Investigation of factors affecting fresh herbage yield in pea (Pisum arvense L.) using data mining algorithms

This study was carried out to determine the factors affecting the wet grass yield of pea plants grown in Turkey. Wet grass yield was predicted using parameters such as genotype, crude protein, crude ash, acid detergent fiber (ADF), and neutral detergent fiber (NDF) with some data mining algorithms. These techniques provided easily interpretable data trees and precise cutoff values. This led to a comparison of the predictive abilities of data mining methods, including multivariate adaptive regression spline (MARS), Chi-square automatic interaction detection (CHAID), classification and regression tree (CART), and artificial neural network (ANN). To test the compatibility of the data mining algorithms, seven goodness-of-fit criteria were used. The predictive abilities of the fitted models were assessed using model fit statistics such as the coefficient of determination (R²), adjusted R², root mean square error (RMSE), mean absolute percentage error (MAPE), standard deviation ratio (SD ratio), Akaike information criterion (AIC), and corrected Akaike information criterion (AICc). With the greatest R² and adjusted R² values (0.998 and 0.986) and the lowest values of RMSE, MAPE, SD ratio, AIC, and AICc (10.499, 0.7365, 0.047, 268, and 688, respectively), the MARS method was determined to be the best model for quantifying plant fresh herbage yield. In estimating the fresh herbage production of the pea plant, the results showed that the MARS method was the most appropriate model and a good substitute for other data mining techniques.

1 Introduction

Pea (Pisum sativum L.), is an indigenous plant throughout southwest Asia and was among the earliest crops that people farmed, with wild varieties still found in Ethiopia, Afghanistan, and Iran (Maria et al., 2009). According to Berhane et al. (2016), legumes such as peas are essential for crop rotations because they contribute to the breakdown of disease and pest cycles, supply nitrogen, enhance soil microbial activity and multiplicity, improve soil composition, conserve soil water, and provide economic variety. Peas are a cool-season annual crop that fixes nitrogen and has a high ratio of edible protein (23%–33%), along with other biomolecules such as vitamins and carbohydrates (Hafiz et al., 2014).

According to Borreani et al. (2009), field pea, faba bean, and white lupin can all be effectively ensiled with the addition of a lactic acid bacteria inoculum and after a brief wilting period in favorable weather. However, white lupin can only be effectively ensiled with the application of a lactic acid bacteria inoculant due to the low dry matter content at cutting and the quick wilting phase, resulting in a very low dry matter content of wilted and unwilted silages.

When harvested as ensiled feed, legume pulses, including field peas, faba beans, and lupins, are annual crops that are well-suited for brief crop rotations (Borreani et al., 2007). Ensiling pulses as a whole-crop forage provides livestock with less expensive, traceable, and nonanimal-based home-grown protein and starch (Cavallarin et al., 2007). This can also increase the efficiency of the production system in dairy farms by reducing the amount of purchased concentrates fed to the animals (Adesogan et al., 2004).

Forage pea (Pisum arvense L.), a highly nutritious and palatable annual legume forage plant, rich in protein within its seeds. After crushing, it can be mixed with roughage. All pea varieties grown in Europe today have flowers in white, green, or yellow colors. Seeds of varieties known in the feed industry in almost all of Europe are evaluated as protein-based feed. If the dried grass of pea is harvested at the appropriate time, it contains about 20% crude protein. Similarly, its seeds contain 20%–30% crude protein, making them a high-quality, nutritious protein source for animals. Peas are used both as dry hay and green seeds for feed, and are valued as a green forage plant in pastures and as green manure to increase nitrogen levels (Özkaynak, 1980; Açıkgöz, 2001).

The aim of this study was to determine the factors affecting green grass yield in pea plants and to predict yield using data on crude ash, crude protein, neutral detergent fiber (NDF), and acid detergent fiber (ADF).

2 Materials and methods

2.1 Experimental materials

The study was conducted on 14 different pea lines and varieties at the Bingöl University Research and Application Field, located 10 km from the Bingöl city center.

The long-term average temperature in Bingöl province is 12.0°C, whereas in 2015, it was 13.7°C. Similarly, the long-term average maximum temperature is 18.4°C, compared to 19.8°C in 2015. The long-term average minimum temperature is 6.4°C, while in 2015 it was 7.2°C. Annual precipitation also showed a decrease, with a long-term average of 933.9 mm, compared to and 801.8 mm in 2015. These values indicate that 2015 was both warmer and received less rainfall than the long-term averages.

Soil analyses were conducted in the soil analysis laboratory of the Department of Soil Science and Plant Nutrition, Faculty of Agriculture, Bingöl University. The results indicate that the soil has a clayey texture, low organic matter, low salinity, basic pH, deficiency in calcium and potassium, and sufficient phosphorus content.

A field experiment was established in 2015 on a field that had been deep-plowed and tilled with a cultivator and harrow. The experiment followed a randomized block design with three replications. Plot sizes were 5 m in length with 30 cm row spacing, and each plot contained four rows. A seeding rate of 15 kg/ha was used. After planting, sprinkler irrigation was applied to ensure emergence, and weed control was conducted manually throughout the growing season using hand hoeing.

It can be generally accepted that the dependent variable, the wet grass yield of pea plants, is influenced by the genotype predictor variables (crude protein, crude ash, ADF, and NDF).

Crude protein values are shown for each feed. To calculate crude protein, multiply the Kjeldahl nitrogen by either 6.25 or 100/16. On average, proteins contain 16% nitrogen. Crude protein provides little insight into a feed’s true protein and nonprotein composition. Many feed composition charts include digestible protein, but it is more deceptive than crude protein due to the significant contribution of body protein to the apparent protein in feces (Stanton and LeValley, 2006). Crude ash is a proximate chemical composition, similar to crude protein.

Animal digestibility is closely linked to ADF. The availability of net energy from digestible energy and voluntary feed intake are associated with NDF. Both metrics have a stronger correlation with expected animal performance (Stanton and LeValley, 2006).

2.2 Statistical methods

2.2.1 Chi-squared automatic interaction detection

For paired-variable assessment, the Chi-squared automatic interaction detection (CHAID) approach may reveal a more trustworthy representation of the unmasked link than either the scatterplot or the smoothed scatterplot. Due to its simplicity in construction, comprehension, and application, CHAID regression tree models are a well-liked approach, particularly among aspiring regression modelers lacking substantial statistical expertise. The foundations of CHAID are also quite appealing: it is an assumption-free technique (i.e., it does not require formal theoretical assumptions to be satisfied) and it is very effective at managing a large number of predictor variables in “big data”. Traditional regression models, on the other hand, are assumption-full, which leaves them vulnerable to unpredictable outcomes and ineffective in handling a large number of predictor variables (Ratner, 2017). Gallagher et al. (2005) state that the CHAID approach is based solely on the classification of categorical dependent variables and uses the Chi-square test to identify categorical independent variables. A representation of the Chi-square test of independence is as follows:

$${\chi }^2=\sum _{ij}\frac{{(O_{ij}-E_{ij})}^2}{E_{ij}}$$

where $\text{i}=1,2,\dots ,\text{r and j}=1,2,\dots ,$

$\text{c}{O}_{ij}$: reflects the cell’s observed frequency.

