The Effects of Winter Cover Crops on Maize Yield and Crop Performance in Semiarid Conditions—Artiﬁcial Neural Network Approach

: Maize is the most widespread and, along with wheat, the most important staple crop in the Republic of Serbia, which is of great signiﬁcance for ensuring national food security. With the increasing demand for food and forage, intensive agricultural practices have been adopted in the maize production systems. In this direction, considerable research efforts have been made to examine the effects of different types of cover crops as a green manure on maize productivity; however, no consistent conclusions have been reached so far. Therefore, the objective of the present study is to examine the possibility of predicting the effects of winter cover crops (CC) integrated with different management practices on the morphological traits, yield, and yield components of maize. The experiment was carried out on chernozem soil from 2016 to 2020 as a randomized complete block design arranged as a split-split-plot with three replicates. The pea as a sole crop (P) and the mixture of pea and triticale (PT) are sown as winter CC with the following subplots: (i) CC used as green manure, and (ii) CC used as forage and removed before maize sowing. The artiﬁcial neural network is used for exploring nonlinear functions of the tested parameters and 13 categorical input variables for modeling according to the following factors: CC, way of using CC, N fertilization, and year. The computed maximums of plant height, number of leaves, number of internodes, plant density, number of ears, grain yield, 1000-grain weight, hectolitre weight, dry matter harvest residue, harvest index, leaves percentage, stems percentage, and ears percentage are as follows: 232.3 cm; 9.7; 10.2; 54,340 plants ha − 1 ; 0.9; 9.8 t ha − 1 ; 272.4 g; 67.0 kg HL − 1 ; 9.2 t ha − 1 ; 0.52; 18.9%; 36.0%, and 45.1%, respectively. The optimal result is obtained with peas used as green manure, with 50 kg N ha − 1 and in the climatic conditions of 2018. Consequently, maize production under subsequent sowing periods can be successfully optimized by adapting selected management options for higher yield accomplishment.


Introduction
The main challenge for agriculture in the 21st century will be to increase food production with minimal environmental impact [1,2]. Therefore, it is necessary to analyze the current approach to agriculture and gradually introduce new solutions and technologies. Conventional farming over a long time has been proven to lead to soil degradation, reducing soil buffering capacity in the variable environmental conditions [3], and increasing sensitivity to extreme weather events [4]. This coincides with the state of the arable land-use systems in the Pannonian Plain, where significant changes in soil properties have been reported (soil organic matter loss, deterioration of physical and biological properties, salinization, etc.) [5][6][7]. An additional concern arises from the limited use of manure, which part of this study, the solution of the MOO was estimated by introducing a Pareto optimal method [39].
Based on all of the above, the goal of this study was to examine the possibility of predicting morphological traits, grain yield and yield components, and some crop performance of maize; i.e., the following 13 variables: the plant height (H), number of leaves (NL), number of internodes (NI), plant density (PD), number of ears (NE), grain yield (GY), 1000-grain weight (TGW), hectolitre weight (HW), dry matter harvest residue (DMHR), harvest index (HI), leaves percentage (L), stems percentage (S) and ears percentage (E); as a function of cover crops type (CC), way of using cover crops (WUCC), N fertilization (NF), and year of maize production (YEAR) under semiarid conditions of the experimental site.

Experimental Site Description-Location, Climate and Soil Characteristics
The research was carried out in the period from 2016 to 2020, at the Rimski Šančevi experimental station (45 • 19 N; 19 • 50 E; 80 m a.s.l.) of the Institute of Field and Vegetable Crops in Novi Sad; the National Institute of the Republic of Serbia (in Vojvodina, Northern Province of Serbia), in the typical chernozem zone of the southern part of the Pannonian Plain ( Figure 1). According to the WRB classification (2014), the soil type is medium-deep chernozem, formed on loess and loess-like sediments; is classified as Calcareous Chernozem (aric, loamic, pachic), abbreviated as CH-cc-ai.lo.ph.
The climate is characterized as a moderate continental climate with extreme seasonal variability in temperature and precipitation. Based on long-term data , the mean annual air temperature is 11.4 • C and the total annual precipitation sum is 640.5 mm. Monthly precipitation and temperature data were collected from the weather station at Rimski Šančevi experimental field ( Figure 1).

