.183	0.173	0.206
DW	R2	0.828	0.827	0.960	0.895	0.957	0.896
	RMSE	0.027	0.027	0.013	0.021	0.014	0.021
	MAPE	22.138	21.838	8.294	12.395	9.244	14.362
	MAD	0.020	0.020	0.009	0.013	0.010	0.014
	SD ratio	0.415	0.416	0.200	0.324	0.208	0.321

Germination percentage (GP), root length (RL), shoot length (SL), fresh weight (FW), dry weight (DW), mean absolute deviation (MAD), mean absolute percentage error (MAPE), root mean square error (RMSE), R-squared (R²), standard deviation (SD), multivariate adaptive regression spline (MARS), extreme gradient boosting (XGBoost), and Gaussian processes classifier (GPC).

provides a thorough summary of the study’s results and displays the results produced by the machine learning models used in the research. Metrics such as the root mean squared error (MSE), mean absolute percentage error (MAPE), and mean absolute deviation (MAD) are often used to assess the overall performance of an algorithm. A decrease in the values of these measures indicates that the model’s predictions are becoming more accurate and aligning with the observed values. Moreover, it evaluates the degree to which the R-squared (R²) model can explain the variation seen between the independent components and the dependent variable being studied.

A grid search cross-validation (GCV) technique was used to assess the performance of the MARS, XGBoost, and GPC models throughout the evaluation process. The XGBoost model had the lowest MSE, MAPE, and MAD values, which indicates that it has greater prediction accuracy for training data of all observable parameters. For the variables GP, SL, RL, FW, and DW, the root mean square error (MSE) values were determined to be 3.990, 0.264, 0.657, 0.123, and 0.013, respectively. According to the findings, the mean absolute percentage error (MAPE) values for the same variables were found to be 4.360, 13.319, 12.643, 9.694, and 8.294, respectively. In addition, the values of the root mean absolute deviation (MAD) for the same variables were examined and found to be 2.948, 0.155, 0.401, 0.078, and 0.009, respectively. Furthermore, the model displayed the greatest R² coefficient when generating predictions for the variables (GP, SL, RL, FW, and DW) using the training dataset (0.926, 0.980, 0.933, 0.973, and 0.960, respectively). When compared to other models, the XGBoost model was shown to have the highest level of performance for the training dataset (Table 2).

Each ML model underwent assessment by generating predictions using the test dataset (Table 2). This aided in predicting the potential performance of the model on unfamiliar data. The MARS model exhibited higher predictive ability for germination power (GP) based on the examination of the R², MSE, MAPE, and MAD metrics. Moreover, when assessing the criterion metrics, it was determined that the GPC model had superior performance in predicting shoot length (SL) and root length (RL), while the Xgboost model showed stronger performance in estimating fresh weight (FW) and dry weight (DW). Figure 5 displays a linear regression graph illustrating the models’ predicted values, which offer the most precise prediction for the variables being studied, along with the observed actual values.

The MARS analysis identified the most effective models, which consisted of 11 terms for GP, 10 terms for SL, 8 terms for RL, 11 terms for FW, and 11 terms for DW (

Table 3. Outcome of the MARS algorithm for the observed variables.

	GP		SL		RL		FW		DW
Terms	Model	Coefficient	Model	Coefficient	Model	Coefficient	Model	Coefficient	Model	Coefficient
1	Intercept	68.800	Intercept	0.103	Intercept	5.107	Intercept	0.963	Intercept	0.134
2	Ozkan	19.800	15 °C	2.557	Tore	0.496	15 °C	0.633	Ozkan	−0.014
3	Tore	−5.600	20 °C	6.962	15 °C	2.533	20 °C	1.988	15 °C	0.076
4	15 °C	13.800	15 °C × 10 EC	−1.468	20 °C	3.707	5 EC	−0.346	20 °C	0.099
5	20 °C	14.850	15 °C × 15 EC	−2.335	5 EC	−1.410	10 EC	−0.560	5 EC	−0.061
6	5 EC	−6.333	15 °C × 20 EC	−2.611	10 EC	−2.025	15 EC	−0.781	10 EC	−0.070
7	10 EC	−8.000	20 °C × 5 EC	−4.464	15 EC	−3.213	20 EC	−0.980	15 EC	−0.096
8	15 EC	−10.667	20 °C × 10 EC	−4.597	20 EC	−4.877	20 °C × 5 EC	−1.159	20 EC	−0.097
9	20 EC	−22.000	20 °C × 15 EC	−5.876			20 °C × 10 EC	−1.034	Tore × 15 °C	0.026
10	Tore × 20 °C	12.200	20 °C × 20 EC	−6.628			20 °C × 15 EC	−1.237	15 °C × 20 EC	−0.042
11	20 °C × 20 EC	5.750					20 °C × 20 EC	−1.422	20 °C ×20 EC	−0.054

Germination percentage (GP), root length (RL), shoot length (SL), fresh weight (FW), and dry weight (DW).