$E_{ij}$: reflects the cell’s expected frequency.

The CHAID approach consists of three steps: merging, splitting, and stopping (Alkhasawneh et al., 2014).

Continuous variables are converted into ordinal variables before the following algorithm is activated. The mapping of a given x into category C(x) is as follows for a given set of break points $a_1,a_2,\dots ,a_{K-1}$ an (in ascending order):

$$C(x)=\{\begin{array}{cc}1& x\le a_1\\ k+1& a_k<x<a_{k+1},k=1,\dots K-2\\ K& a_{K-1}<x\end{array}$$

When estimating the ranks, if K is the desired number of bins, $x_i$ frequency weights are taken into account for the computation of the break points. The average rank is used if there are ties. The following is an ascending order of the rank and accompanying values: ${\left\{r_{(i)},x_{(i)}\right\}}_{i=1}^n$

For k = 0 to (K−1), set $I_k=\left\{i:[r_{(i)}\frac{K}{N_{f}1}]=k\right\}$ where (x) displays the floor integer of x. If $I_k$ is not empty. $i_k=max\left\{i:i\in I_k\right\}$.

In order to exclude the largest, the break points are made equal to the x values corresponding to the $i_k$ (Breiman et al., 1984; Orhan et al., 2016; Gözüaçık et al., 2018).

2.2.2 Classification and regression tree

Classification and regression tree (CART) is a rule-based, nonparametric machine learning technique that looks for relationships inferred from input characteristics (predictor variables) to target attributes. To improve the accuracy of the target variable prediction, the predictor variable is divided into many areas using this approach (Breiman et al., 1984; Steingberg and Colla, 2016).

Numerous fields, such as agricultural and veterinary sciences, extensive use it (Cak et al., 2013; Eyduran et al., 2013; Çelik and Yilmaz, 2018; Çelik et al., 2018). By locating the primary patterns within the collection of independent variables, the CART technique can be categorize and forecast the values of a specific dependent variable, Y. The dependent variable in binary classification problems is binary-valued, while in regression problems, it is continuous or interval-type. The independent variables may be continuous, ordinal, or nominal in nature. Recursive partitioning, the methodical process of building a binary decision tree by dividing each node into two child nodes or not, is the foundation of the CART (Yordanova et al., 2015).

In a regression problem, the mean value of all cases in each terminal node of the decision tree constitutes the projected value. The mean squared error from all variables and threshold values is minimized at each step by determining one independent variable and its suitable threshold value, such as ${\theta }_k$. For all circumstances where there are two viable answers, “yes” or “no”, the splitting rule has the form $x_{ki}<{\theta }_k$. In this manner, the independent variable input space is divided into multidimensional, nonoverlapping 2D rectangular or hypercuboid sections. A decision tree is a flow diagram that shows the dependent variable’s categorization and regression prediction models (Yordanova et al., 2015). All beginning cases are dispersed into the regression tree’s terminal nodes.

The following stages can be used to describe the CART method (Gupta et al., 2017):

• The following formal formula is used to calculate the impurity of D and the potential result.

$$Gini(D)=1-{\displaystyle {\sum }_i^m{p}_i^2}$$

Where $p_i$ is the probability that a tuple in data D belongs to class $C_i$, and it is given by $\left|C_{i,D}\right|/\left|D\right|$.

• The following formula is used to compute each partition attribute’s impurity:

$$Gin{i}_A(D)=\frac{\left|D_1\right|}{\left|D\right|}Gini(D_1)+\frac{\left|D_2\right|}{\left|D\right|}Gini(D_2)$$

The optimum binary split should then be chosen for use in the following phase by selecting the partition attribute with the lowest Gini index.

• The following formula is utilized to determine the impurity reduction:

$$\Delta Gini(A)=Gini(D)-Gin{i}_A(D)$$

The splitting attribute is determined by selecting the feature with the lowest Gini index and the largest drop in impurity (Gupta et al., 2017).

2.2.3 Artificial neural networks

An information processing system called an artificial neural network (ANN) is modeled after biological systems, such the human brain. The brain’s distinctive characteristics include learning new ideas, making judgments, and deriving conclusions from complex and perhaps irrelevant or incomplete data. The widespread use of ANN stems from their limited capacity to mimic brain functions, albeit in a limited capacity (Samarasinghe, 2007). ANNs, therefore, provide an alternative methodology to conventional statistical techniques, which call for the definition of an algorithm and its recording as a computer program. ANNs are instead given example tasks, and their weight coefficients and connections between network parts are automatically adjusted based on the training method (Tadeusiewicz and Lula, 2007).

A typical artificial neuron and the modeling of a multilayered neural network are as follows. The signal flow from inputs $x_1,x_2,.,x_n$ is considered to be unidirectional, which is a neuron’s output signal flow (O). The neuron output signal O is given by the following relationship:

$$O=f(net)=f\left({\displaystyle {\sum }_{i=1}^n{w}_i{x}_i}\right)$$

The weight vector in this case is denoted by $w_i$, and the function $f(net)$ is known as an activation (transfer) function. A scalar product of the weight and input vectors defines the variable net,

$$net=w^{T}x=w_1{x}_1+w_2{x}_2+\dots +w_n{x}_n$$

When T is a matrix’s transpose, and the output value O is calculated in the most basic scenario as

$$O=f(net)=\{\begin{array}{c}1 if w^{T}x\ge \theta \\ 0 otherwise\end{array}$$

where the node type is referred to as a linear threshold unit and θ is known as the threshold level (Abraham, 2005).

2.2.3.1 Perceptron neural networks

By establishing the weights and relevant functions, a multilayer perceptron network can be utilized to solve many exceedingly complex mathematical problems that involve complex nonlinear equations. Different activation functions can be utilized in neurons depending on the kind of issue. In these networks, there are three layers: an input layer that introduces issue inputs, a hidden layer, and an output layer that offers the solution. Backpropagation is a popular training technique for these networks (Manhaj, 2002).

2.2.3.2 Artificial neural network structures

The linked collection of neural networks often uses mathematical techniques to handle data. This multilayer perceptron network (MLP) has three layers: an input layer, a hidden layer, and an output layer for input data, data processing, and output data, respectively. Each layer is composed of many artificial neurons, or nodes. All neurons are linked to each other, except within the same layers. The results are categorized and moved to the output layer using hidden layers. The target variable’s anticipated values are likewise displayed in the output layer. An estimate of the stations’ daily discharge is displayed in the output layer of the current research. The backpropagation technique is used in the multilayer perceptron network’s training process. This algorithm defines the starting weights, which are then assigned to the knots. Next, the model is updated to include the learning samples, after which the output is produced and contrasted with trial samples. When differences exceed the designated cutoff point, the weights are adjusted until the difference between the intended and actual outputs is minimized. This procedure is carried out until the maximum number of iterations or a previously established level of precision is reached (Islam et al., 2001).