Weather Conditions during the Research Period
During the investigated period, temperatures, and precipitation showed differences among investigated growing seasons and deviations from the long-term averages (LTA) (Figure 2). At the beginning of the experiment, from December 2016 to March 2017, the lower precipitation was recorded compared with the LTA . That caused a shortage of water in the soil, and drought continued in summer 2017 (July-August), while in autumn 2017 precipitation increased. Temperatures in the period October-September 2017/18 were higher by 1.4 • C compared with LTA. The amounts of precipitation during the winter period compensated for the lack of soil moisture in 2017 and provided sufficient soil moisture during the spring season of 2018. Total precipitation in the hydrological year 2017/18 (753.4 mm) was above the average values for the experimental site. The small amount of precipitation (7.4 mm) at the beginning of the hydrological year 2018/19 (October) resulted in decreased soil moisture in autumn 2018. In February and March, a period with a lower amount of precipitation continued, and from the spring the precipitation increased, particularly in May (147.6 mm). Total precipitation for the hydrological year 2018/19 was lower by 42.9 mm than the LTA. Based on the meteorological conditions, 2019 was the warmest in the last 50 years, with an average annual temperature of 13.4 • C. During the hydrological year of 2019/20, the amount of precipitation was 72.5 mm higher than LTA. In June and August, the amount of precipitation (163 and 138 mm, respectively) was significantly above the average values. Significant deviations from the monthly temperatures were recorded in November and December 2019, as well as in February 2020. The warmest month in 2020 was August, with an average monthly temperature of 24.1 • C. Figure 1. Location of the trail and experimental design showing the main analyzed factors: cover crops (CC) type and way of using CC (PT: winter pea + triticale, for green manure and for forage; P: winter pea as a sole crop, for green manure, and for forage; Ø : control treatment-without CC), and N fertilization treatments (N50: with N fertilizer of 50 kg ha −1 , NØ : without N fertilizer application).
The climate is characterized as a moderate continental climate with extreme seasonal variability in temperature and precipitation. Based on long-term data , the mean annual air temperature is 11.4 °C and the total annual precipitation sum is 640.5 mm. Monthly precipitation and temperature data were collected from the weather station at Rimski Šančevi experimental field ( Figure 1).

Weather Conditions during the Research Period
During the investigated period, temperatures, and precipitation showed differences among investigated growing seasons and deviations from the long-term averages (LTA) (Figure 2). At the beginning of the experiment, from December 2016 to March 2017, the lower precipitation was recorded compared with the LTA (1970-2017). That caused a Figure 1. Location of the trail and experimental design showing the main analyzed factors: cover crops (CC) type and way of using CC (PT: winter pea + triticale, for green manure and for forage; P: winter pea as a sole crop, for green manure, and for forage; Ø: control treatment-without CC), and N fertilization treatments (N50: with N fertilizer of 50 kg ha −1 , NØ: without N fertilizer application).
was the warmest in the last 50 years, with an average annual temperature of 13.4 °C . During the hydrological year of 2019/20, the amount of precipitation was 72.5 mm higher than LTA. In June and August, the amount of precipitation (163 and 138 mm, respectively) was significantly above the average values. Significant deviations from the monthly temperatures were recorded in November and December 2019, as well as in February 2020. The warmest month in 2020 was August, with an average monthly temperature of 24.1 °C .

Experimental Design
The experimental plots were established using a randomized complete block design (RCBD) arranged as a split-split-plot, with 3 replicates (Figure 1). Maize was grown in a 3-year rotation with Sudan grass and soybean. Cover crops (CC) were considered as the main plots, which consisted of the following: (a) the winter pea (Pisum sativum ssp. Arvense L.) (P), (b) the mixture of winter pea + triticale (×Triticosecale) (PT), and (c) control treatment (Ø, without CC). The CC plots were divided into the following two sub-plots (way of using cover crops, WUCC): in the first sub-plot, winter CC was used as green manure with plowing, while in the second, the cover crops were mowed and forage was taken away. Based on above factors, the trial anticipated 5 treatments as follows: (i) winter pea + triticale (PT) for green manure, (ii) winter pea + triticale (PT) for forage, (iii) winter pea as a sole crop (P) for green manure, (iv) winter pea (P) for forage, and (v) control (Ø)-