). These models included intercept terms without any interaction effects. Upon analyzing the ninth term with the lowest coefficient and the tenth term with the highest coefficient for GP, it was anticipated that the 20 EC treatment would result in the greatest loss, while the Ozkan cultivar would result in the greatest rise. When examining the tenth term with the lowest coefficient and the third term with the highest coefficient for shoot length, it was predicted that the 20 °C × 20 EC interaction would lead to the most significant decrease, while the 20 °C treatment alone would result in the most significant increase. When examining the eighth term with the smallest coefficient and the fourth term with the largest coefficient for root length, it was predicted that the 20 EC treatment would lead to the most significant decrease, while the 20 °C treatment would result in the most significant increase. By examining the eleventh term with the lowest coefficient and the third term with the highest coefficient for fresh weight, it was predicted that the 20 °C × 20 EC interaction would lead to the most significant decrease, while the 20 °C treatment alone would result in the most significant increase. When examining the eighth term with the smallest coefficient and the fourth term with the largest coefficient for fresh weight, it was predicted that the 20 °C × 5 EC interaction would lead to the most significant decrease, while the 20 °C treatment alone would lead to the most significant increase.

Figure 6 presents the variable importance rankings for predicting the output variable using input variables in the XGBoost and GPC models. The analysis revealed that the most effective factors determining the germination percentage, shoot length, root length, fresh weight, and dry weight in the XGBoost model were cultivar Ozkan at 20 °C, 20 EC, 20 °C, and 15 °C, respectively. Within the GPC model, the variable that had the greatest impact on dry weight was 20 °C, although the outcomes for other factors had similarities with those of the XGBoost model.

Using the predictive capabilities of machine learning algorithms, expected germination parameters between 0 and 20 °C were predicted to optimize germination parameters under changing salt stress and temperature treatment conditions and to enable more effective agricultural practices (Supplementary Table S5). The aim was to predict the expected values of the temperature process on unobserved data based on the findings. According to the results, it was predicted that the application of machine learning algorithms might result in the failure of fodder pea plants to germinate at temperatures of 9 °C or lower for the SL parameter, 1 °C or lower for the RL parameter, 5 °C or lower for the FW parameter, and 1 °C or lower for the DW parameter.

4. Discussion

The current research demonstrates that salt has a negative impact on the germination and development of fodder pea cultivar seeds, as shown in Table 1. The germination and seedling phases are crucial for the survival and successful development of plants, especially under challenging conditions. The results of this investigation revealed that the germination and establishment of fodder pea seeds were progressively hindered as the salt stress increased. Several studies have found that higher salinity levels have a negative impact on germination percentage and speed in field pea [33], chickpea [34], wheat [35], and other legumes [36,37,38]. There are a few reasons that contribute to salinity’s ability to delay and impede seed germination. These variables include a reduction in water intake, alterations in the mobilization of stored food, and disruption of the structural organization of proteins [39]. The findings of this study were in line with studies that had been discussed and published in the past. It is possible that the osmotic impact of salt stress, which results in a reduction in growth promoters, is responsible for the shortening of growth characteristics [8]. There is also the possibility that the growth suppression is caused by a lack of water, ion toxicity, and nutritional imbalance because of the obstruction of other nutrients, including nitrogen, phosphorus, potassium, calcium, and nitrogen oxides [40].

As demonstrated in Table 1, the fodder pea cultivars that were treated to germination at temperatures of 10, 15, and 20 °C had the greatest germination values at the 20 °C temperature. Moreover, it was discovered that the Ozkaynak cultivar had the greatest percentage of germination among the cultivars that sprouted under the circumstances of the maximum temperature (20 °C). High ambient temperatures pose a substantial threat to agricultural productivity worldwide throughout plant growth and germination phases [38]. Additionally, plants exposed to low temperatures within the temperature range of 0 to 15 °C face challenges in productivity due to osmotic and oxidative stress, which necessitates the activation of various stress-related signaling pathways to restore cellular balance [41,42]. For plants that are susceptible to heat limitations, germination may provide early research for determining the threshold tolerance of the plant. Çaçan et al. [43] investigated the germination abilities of several fodder pea lines and kinds under varied temperature conditions (10, 23, 27, and 30 °C). According to their study, the rate of germination typically fell at a temperature of 30 °C, but the time it took for germination to occur was reduced at a temperature of 10 °C. Sivritepe [44] declared that the optimal temperature for conducting germination experiments on Pisum sativum seedlings is 20 °C. According to Brar et al. [45], the maximum germination rate of Pisum arvense seeds was seen at a temperature of 20 °C. The results of our investigation were consistent with those of earlier research.