The input layer of a feed-forward backpropagation neural network receives external evidence. These inputs are then moved to input variables via the identity transfer function. Through the connections between input layer and hidden layer neurons, scientists were able to access the hidden layers. The basic calculation of ANNs performed in these layers is achieved by connecting weights between the neurons of hidden layers (Nowruzi and Ghassemi, 2016). In order to weight the summations of the outputs from the preceding layer in the neurons of the hidden layer, they are adding biasedly. This total is then transferred using a transfer function. For a neuron in the buried layer, the hyperbolic tangent sigmoid transfer function is implemented by

$$n_j=\frac{2}{1+e^{-2Z}}-1$$

where Z will be ascertained as follows, and n_j represents the jth neuron output.

$$Z=\sum _{i=1}^r{w}_{ij}{p}_i+b_j$$

Here, p_i is the ith neuron’s output, and ω_ij are the ith neuron’s connectivity weights from the previous layer to the jth neuron. Furthermore, b_j is the bias, and r is the number of neurons in the preceding layer.

In addition to hyperbolic tangent activation, other activation functions such as linear activation function, sigmoid function, exponential linear unit, and Softmax function can be used in artificial neural networks.

The linear activation function can be defined as:

$$F(Z)=aZ$$

Any constant value that the user selects can be the value of variable a (Sharma et al., 2020).

The sigmoid function can be defined as follows (Sibi et al., 2013):

$$sig(Z)=\frac{1}{1+e^{-Z}}$$

Exponential linear unit introduces a parameter slope for the negative values of x. It uses a log curve for defining the negative values (Sharma et al., 2020).

$$f(Z)=Z, Z\ge 0$$

$$f(Z)=a(e^Z-1), Z<0$$

The Softmax function is a combination of multiple sigmoid functions. Since a sigmoid function is known to yield values between 0 and 1, these values may be interpreted as the probability of the data points belonging to a certain class. The Softmax function can be used for multiclass classification issues, in contrast to sigmoid functions, which are utilized for binary classification. The function yields the probability for each data point across all classes. It can be stated as follows: (Sharma et al., 2020).

$$\sigma {(Z)}_i=\frac{e^{Z_i}}{{\sum }_{k=1}^K{e}^{Z_k}} for i=1,2,\dots ,\mathit{K.}$$

In this study, the highest R² and adjusted R² and the lowest RMSE, MAPE, SD ratio, AIC, and corrected Akaike information criterion (AICc) values were achieved using the hyperbolic tangent activation function. The hyperbolic tangent activation function was used because it provided the best prediction.

The output layer will receive the results. Thus, the output layer will obtain the output variable. In the output layer, the linear transfer function (λ) is applied as,

$$g=\lambda (w_{L}Z+b_0)$$

where the output layer and the final hidden layer’s connectivity weights are denoted by $w_L$. Furthermore, the output layer bias is b₀ (Rumelhart et al., 1986).

2.2.4 Multivariate adaptive regression spline

Multivariate adaptive regression spline (MARS) is a nonparametric modeling technique that adds nonlinearities and variable interactions to the linear model. This approach is an extension of recursive partitioning regression (RPR), which creates distinct subregions inside the predictor variable space (Friedman, 1991; Montero, 2013). The model is expressed as follows:

$$y_t=f(x_t)={\beta }_0+\sum _{i=1}^k{\beta }_{i}B(x_{it})$$

where ${\beta }_i$, which range from i = 1,…, k, are the model parameters for the corresponding x_it variables and $y_t$ is the response variable at instant t. The intercept is represented by the value ˇ0, and the basis functions each $B(x_{it})$ may be expressed as

$$B(x_{it})=\text{max}(0, x_{it}-c)$$

or

$$B(x_{it})=\max (0, c-x_{it}),$$

where c is a threshold value and k is the number of explanatory variables, which includes interactions of the predictor variables (Salford Systems, 2001a). By using only a small number of knots c, the MARS algorithm (Friedman, 1991) aims to fit splines of the form $(x_{it}-c)$ and $(c-x_{it}),$ to high-dimensional data. Thus, in a forward stepwise fashion, the algorithm looks for the ideal c to approximate the relationship between $y_t$ and the predictor variables x_it. It begins with an empty model and adds knots to the model recursively for each of the predictor variables in x_it. The variable and knot selected at each phase are chosen to produce the greatest reduction in the final model’s error (Friedman, 1991). For both forms, consider it as a functional value x_it. In the first version, x_it equals $x_{it}-c$ for all values of x greater than c and is set to 0 for all values of x_it up to a threshold value, c. In the second form, x_it equals $c-x_{it}$ for all values of x less than c and is set to 0 for all values of x_it greater than a threshold value, c (Abraham et al., 2001). Every function has a knot at value c and is piecewise linear. Linear nonsmooth splines are these transformed functions (Hastie et al., 2009). $B(x_{it})$ are functions that rely on the corresponding x_it variables. The data analysis yields the space partition points and model parameters. The complexity of the model is indicated by the number of derived basis functions (Salford Systems, 2001a).

The least squares approach is used to identify the functions with the best estimate performance once the fundamental functions and knots have been identified (Friedman et al., 2001). Generalized crossvalidation (GCV) measurement serves as the basis for model selection (Salford Systems, 2001b).

$$GCV=\sum _{i=1}^n\frac{{(y_i-\widehat{y_i})}^2}{{(1-C(M)/n)}^2}$$

where C(M) displays a penalty measure correlated with the number of chosen parameters, and $\widehat{y}$ denotes the projected values.

The differences between each method used were displayed in Table 1.

Table 1

Table 1. Differences between MARS, CHAID, CHART, and ANN methods.

The following goodness-of-fit criteria were computed in order to compare the prediction performance of the approaches in 10-fold crossvalidation (Willmott and Matsuura, 2005; Takma et al., 2012):

Pearson correlation coefficient (r) between the actual and predicted yield (WGY) values,

1. Akaike information criterion (AIC) calculated as:

$$AIC=nln\left[\frac{1}{n}{(y_i-\widehat{y_i})}^2\right]+2k, if\frac{n}{k}>40$$

or:

$$AI{C}_c=nln\left[\frac{1}{n}{(y_i-\widehat{y})}^2\right]+2k+\frac{2k(k+1)}{n-k-1}, \text{otherwise}$$

2. Root-mean-square error (RMSE) given by the following formula:

$$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n{(y_i-\widehat{y_i})}^2}$$

3. Standard deviation ratio (SD_ratio):

$$S{D}_{ratio}=\frac{s_m}{s_d}$$

4. Mean absolute percentage error (MAPE):

$$MAPE=\frac{1}{n}\sum _{i=1}^n\left|\frac{y_i-\widehat{y_i}}{y_i}\right|.100$$

5. Coefficient of determination

$$R^2=1-\frac{{\sum }_{i=1}^n{(y_i-\widehat{y_i})}^2}{{\sum }_{i=1}^n{(y_i-{\overline{y}}_i)}^2}$$

6. Adjusted coefficient of determination

$$Adj.R^2=1-\frac{\frac{1}{n-k-1}{\sum }_{i=1}^n{(y_i-\widehat{y_i})}^2}{\frac{1}{n-1}{\sum }_{i=1}^n{(y_i-{\overline{y}}_i)}^2}$$

where n is the number of cases in a set, k is the number of model parameters, $y_i$ is the output variable’s actual (observed) value, $\widehat{y}$ is its predicted value, $s_m$ is the standard deviation of model errors, and $s_d$ is its output variable’s standard deviation (WGY).