Experimental Design
The experimental plots were established using a randomized complete block design (RCBD) arranged as a split-split-plot, with 3 replicates (Figure 1). Maize was grown in a 3-year rotation with Sudan grass and soybean. Cover crops (CC) were considered as the main plots, which consisted of the following: (a) the winter pea (Pisum sativum ssp. Arvense L.) (P), (b) the mixture of winter pea + triticale (×Triticosecale) (PT), and (c) control treatment (Ø, without CC). The CC plots were divided into the following two sub-plots (way of using cover crops, WUCC): in the first sub-plot, winter CC was used as green manure with plowing, while in the second, the cover crops were mowed and forage was taken away. Based on above factors, the trial anticipated 5 treatments as follows: (i) winter pea + triticale (PT) for green manure, (ii) winter pea + triticale (PT) for forage, (iii) winter pea as a sole crop (P) for green manure, (iv) winter pea (P) for forage, and (v) control (Ø)-without CC. All sub-plots were additionally divided into two sub-subplots (N fertilization), namely, with nitrogen fertilizer of 50 kg ha −1 (N50), and without N fertilizer application (NØ). The dimensions of each individual basic plot were 6 × 4 m, i.e., 24 m 2 . Winter CC was sown in autumn, in the last week of October/beginning of November, and cutting for forage and plowing as green manure was performed in May. After plowing and seedbed preparation, in the first decade of June, maize (hybrid NS-4051) was sown at a distance of 23 cm between plants in a row and 70 cm between rows, with a seeding rate of 62,000 plants ha −1 . Nitrogen (Urea, 46% N) was applied in maize vegetative phase BBCH 34, and the harvest was performed in October. The description of field operations by analyzed years is shown in Table 1.

Measurements and Analytical Determination
The disturbed field moist soil samples (1 kg of soil per plot) were collected from the topsoil (0-30 cm) in autumn, after maize harvest ( Table 2). The pH value of the soil was determined in the suspension of soil in KCl and H 2 O, by Metrel MA 3657 pH meter. Humus content was determined by oxidizing organic matter with potassium bichromate [47]. Calcium carbonate (CaCO 3 ) content was determined volumetrically using Scheibler calcimeter. Mineral N in the soil was extracted using 2 M KCl (1:4 soil-to-solution ratio, on weight basis) and determined by steam distillation [48]. The available phosphorus (P) and potassium (K) content were measured by the Ammonium-Lactate (AL) method. The concentration of P 2 O 5 was measured by spectrophotometry, while the concentration of K 2 O was measured by flame photometry [49]. The dry matter yield of aboveground biomass of cover crops (t ha −1 ) was evaluated before incorporation by cutting the crop to a stubble height of 5 cm. Samples were collected from 1 m 2 from each basic CC plot and each replicate. Samples were taken from the central part of each plot. The dry matter yield of cover crops was obtained by drying samples to a constant mass at 105 • C in the oven. The morphological traits and yield components of maize were measured at the maize-harvesting time (October) by randomly choosing 15 plants from each plot in 3 repetitions. Grain yield was adjusted to a moisture content of 14%.

ANN Modelling
A multi-layer perception model (MLP), with three layers (input, hidden, and output) was implemented to construct 13 ANN models for the following output variables: H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E; according to the following factors: cover crops type (CC), way of using cover crops (WUCC), N fertilization (NF) and YEAR. This format of the ANN model is approved for its high potential of estimating nonlinear functions [50][51][52][53].
The maize crop performance database, which was employed for the 13 ANN models, was stochastically segmented into the following: training, cross-validation, and testing data (70%, 15%, and 15% of experimental data, respectively). The training data set was applied during the learning cycle of the ANNs calculation and also used to evaluate the optimal number of neurons in the hidden layer and compute the weight coefficient of individual neurons in the network [58]. The weight coefficients and biases connected to the hidden and output layers of the ANN models were recorded in matrices and vectors W 1 and B 1 , and W 2 and B 2 , individually. The neural network model can be displayed by the following matrix equation: where Y is the matrix of the outputs, f 1 and f 2 are transfer functions in the hidden and output layers, respectively, and X is the matrix of inputs [59]. The elements of matrices W 1 and W 2 for each ANN model were computed during the learning cycle, in which the elements are constantly introduced applying an optimization method to minimize the disagreement between the data and the models [54,60,61]. The BFGS algorithm was implemented to enhance the evaluation and stabilize the solution's convergence [62]. The coefficients of determination were utilized as parameters to monitor the execution of the achieved ANN model. The ANN model was created to foresee and optimize the following parameters: H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E, according to the following treatments (factors): CC, WUCC, NF, and YEAR.