The results of our research shed light on the significance of using machine learning (ML) strategies, in particular ML algorithms like XGBoost, MARS, and GPC, to enhance the comprehension and optimization of germination in forage pea when it is subjected to salt stress and temperature treatment. The outcomes of our study indicated that these computational methods are useful in predicting and improving important factors associated with germination. In the process of analyzing the prediction performances and R² of the models, it was found that the model performances were comparable to those of earlier research that has been published in the literature. Benlioğlu et al. [23] effectively used machine learning methods to clarify the germination traits of tetraploid wheat in the presence of drought-induced stress. In their investigation, they reported that the elastic net (ELNET) model showed more efficiency in predicting germination speed (R² = 0.762, RMSE = 4.873, MAD = 3.755), germination power (R² = 0.514, RMSE = 4.946, MAD = 3.581), and water absorption capacity (R² = 0.902, RMSE = 5.140, MAD = 2.492), the Gaussian processes classifier (GPC) model showed superior performance in predicting root length (R² = 0.981, RMSE = 0.552, MAD = 0.407), and the XGBoost model showed better prediction performance for shoot length (R² = 0.962, RMSE = 0.761, MAD = 0.457), fresh weight (R² = 0.962, RMSE = 0.077, MAD = 0.056), and dry weight (R² = 0.944, RMSE = 0.018, MAD = 0.014). In their investigation, Aasim et al. [46] examined the potential impact of different concentrations of hydrogen peroxide (H₂O₂) on the germination and morphological attributes of in vitro-grown cannabis seedlings. The researchers employed four distinct machine learning algorithms to analyze the data: random forest, extreme gradient boosting, support vector classifier, and Gaussian process. The results of the research demonstrated that the RF model demonstrated a higher level of accuracy in forecasting the output variable. Machine learning models, in contrast to more conventional approaches, provide a strategy that is both more effective and more precise for predicting the outcomes of complex and nonlinear biological processes in the fields of agriculture and applied sciences [25,47,48]. All the germination parameters that were observed in our research showed that the XGboost model had the greatest performance in terms of prediction. Previous research has also indicated that the XGBoost model, which is one of the machine learning algorithms, has higher prediction performance when it comes to optimizing tissue culture parameters and forecasting the outcomes of regeneration under tissue culture circumstances [25,30,49]. Chen and Guestrin [28] developed the XGBoost algorithm, a powerful tool for addressing regression and classification problems. It is a member of the gradient boosting decision tree category and is well-known for its exceptional performance and speed [28]. Within a gradient boosting architecture, XGBoost excels at leveraging errors to iteratively reduce the error rate over several iterations.

The efficiency of an algorithm might fluctuate based on the attributes of the dataset and the complexity of the issue being addressed [25]. Within this particular context, some algorithms may be better suited for certain sorts of data and issues, while others may be more efficient for distinct data formats and difficulties. For instance, while certain methods may perform well in datasets with many dimensions, algorithms that have a tendency to overfit may not be the best choice for tiny datasets. In order to obtain the best possible performance, it is often essential to experiment with various methods and carefully scrutinize model assessment criteria [27,30,48,49].

Based on the findings and unobserved data, the expected estimated values of the temperature process were determined. According to the results, it was predicted that ML algorithms at 9 degrees or less according to the SL parameter, 1 degree or less according to the RL parameter, 5 degrees or less according to the FW parameter, and 1 degree or less according to the DW parameter may result in the failure of fodder pea plants to germinate. When the temperature drops below the ideal range for germination, fodder pea plants either do not germinate at all or have a very poor germination rate [3,45].

5. Conclusions

Overall, the study’s results indicate that when temperature (from 10 °C to 20 °C) conditions rise, the cultivars become less sensitive to salt stress in terms of germination characteristics such as shoot length, root length, fresh weight, and dry weight. When compared to the cultivars Arda, Ozkaynak, and Taskent, the Tore demonstrated a higher tolerance to salt stress. In addition, the results of this research demonstrate that making it possible for forage pea cultivars to overcome the constraints that are imposed by salt stress and temperature treatment is essential, as is the ability to identify the abiotic stress conditions that are associated with the plant species and variety. In addition to these achievements, the incorporation of machine learning (ML) methods into the process of optimization and prediction, as well as the prediction of germination for cultivars such as fodder peas based on environmental circumstances, may positively boost efficiency in terms of both time allocation and workload. From this point, XGBoost, MARS, and GPC are examples of machine learning models that have successfully predicted crucial parameters in procedures that include complicated biological processes. As more research is conducted to investigate the interaction between machine learning and seed germination, there is a possibility that unique insights and strategies could be discovered. These discoveries might be used to enhance the sustainability and efficiency of plant propagation methods.