Data mining algorithms are non-parametric statistical techniques that do not require a normality assumption for the dependent variables. These algorithms perform well in cases of missing data for independent variables and can be applied to both large and small datasets (Yordanova et al., 2015).

For algorithms that have varying input numbers, the adjusted coefficient of determination can be utilized as a goodness-of-fit criteria. This adjustment accounts for difference in the number of input variables. The MARS algorithm defines k as the number of phrases. IBM SPSS 26 (IBM Corp. Released, 2019) was used for statistical analyses of the CHAID, CART, and ANN algorithms, while the R Studio software (R Core Team, 2022) described the MARS method.

3 Results

3.1 Descriptive statistics

Table 2 provides descriptive information about the traits (wet grass yield, crude protein, crude ash, ADF, and NDF) of peas grown in 14 distinct genotypes.

Table 2

Table 2. Descriptive statistics of characteristics of peas.

3.2 Results of correlation matrix and principal component analysis

The correlation matrix for the characteristics of peas is presented in Figure 1.

Figure 1

Figure 1. Correlation coefficients between features.

When examining the correlation coefficients in Figure 1, the highest correlations are observed between the ADF-NDF (0.490) and ADF-crude protein (− 0.450) variables. Correlation coefficients between other traits are low and statistically insignificant. The lowest correlations were between WGY-ADF (0.001), WGY-NDF (0.024), WGY-Protein (0.030), and ADF-Ash (− 0.044), respectively. The representation of the principle component analysis (PCA) graph for the same variables is presented in Figure 2.

Figure 2

Figure 2. Relationship among traits according to principal component analysis.

In PCA analysis, the first principal component (PC1) accounted for 36%, while the second principal component (PC2) accounted for 23.1%. Together, PC1 and PC2 explained a total of 59.1% of the variation. An angle between the slices between 0° and 90° is interpreted as a positive correlation among the traits within those slices, whereas an angle between 90° and 180° is interpreted as a negative association. If the angle is exactly 90°, it indicates no relationship between the traits (Yan and Tinker, 2005: Aktaş, 2017). As the vector moves away from the origin, the variation between variables increases according to the trait examined, whereas the variation decreases as the vector approaches the origin (Abate et al., 2015). Accordingly, the relationship between ADF and NDF is positive. In contrast, the relationships between ADF-WGY and ADF-Crude Ash variables are positive but very weak. Conversely, the relationship between ADF-Crude Protein is negative. While the relationship between NDF and WGY and crude ash variables is positive, the relationship between NDF and crude protein is negative. The relationships between WGY and both crude protein and crude ash are positive. Additionally, the relationship between crude protein and crude ash is also positive.

CHAID, CART, ANN, and MARS techniques were used to investigate the impacts of characteristics on wet grass yield in peas, and the findings are given below.

3.3 Result of CHAID algorithm

The CHAID method was used to assess the impacts of various factors on wet grass yield. The parent node to child node ratio was set at 8:4, as this configuration provided better goodness-of-fit criteria within the CHAF algorithm. Crossvalidation was performed with a setting of 10. The regression tree diagram resulting from the CHAID method is shown in Figure 3.

Figure 3

Figure 3. CHAID classification tree diagram of the estimation in wet grass yield.

An analysis of the CHAID diagram revealed that genotype (Adj. p-value = 0.003, F = 33.192) was the first-order effective independent variable influencing wet grass yield of peas, followed by ADF (Adj. p-value = 0.008, F = 21.306) as the second-order variable (Figure 3). Throughout the whole tree construction process, the branches produced by independent variables were statistically significant (p < 0.05). In terms of R², SD ratio, RMSE, MAPE, AIC, and AICc, the CHAID algorithm’s performance was determined to be 0.759, 0.564, 109.75, 9.198, 443, and 876, respectively. The results of the CHAID algorithm indicated that the highest yield in peas was found to be 1,329.889 kg for the RETNA, GATEM-101, ÜRÜNLÜ, and GÖLYAZI lines, with ADF > 33.59.

3.4 Result of CART algorithm

The CART technique was used to determine the effects of various variables on wet grass yield. The ratio of parent nodes to child nodes was established at 8 to 4, which resulted in improved goodness-of-fit criteria for the CART algorithm. Figure 4 illustrates the regression tree diagram generated by the CART algorithm. A crossvalidation approach was set to 10.

Figure 4

Figure 4. CART classification tree diagram of the estimation in wet grass yield.

The CART diagram revealed that genotype (improvement = 24.091) was the first-order effective independent variable affecting the wet grass yield of peas, closely followed by ADF (improvement = 24.091) (Figure 4). Independent factors created significant branches over the tree construction process (p < 0.05). The CHAID method performed well in terms of R², SD ratio, RMSE, MAPE, AIC, and AICc, with values of 0.752, 0.576, 111.526, 9.791, 499, and 918, respectively. According to the CART algorithm, the greatest pea yield was recorded at 1,329.889 kg when ADF > 33.625.

3.5 Result of artificial neural network

The multilayer perceptron artificial neural network model was chosen for its suitability to the data. A training ratio of 70% and testing ratio of 30% were used, with the scaled conjugate gradient selected as the optimization algorithm. In the present study, an ANN with 10-fold crossvalidation was applied. Figure 5 shows the connections within the ANN.

Figure 5

Figure 5. Artificial neural network architecture.

In Figure 5, the activation function used in the hidden layer of the artificial neural network architecture is the hyperbolic tangent, while the output employs the identity activation function. The parameter estimates of the ANN are presented in Table 3.

Table 3

Table 3. Parameter estimates of ANN.

The connections between each neuron in Table 3 are as follows:

The connection weight value between protein in the input layer and H(1:1) of the first neuron in the hidden layer is − 0.299. The connection weight value between H(1:2) of the second neuron in the hidden layer is 0.699, while the connection weight value between H(1:3) of the third neuron in the hidden layer is 0.084.

The connection weight value between the ash in the input layer and the H(1:1) of the first neuron in the hidden layer is − 0.193. The connection weight value between the H(1:2) of the second neuron in the hidden layer is − 0.577, and the connection weight value between H(1:3) of the third neuron in the hidden layer is 0.240.