Global Sensitivity Analysis
The Yoon's global sensitivity formula for the obtained ANN model was exploited to evaluate the relative influence of the input parameters on output variables, based on weight coefficients of the developed ANN models [63].
where: w-weight coefficient in ANN models, i-input variable, j-output variable, khidden neuron, n-number of hidden neurons, m-number of inputs.

Error Analysis
The numerical confirmation of the developed model was investigated applying the coefficient of determination (R 2 ), reduced chi-square (χ 2 ), mean bias error (MBE), root mean square error (RMSE), and mean percentage error (MPE). These frequently used parameters can be obtained according to the following equations [64]: where x exp,i marks the experimental values and x pre,i present value computed by the model, N and n are the number of observations and constants, accordingly.

Multi-Objective Optimization
The obtained ANN models were employed for multi-objective optimization (MOO) calculation, with the aim to gain the specific set of production parameters, which would extract the maximal values of PD, GY, TGW, HW, DMHR, and HI. The final result of MOO was derived using a Pareto front algorithm, which existed in the case of one objective function improvement without deteriorating the others [39]. The genetic algorithm (GA) was used to find the solutions to the MOO problem by a stochastic method inspired by natural evolution applying the mutation, selection, inheritance, and crossover [65]. For the MOO computation, MATLAB software (version 7.10.0 (2010), The MathWorks Inc., Natick, MA, USA) was used, and gamultiobj algorithm was used for MOO optimization. The primary population is formed by chance and then introduced to a set of points in the design area. The populations of the next generations were determined using distance measures and non-dominated ranking of the particular points within the existing generation [39,45,46]. Modeling, statistical analysis, and multi-objective optimization of the collected data in this study were processed statistically using the software package STATISTICA version 14.0. (TIBCO Software Inc., Palo Alto, CA, USA).

Dry Matter Yield of Cover Crops
The dry matter yield of cover crops differed and depended on a complex interaction of years and CC. Due to favorable weather conditions, the highest dry matter yield of CC was achieved in 2020 with the mixture of PT (4.4 t ha −1 ) and was higher by 0.9 t ha −1 compared to the sole pea crop (Figure 3). The lowest yield of CC dry matter was achieved in 2017, in the pea crop (1.2 t ha −1 ), which differed by 0.5 t ha −1 in relation to the mix PT. In the years 2018, 2019, and 2020, the CC yield was higher than in 2017. The yield of PT was higher by 1.7 t ha −1 in 2018, i.e., by 2.7 t ha −1 in 2020 compared to the yield in 2017.

ANN Model
The attained ANN models showed sufficient generalization ability for experimental data prediction. Based on the ANN models' performance, the optimal number of neurons in the hidden layer for calculation of the following selected parameters: H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E, the prediction was between 5 and 10 to attain the required high values of R 2 (in the range from 0.487 to 0.966, for the training period) and as low as possible sum of squares (SS) values (Tables 3 and 4).

ANN Model
The attained ANN models showed sufficient generalization ability for experimental data prediction. Based on the ANN models' performance, the optimal number of neurons in the hidden layer for calculation of the following selected parameters: H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E, the prediction was between 5 and 10 to attain the required high values of R 2 (in the range from 0.487 to 0.966, for the training period) and as low as possible sum of squares (SS) values (Tables 3 and 4).  The goodness of fit among experimental computations and model estimated outputs, described as the ANN models' performance (sum of R 2 within measured and calculated parameters), through training, testing, and validation actions, are displayed in Table 4.
The used ANN models predicted the experimental variables quite well for a wide range of the process parameters (which can be seen in Figures 4 and 5, where the experimentally estimated and predicted values of the ANN models are shown). The goodness of fit among experimental computations and model estimated outputs, described as the ANN models' performance (sum of R 2 within measured and calculated parameters), through training, testing, and validation actions, are displayed in Table 4.
The used ANN models predicted the experimental variables quite well for a wide range of the process parameters (which can be seen in Figures 4 and 5, where the experimentally estimated and predicted values of the ANN models are shown).

Global Sensitivity Analysis-Yoon's Interpretation Method
In this segment, the impact of input variables (i.e., factors CC, WUCC, NF, and YEAR) on H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E, throughout the maize production, the Yoon's interpretation method of the generated ANN models were analyzed. The graphical display of the ANN model results is shown in Figure 6.