The connection weight value between ADF in the input layer and H(1:1) of the first neuron in the hidden layer is 0.077. For the second neuron in the hidden layer, H(1:2), the connection weight value between is 0.444, while for the third neuron, H(1:3), the connection weight value is 0.389.

The connection weight value between NDF in the input layer and H(1:1) of the first neuron in the hidden layer is 0.427, while H(1:2) of the second neuron has weight of 0.141, and H(1:3) of the third neuron has a weights of − 0.430.

The learning sum of squares error (SSE) in the ANN model was 3.638, with a relative error of 0.269. For the test data, the SSE was 6.577, and the relative error was 0.594.

Table 4 shows the percentage of importance of independent variables.

Table 4

Table 4. Independent variable importance.

As shown in Table 4, the independent variables affecting wet grass yield in the output layer include genotype (line) with a coefficient of 0.409, protein at 0.317, ash at 0.083, ADF at 0.046, and NDF at 0.145. Figure 6 presents a percentage column graph illustrating the influence of these variables on the prediction.

Figure 6

Figure 6. Importance of variables.

As can be seen in Table 5, genotype (line) has the highest influence, accounting for 100% of the effect on the fresh herbage yield of pea plants sold from the terminal on this model. In addition, crude protein is the second most important independent variable with a rate of 77.4%, while NDF accounts for 35.5%. Crude ash has an effect of 20.2%, and ADF has the least impact, with a rate of 11.3% on the fresh herbage yield from the terminals.

Table 5

Table 5. MARS, CHAID, CART, and ANN types’ predictive performance.

3.6 MARS algorithm results

The model estimation coefficients of the MARS algorithm, which predicts the fresh herbage yield of pea plants, are provided in Table 6. A penalty of − 1 penalty and 10-fold crossvalidation were applied in the R studio free software to improve the predictive accuracy of the MARS algorithm.

Table 6

Table 6. Results of MARS algorithm in the prediction of fresh herbage yield of pea plants.

According to presented results in Table 4, all coefficients concerning MARS predictive model were statistically significant (p < 0.05, p < 0.01, p < 0.001). The desirable predictive quality of the MARS equation produced here was obtained with ensuring the smallest GCV (110). The recorded or observed values in fresh herbage yield of pea plants were correlated very strongly with those predicted by the MARS model (p < 0.001), indicating effective plant yield modeling. For prediction equation of MARS model with 35 terms, no overfitting problem was recorded, as evidenced by the R² estimate (0.998) being close to the CVR² estimate (0.782). The present SD ratio of 0.047, RMSE of 10.499, MAPE of 0.7365, AIC of 268, and AICc of 688 indicate that the MARS model, which captures influential factors, demonstrates an excellent fit.

According to the MARS method, several terms and coefficients can be read as follows: The equation derived by incorporating the interaction effect of the model’s coefficients is shown in detail below. The effect and corresponding positive coefficient (69.189) on fresh herbage yield were shown to be favorably correlated when ADP ≤ 31.02 in peas; on the other hand, an adverse corresponding negative coefficient (− 126.666) on fresh herbage yield was identified when protein > 9.67. The greatest positive effect on fresh herbage yield in peas was 36,677.515, when the genotype was Özkaynak. The second highest favorable effect, with an increase of 3,143.449, occurred when the genotype was Gate101 and protein > 9.67 have. The third largest positive effect was noted when the genotype was Özkaynak and Cash was present, leading to an increase in fresh herbage yield of 1,575.428.

The greatest negative effect on fresh herbage yield is – 29,761.909 if genotype = P104. The second and third largest negative effects on fresh herbage yield are – 17,608.673 if genotype = Golyazi cm and – 17,220.45 when CAsh > 9.87 cm, respectively.

WGY = 1.23e+03 − 72.4*GenotypGate101 – 17,609*GenotypGolyazi + 36,678*GenotypOzkayna

− 29,762*GenotypP104 – 2,525*max(0, 9.67 − protein) − 127*max(0, protein − 9.67)

− 345*max(0, 9.87 − CAsh) – 17,220*max(0, CAsh − 9.87) + 69.2*max(0, 31 − ADF)

− 119*max(0, ADF − 31) + 95.9*max(0, 41.7 − NDF) + 266*max(0, NDF − 41.7)

− 219*max(0, NDF − 42.1) − 60*max(0, NDF − 43.1) + 447*GenotypGolyazi * NDF

+ 1575*GenotypOzkayna*CAsh – 1,192*GenotypOzkayna*NDF

+ 651*GenotypP104*NDF − 202*GenotypATOS*max(0, protein − 9.67)

+ 71*GenotypATOS*max(0, 9.87 − CAsh) + 3,143*GenotypGate101*max(0, protein − 9.67)

+ 162*GenotypP101*max(0, 9.87 − CAsh) – 1,609*GenotypP51*max(0,CAsh − 9.87)

+ 669*GenotypRETNA*max(0, CAsh − 9.87) + 335*GenotypRETNA*max(0, ADF − 31)

− 3,132*GenotypSprinP*max(0, 9.67− protein) − 99.9*GenotypSprinP*max(0, 31 − ADF)

+ 227*GenotypSpring max(0, ADF − 31) + 571*GenotypUrunlu*max(0, protein − 9.67)

− 262*GenotypUrunlu * max(0, ADF − 31) + 14.7*protein*max(0, ADF − 31)

+ 81.1*max(0, 9.67 − protein)*ADF + 385*max(0, CAsh − 9.87) * NDF

+ 142*GenotypGolyazi*max(0, 9.67 − protein)*ADF

The relative importance of the variables predicting fresh herbage yield as a result of the MARS algorithm is given in Figure 7.

Figure 7

Figure 7. Graph of relative importance for fresh herbage yield.

The graph of the estimated values generated by the MARS algorithm alongside the observed values is shown in Figure 8.

Figure 8

Figure 8. Agreeable expectations and observed fresh herbage yield values.

Assuming the plant has the following characteristics (genotype = “Özkaynak”, protein = 10.5, Cash = 8, ADF = 32, and NDF = 41.3), the predicted fresh herbage yield of the pea is 992.648 kg.

When predicting fresh herbage production, all algorithms yielded suitable results (Table 5). The predicted accuracy of the algorithms was ranked in the following order of superiority: MARS > CHAID > CART > ANN.

MARS, multivariate adaptive regression spline; CHAID, Chi-square automatic interaction detection; CART, classification and regression tree; ANN, artificial neural network; AIC, Akaike information criterion; AICc, corrected Akaike information criterion; RMSE, root-mean-square error; SD, standard deviation; MAPE, mean absolute percentage error; R², coefficient of determination.

4 Discussion

The average green yield (kg/ha) of pea plants grown in different varieties was found to be between 735.89 and 1,273.56. The crude protein content (%) was 7.64–13.81, crude ash content (%) was 7.52–12.09, NDF content (%) was 37.18–46.03, and ADF content (%) was 27.75–35.04. Notably, the crude protein content observed in this study was lower than the findings of Alatürk et al. (2021) and Uzun et al. (2012), who reported crude protein content percentages of 16.8–20.5 and 9.5–20.4, respectively. The crude ash content in this study was also lower than the findings reported by Alatürk et al. (2021), which ranged from 10.9% to 13.0%. However, NDF and ADF contents of pea forage were in accordance with the study of Alatürk et al. (2021), who reported values of 38.4%–43.2% and 28.6%–34.5%, respectively. In a related study, Çaçan et al. (2018) investigated several forage pea lines and cultivars to determine seed yield, hay yield, and hay quality. They found that seed yield ranged from 33.8 to 180.2 kg ha⁻¹, hay yield ranged from 160.3 to 887.0 kg ha⁻¹, ash content ranged from 9.42% to 11.19%, crude protein content ranged from 6.54% to 11.91%, crude protein yield ranged from 11.9 to 104.9 kg ha⁻¹, ADF ranged from 29.5% to 39.8%, and NDF ranged from 39.1% to 51.2%, respectively. The Gatem, Ürünlü, Gölyazı, and Spring Pea 3-638 genotypes exhibited superior characteristics under Bingöl ecological conditions. In a study by Karadeniz and Bengisu (2022), which investigated the effects of row spacing on yield and quality in green peas (Pisum sativum ssp. arvense), crude protein was found to range from 20.2% to 22.5%, crude ash from 8.0% to 8.9%, ADF from 32.00% to 33.65%, and NDF from 42.4% to 44.9%. Additionally, a study by Victor (2022) assessed the feed value of green mass of annual legume species and reported the crude protein, ADF, and NDF values of fodder pea as 142 g/kg DM, 392 g/kg DM, and 598 g/kg DM, respectively. Çalık (2020) determined crude protein levels of 10.3%–20.1%, ADF ratios of 21.7%–36.4%, and NDF ratios of 33.2%–43.4% for fodder pea, based on an animal nutrition study assessing various legumes consumed as roughage in Şanlıurfa. In addition, Sarıkaya et al. (2023) conducted a study to determine the effects of different sowing times and plant densities on dry grass yield and quality in some forage pea varieties, finding average values of 14.16% for crude protein, 27.98% for ADF, and 37.64% for NDF. Kara and Sürmen (2023) examined eight different forage pea cultivars (Kirazlı, Ulubatlı, Ürünlü, Gölyazı, Özkaynak, Töre, Taşkent, GAP Pembesi) and subjected them to mowing treatments in three different phenological periods (10%, 50%, and 100% flowering) under Aydın ecological conditions, finding crude protein levels of 19.88%, ADF ratios of 32.57%, and NDF ratios of 45.25%. In a separate study, Kır (2022) investigated the appropriate mixing ratios of fodder pea (Pisum sativum ssp. arvense L.) and rye (Secale cereale L.) under rainy conditions in Kırşehir province during the 2018–2019 vegetation period, reporting crude protein levels of 15.6%, ADF ratios of 31.1%, and NDF ratios of 39.3%.

In the study of Çelik et al. (2024), the factors influencing fresh herbage yield in sorghum–sudangrass hybrid plants were analyzed using CHAID, CART, MARS, and Bagging MARS algorithms. MARS, Bagging MARS, CART, and CHAID were found to be the most appropriate algorithms for predicting the dependent variable. In this study, the MARS algorithm emerged as the best predictor of crop yield, followed by CHAID, CART and ANN methods, respectively. As in the authors’ study, the CART algorithm ranked as the third best method in this study. The R², adjusted R², and SD ratio statistics were closely aligned in both this study and the authors’ research.

The findings obtained of this study show some similarities and some differences with those of previously mentioned researchers. These differences are likely attributed to a combination of factors, with one of the most significant being the variations in climate and soil structures among the study areas. Different climatic and soil conditions can directly affect the development of vegetation and the application of agricultural practices. This can lead to differences in the findings. Another factor is the different practices used by different researchers (such as different fertilization methods, irrigation techniques, weed control methods, and harvest times).

5 Conclusion

In this study, the performances of CHAID, CART, ANN, and MARS methods were analyzed to predict wet grass yield in pea plants. The input variables included genotype (line), crude protein (%), crude ash (%), ADF (%), and NDF (%). The results were compared using different goodness-of-fit tests, including the coefficient of determination (R²), adjusted R², RMSE, MAPE, SD ratio, AIC, and AICc. The results of this study are presented below.

According to the results of the MARS algorithm, the variables that contributed the most to wet herbage yield in pea plants were genotype, crude protein, crude ash, NDF and ADF. As a result of the application of artificial neural network method, the order of importance of the variables affecting wet grass yield in pea was identified as genotype, crude protein, NDF, crude ash, and ADF. The CHAID algorithm estimated the highest fresh herbage yield of pea at 1,329.889 kg in RETNA, GATEM-101, ÜRÜNLÜ, and GÖLYAZI lines, with ADF > 33.59. When the CART algorithm was applied, the highest herbage yield was reached when ADF > 33.625, resulting in an estimated yield of 1,329.889 kg. In this case, the results from the CHAID and CART algorithms were very close to each other. The performance findings are as follows: MARS > CHAID > CART > ANN (best to worst).

It was determined that mining approaches are quite effective in field agricultural data for identifying factors influencing plant production and predicting any variables.

Data availability statement

Publicly available datasets were analyzed in this study. For more information on the original contributions presented in this study, please contact the corresponding authors.

Ethics statement

The manuscript presents research on animals that do not require ethical approval for their study.

Author contributions

MÇ: Data curation, Investigation, Methodology, Resources, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. ŞÇ: Conceptualization, Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review & editing. AB: Conceptualization, Data curation, Investigation, Methodology, Resources, Validation, Writing – original draft, Writing – review & editing.

Funding

The author(s) declare that no financial support was received for the research, authorship, and/or publication of this article.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that no Generative AI was used in the creation of this manuscript.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Abate, F., Mekbib, F., Dessalegn, F. (2015). GGE biplot analysis of multienvironment yield trials of durum wheat (Triticum turgidum Desf.) genotypes in North Western Ethiopia. Am. J. Exp. Agric. 8, 120–129.

Google Scholar

Abraham, A. (2005). “Artificial neural networks,” in Handbook of measuring system design. Eds. Sydenham, P. H., Thorn, R. (USA: John Wiley & Sons, Ltd).

Google Scholar

Abraham, A., Steinberg, D., Philip, N. S. (2001). Rainfall forecasting using soft computing models and multivariate adaptive regression splines. IEEE SMC Transactions Special Issue Fusion Soft Computing Hard Computing Ind. Appl. 1, 1–12.

Google Scholar

Açıkgöz, E. (2001). Yem bitkileri (Uludağ Üniversitesi Güçlendirme Vakfı Yayın), 584. No: 182 Vipaş AŞ Yayın No: 58 (Bursa: 3. Baskı) s.