Global Sensitivity Analysis-Yoon's Interpretation Method
In this segment, the impact of input variables (i.e., factors CC, WUCC, NF, and YEAR) on H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E, throughout the maize production, the Yoon's interpretation method of the generated ANN models were analyzed. The graphical display of the ANN model results is shown in Figure 6. As presented in Figure 6a,f, 2017 is the most negatively influential parameter on H and GY, with a relative importance of −40.78% and −23.69%, respectively; while the influence of the year 2018 on the NL was quite the opposite, with a relative influence of +24.23% (Figure 6b). The arid period during the summer months of 2017 (Figure 2a) affected the slow growth and development of maize plants, as well as grain yield. The results show that optimal conditions for maize production can be achieved if winter peas are incorporated into the soil as green manure, along with nitrogen fertilization (50 kg N ha −1 ), and in years with a favorable precipitation distribution; such as in 2018. NI was mainly influenced by the 2019 year, with a relative importance of −22.64% (Figure 6c), while the year 2020 was the most significant parameter for plant density (Figure 6d), with a relative influence of 30.68%. On the other hand, according to Figure 6e, the number of ears was negatively influenced by the mixture of pea and triticale, using cover crops as green manure, and the year 2020, with a relative importance of −18.36%, −13.11%, and −13.02%, respectively.
The maize yield dependence on the investigated factors is reflected in 1000-grain weight and hectolitre weight. According to Figure 7a,c,f, the year 2017 had a negative effect on 1000-grain weight, dry matter harvest residue, and stem percentage, with a relative importance of −33.48%, −23.02%, and −26.60%, respectively. The year 2019 had a positive impact on the hectolitre weight (+16.55%) and ear percentage (+29.14%), while it showed As presented in Figure 6a,f, 2017 is the most negatively influential parameter on H and GY, with a relative importance of −40.78% and −23.69%, respectively; while the influence of the year 2018 on the NL was quite the opposite, with a relative influence of +24.23% (Figure 6b). The arid period during the summer months of 2017 (Figure 2a) affected the slow growth and development of maize plants, as well as grain yield. The results show that optimal conditions for maize production can be achieved if winter peas are incorporated into the soil as green manure, along with nitrogen fertilization (50 kg N ha −1 ), and in years with a favorable precipitation distribution; such as in 2018. NI was mainly influenced by the 2019 year, with a relative importance of −22.64% (Figure 6c), while the year 2020 was the most significant parameter for plant density (Figure 6d), with a relative influence of 30.68%. On the other hand, according to Figure 6e, the number of ears was negatively influenced by the mixture of pea and triticale, using cover crops as green manure, and the year 2020, with a relative importance of −18.36%, −13.11%, and −13.02%, respectively.
The maize yield dependence on the investigated factors is reflected in 1000-grain weight and hectolitre weight. According to Figure 7a,c,f, the year 2017 had a negative effect on 1000-grain weight, dry matter harvest residue, and stem percentage, with a relative importance of −33.48%, −23.02%, and −26.60%, respectively. The year 2019 had a positive impact on the hectolitre weight (+16.55%) and ear percentage (+29.14%), while it showed a negative impact of −23.18% on the harvest index (Figure 7b,d,g). High temperatures affected the 1000-grain weight due to the rapid ripening in dry years such as in 2017. The influence of peas as a cover crop on leaf percentage was negative, with a relative importance of −27.55% (Figure 7e). The result indicates higher effects of the winter pea treatment on the ear and stem percentage, which is made possible by the greater availability of soil moisture and nitrogen.  (Figure 7b,d,g). High temperatures affected the 1000-grain weight due to the rapid ripening in dry years such as in 2017. The influence of peas as a cover crop on leaf percentage was negative, with a relative importance of −27.55% (Figure 7e). The result indicates higher effects of the winter pea treatment on the ear and stem percentage, which is made possible by the greater availability of soil moisture and nitrogen.