Google Scholar

Adesogan, A. T., Salawu, M. B., Williams, S. P., Fisher, W. J., Dewhurst, R. J. (2004). Reducing concentrate supplementation in dairy cow diets while maintaining milk production with pea-wheat intercrops. J. Dairy Sci. 87, 3398–3406. doi: 10.3168/jds.S0022-0302(04)73475-X

PubMed Abstract | Crossref Full Text | Google Scholar

Aktaş, H. (2017). Türkiye’de yoğun ekim alanına sahip bazı arpa (Hordeum vulgare L.) çeşitlerinin destek sulamalı ve yağışa dayalı koşullarda değerlendirilmesi. Tekirdağ Ziraat Fakültesi Dergisi. 14, 86–97.

Google Scholar

Alatürk, F., Çınar, Ç., Gökkuş, A. (2021). Farklı Sıra Aralıklarının Bazı Yem Bezelyesi Çeşitlerinin verim ve kalitesi üzerine etkileri. Türk Tarım ve Doğa Bilimleri Dergisi 8, 53–57. doi: 10.30910/turkjans.688894

Crossref Full Text | Google Scholar

Alkhasawneh, M. S., Kalthum Ngah, U., Tien Tay, L., Ashidi Mat Isa, N., Subhi Al-Batah, M. (2014). Modeling and testing landslide hazard using decision tree. J. Appl. Mathematics 929768, 1–9. doi: 10.1155/2014/929768

Crossref Full Text | Google Scholar

Berhane, S., Berhe, H., Gebrekorkos, G., Abera, K. (2016). Determination of planting spacing for improved yield and yield components of Dekoko (Pisum sativum var. abyssinicum) at Raya Valley, Northern Ethiopia. Afr. J. Plant Sci. 10, 157–161. doi: 10.5897/AJPS2016.1428

Crossref Full Text | Google Scholar

Borreani, G., Chion, A. R., Colombini, S., Odoardi, M., Paoletti, R., Tabacco, E. (2009). Fermentative profiles of field pea (Pisum sativum), faba bean (Vicia faba) and white lupin (Lupinus albus) silages as affected by wilting and inoculation. Anim. Feed Sci. Technol. 151, 316–323. doi: 10.1016/j.anifeedsci.2009.01.020

Crossref Full Text | Google Scholar

Borreani, G., Peiretti, P. G., Tabacco, E. (2007). Effect of harvest time on yield and pre-harvest quality of semi-leafless grain peas (Pisum sativum L.) as whole-crop forage. Field Crops Res. 100, 1–9. doi: 10.1016/j.fcr.2006.04.007

Crossref Full Text | Google Scholar

Breiman, L., Friedman, J. H., Olshen, R. A., Stone, C. J. (1984). Classification and Regression Trees (Boca Raton, USA: Chapman and Hall/CRC Press).

Google Scholar

Çaçan, E., Kaplan, M., Kökten, K., Tutar, H. (2018). Evaluation of some forage pea (Pisum sativum ssp. arvense L.) lines and cultivars in terms of seed yield and straw quality. Iğdır Üni. Fen Bilimleri Enst. Der. 8, 275–284. doi: 10.21597/jist.428996

Crossref Full Text | Google Scholar

Cak, B., Keskin, S., Yilmaz, O. (2013). Regression tree analysis for determining of affecting factors to lactation milk yield in brown Swiss cattle. Asian J. Anim. Veterinary Adv. 8, 677–682. doi: 10.3923/ajava.2013.677.682

Crossref Full Text | Google Scholar

Çalık, A. (2020). “The value of some legumes consumed as roughage in şanlıurfa in animal nutrıtıon,” in Academic Studies in Agriculture (USA: Forestry and Aquaculture-II), 57.

Google Scholar

Cavallarin, L., Tabacco, E., Borreani, G. (2007). Forage and grain legume silages as a valuable source of proteins for dairy cows. Ital. J. Anim. Sci. 6, 282–284. doi: 10.4081/ijas.2007.1s.282

Crossref Full Text | Google Scholar

Çelik, Ş., Boydak, E., Fırat, R. (2018). An analysis of factors affecting yield, oil production rate and plant height in sunflowers using selected data mining algorithms. J. Anim. Plant Sci. 28, 1085–1093.

Google Scholar

Çelik, Ş., Tutar, H., Gönülal, E., Er, H. (2024). Prediction of fresh herbage yield using data mining techniques with limited plant quality parameters. Sci. Rep. 14, 21396. doi: 10.1038/s41598-024-72746-9

PubMed Abstract | Crossref Full Text | Google Scholar

Çelik, Ş., Yilmaz, O. (2018). Prediction of body weight of Turkish Tazi dogs using data mining techniques classification and regression tree CART and multivariate adaptive regression splines MARS. Pakistan J. Zoology (PJZ) 50, 575–583. doi: 10.17582/journal.pjz

Crossref Full Text | Google Scholar

Eyduran, E., Yilmaz, I., Tariq, M. M., Kaygisiz, A. (2013). Estimation of 305-D milk yield using regression tree method in brown Swiss cattle. J. Anim. Plant Sci. 23, 731–735.

Google Scholar

Friedman, J. (1991). Multivariate adaptive regression splines. Ann. Statist. 19, 1–141. doi: 10.1214/aos/1176347963

Crossref Full Text | Google Scholar

Friedman, J., Hastie, T., Tibshirani, R. (2001). The elements of statistical learning Vol. 1 (New York: Springer series in statistics).

Google Scholar

Gallagher, C., Monroe, H., Fish, J. (2005). An iterative approach to classification analysis. J. Appl. Stat, 237–280. Available at: http://www.casact.org/pubs/dpp/dpp90/90dpp237.pdf.

Google Scholar

Gözüaçık, C., Eyduran, E., Çami, H., Kara, M. K. (2018). Detection of infection preferences of the alfalfa seed chalcid, Bruchophagus roddi Gussakovskiy 1933 (Hymenoptera: Eurytomidae) in alfalfa (Medicago sativa L.) fields of Igdir, Turkey. Legume Res. 41, 150–154. doi: 10.18805/lr.v0iOF.9099

Crossref Full Text | Google Scholar

Gupta, B., Rawat, A., Jain, A., Arora, A., Dhami, N. (2017). Analysis of various decision tree algorithms for classification in data mining. Int. J. Comput. Appl. 163, 15–19. doi: 10.5120/ijca2017913660

Crossref Full Text | Google Scholar

Hafiz, M. I. U., Rehman, S., Bilal, M., Naqvi, H. A. S., Manzoor, A. S., Ghafoor, A., et al. (2014). Evaluation of genetic diversity in pea (Pisum sativum) based on morphoagronomic characteristics for yield and yield associated traits. J. Biol. Environ. Sci. 4, 321–328.

Google Scholar

Hastie, T., Tibshirani, R., Friedman, J. (2009). The Elements of Statistical Learning, Data Mining, Inference and Prediction (Springer Science+Business Media, LLC), ISBN: 978-0-387-84857-0.

Google Scholar

IBM Corp. Released (2019). IBM SPSS Statistics for Windows, Version 26.0 (Armonk, NY: IBM Corp).

Google Scholar

Islam, M. M., Sado, K., Owe, M., Brubaker, K., Ritchie, J., Rango, A. (2001). Flood Damage and Management Modelling Using Satellite Remote Sensing Data with GIS: Case Study of Bangladesh (USA: IAHS PUBLICATION), 455–457.

Google Scholar

Kara, E., Sürmen, M. (2023). Yield and quality characteristics of forage pea varieties at different phenological stages. Adnan Menderes Üniversitesi Ziraat Fakültesi Dergisi 20, 295–301. doi: 10.25308/aduziraat.1392323

Crossref Full Text | Google Scholar

Karadeniz, E., Bengisu, G. (2022). Effects of Row Spacing on Yield and Quality of Forage Pea (Pisum sativum ssp. arvense). Turkish J. Range Forage Sci. 3, 30–35. doi: 10.51801/turkjrfs.1100519

Crossref Full Text | Google Scholar

Kır, H. (2022). Effects of different forage pea and rye mixtures on forage yield and quality. Turkish J. Range Forage Sci. 3, 11–17. doi: 10.51801/turkjrfs.1073958

Crossref Full Text | Google Scholar

Manhaj, M. (2002). Principles of Artificial Neural Networks (Tehran: Published by Industrial University of Amirkabir (Tehran Polytechnic).

Google Scholar

Maria, A. E., Milan, A. L., Martin, A. E., Cravero, P. V., Anido, L. S. F., Cointry, L. E. (2009). Comparison of morphological and molecular data for pea (Pisum sativum) in low and high yielding environments. New Z. J. Crop Hortic. Sci. 37, 227–233. doi: 10.1080/01140670909510268

Crossref Full Text | Google Scholar

Montero, R. (2013). “Variables no estacionarias y cointegración,” in Documentos de Trabajo en Economía Aplicada (Universidad de Granada, Espana).

Google Scholar

Nowruzi, H., Ghassemi, H. (2016). Using artificial neural network to predict velocity of sound in liquid water as a function of ambient temperature, electrical and magnetic fields. J. Ocean Eng. Sci. 1, 203–211. doi: 10.1016/j.joes.2016.07.001

Crossref Full Text | Google Scholar

Orhan, H., Eyduran, E., Tatlýyer, A., Saygýcý, H. (2016). Prediction of egg weight from egg quality characteristics via ridge regression and regression tree methods. Rev. Bras. Zootecnia 45, 380–385. doi: 10.1590/S1806-92902016000700004

Crossref Full Text | Google Scholar

Özkaynak, İ. (1980). Yem bezelyesi (Pisum arvense L.) yerel çeşitler üzerine seleksiyon ıslah çalışmaları. Ankara Üniversitesi Ziraat Fakültesi Yem bitkileri Çayır ve Mera Kürsüsü Ulucan Matbaası.

Google Scholar

Ratner, B. (2017). “Statistical and machine-learning data mining,” in Techniques for Better Predictive Modeling and Analysis of Big Data, 3rd ed. (USA: CRC Press, Taylor and Francis Group, LLC).

Google Scholar

R Core Team (2022). R: A Language and Environment for Statistical Computing (Vienna, Austria: R Foundation for Statistical Computing). Available at: https://www.R-project.org/.

Google Scholar

Rumelhart, D. E., Hinton, G. E., Williams, R. J. (1986). Parallel Distributed Processing (Cambridge: MIT Press), 318–362.

Google Scholar

Salford Systems (2001a). “Multivariate Adaptive Regression Splines (MARS): user guide. Chapter 3,” in MARS Basics - Smoothing, splines and knot selection (San Diego, California), 9–34.

Google Scholar

Salford Systems (2001b). TreeNet stochastic gradient boosting: An implementation of the MART methodology (San Diego, California).

Google Scholar

Samarasinghe, S. (2007). “Neural networks for applied science and engineering,” in From Fundamentals to Complex Pattern Recognition (Auerbach Neural Publications, Boca Raton, New York).

Google Scholar

Sarıkaya, M. F., İleri, O., Erkovan, Ş., Erkovan, H.İ., Koç, A. (2023). Growing forage pea (Pisum arvense L.) for hay: Different sowing dates and plant densities in Central Anatolia. Atatürk Üniversitesi Ziraat Fakültesi Dergisi 54, 75–80. doi: 10.5152/AUAF.2023.22067

Crossref Full Text | Google Scholar

Sharma, S., Sharma, S., Athaiya, A. (2020). Activation functions in neural networks. Int. J. Eng. Appl. Sci. Technol. 4, 310–316. doi: 10.33564/IJEAST.2020.v04i12.054

Crossref Full Text | Google Scholar

Sibi, P., Allwyn Jones, S., Siddarth, P. (2013). Analysis of different activation functions using back propagation neural networks. J. Theor. Appl. Inf. Technol. 47, 1264–1268.

Google Scholar

Stanton, T. L., LeValley, S. B. (2006). Feed composition for cattle and sheep (USA: Colorado State University Extension Service).

Google Scholar

Steingberg, D., Colla, P. (2016). (CART). Classification and Regression Trees Vol. 2016 (San Diego, CA, USA: Salford Systems).

Google Scholar

Tadeusiewicz, R., Lula, P. (2007). Neural Networks (Kraków: StatSoft Poland).

Google Scholar

Takma, C., Atil, H., Aksakal, V. (2012). Comparison of multiple linear regression and artificial neural network models goodness of fit to lactation milk yields. Kafkas Üniversitesi Veteriner Fakültesi Dergisi 18, 941–944. doi: 10.9775/kvfd.2012.6764

Crossref Full Text | Google Scholar

Uzun, A., Gün, H., Açıkgöz, E. (2012). Farklı Gelişme Dönemlerinde Biçilen Bazı Yem Bezelyesi (Pisum sativum L.) Çeşitlerinin Ot, Tohum ve Ham Protein Verimlerinin Belirlenmesi. U. Ü. Ziraat Fakültesi Dergisi 26, 27–38.

Google Scholar

Victor, ȚÎ.Ț.E. I. (2022). The biochemical composition of some annual fabaceae species and their potential application in Moldova. Agron. Ser. Sci. Research/Lucrări Ştiinţifice Seria Agronomie 65, 183–188.

Google Scholar

Willmott, C., Matsuura, K. (2005). Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance. Climate Res. 30, 79–82. doi: 10.3354/cr030079

Crossref Full Text | Google Scholar

Yan, W., Tinker, N. A. (2005). A biplot approach for investigating QTL-by-environment patterns. Mol. Breed. 15, 31–43.

Google Scholar

Yordanova, A., Gocheva-Ilieva, S., Kulina, H., Yordanova, L., Marinov, I. (2015). Classification and regression tree analysis in modeling the milk yield and conformation traits for Holstein cows in Bulgaria. Agric. Sci. Technol. 7, 208–213.

Google Scholar