Multi-Objective Optimization (MOO) of the Outputs of ANN
One of the main goals of this research was to optimize PD, GY, TGW, HW, DMHR, and HI throughout different field management practices and years, synchronously, employing the ANN models and varying input variables. These numerical assignments were solved for the ANN models by applying the MOO computation in Matlab. The MOO method was set to obtain the best combinations of different parameters by maximizing the output variables in the ANN models. The optimization method's constraints were used in the experimental series of parameters. The number of generations achieved was 723 for the ANN models, while the population dimension was set to 200 for all input variables. Thus, the number of points on the Pareto front was 69 for the ANN models. The computed maximums of H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E were the following: 232.3 cm; 9.7; 10.2; 54.340 plants ha −1 ; 0.9; 9.8 t ha −1 ; 272.4 g; 67.0 kg HL −1 ; 9.2 t ha −1 ; 0.52; 18.9%; 36.0%, and 45.1%, respectively. The optimal result was obtained with the following parameters: peas used as green manure, with an application of 50 kg N ha −1 , and in the climatic conditions of 2018.

Discussion
Our study showed that climatic conditions significantly modified overall cover crop performance while increasing the sensitivity of the main crop to weather conditions ( Figure 3). This is particularly noticeable in 2017, when the lack of precipitation (first during the winter period, and later during the spring) suppressed their growth, and decreased the yield. During the four years of research, the average dry matter yield of cover crops was higher in the mixture of pea and triticale (3.4 t ha −1 ) compared to peas as a sole crop (2.1 ha −1 ) (Figure 3). Similar to our results, Ref. [66] states that mixed cover crops have a higher forage yield and a better nutrient balance for ruminant feed than sole cropping. On the other hand, the plowing of a cereal/legume mixture results in a larger amount of dry matter incorporated into the soil as green manure [67], which affects the soil's physical properties and soil air regime [68].
The selection of proper cover crops and timely incorporation of winter CCs can increase or adversely affect the crop yield of the subsequent crop in rotation. Chen et al. [69] registered an increased maize yield from 1.60 to 2.43 t ha −1 following CC, which is consistent with the results of our study. Increased maize yield was also obtained in the research of [70], in which a mean increase of grain yield was 78, 91, and 66% with the inclusion of cowpea, pigeon pea, and hemp, compared with the fallow system. On the contrary, Ref. [71] reported a decrease in the yield of cash crops (maize and sunflower) by 0.5-3.0 t ha −1 . In our study, the sowing of maize was performed after the recommended date to give CCs more time to produce higher biomass. Currently, there is a lack of information on crop cultivation in the subsequent sowing period in the temperate conditions of Vojvodina province. Nevertheless, the assumption is that hybrids of shorter FAO maturity groups can be established significantly later than the optimal term with acceptable risks. Mahama et al. [70] confirm that the sowing delay of main summer crops is related to the necessity to provide time for winter CCs growth and higher biomass production, and in the case of legumes, to accumulate N. However, in this study, it was shown that this practice could be successfully implemented for the winter cover crops. Out of four years of research; only in 2017 arid periods were recorded in July and August with 29.4 mm of precipitation (90 mm less than the LTA).
Consequently, the yield components of maize were lower and affected the yield accomplishment. In conditions with uncertain amounts of precipitation necessary for the plants' growth and development, it is essential to make a trade-off between the benefits provided by CCs and the possibility of obtaining a reduced yield of the main crop. Given that the average maize yield in our experiment is not considerably lower than the average grain yield of the recommended sowing dates, this approach could be a crucial point in designing future cropping systems in semiarid conditions. ANN models can help predict maize behavior under our experimental setup. In most cases, the predicted values were approaching the desired R 2 value for the ANN models. The SS achieved by the ANN models is of the same order of magnitude as experimental errors for H, NL, NI, PD, NE, GY, TGW, HW, DMHR, HI, L, S, and E, which were also observed in similar studies [54,62,72]. The ANN protocols are challenging (71-141 weights-biases) due to the high nonlinearity of the studied system [54,73]. Nevertheless, the character of the ANN models fit as observed in Table 4, where χ 2 , MBE, RMSE, and MPE were lower [64]. The residual analysis of the developed model was additionally conducted to provide insight into maize response to key variables and the limitations of ANN analyses. In addition, skewness was used to estimate the deviation of the distribution from regular symmetry. Skewness showed different values for some variables (NI, L, H, HI, E, etc.), indicating asymmetrical distributions, although normal distributions are ideally symmetrical. Accordingly, these indicators are developed as a consequence of complex interactions between the observed parameters. The data set was checked for kurtosis (the "peakeness" of a distribution). In our study, leaf and ear percentages are distinct in kurtosis values compared with other parameters, indicating heavier tails than normal distribution, with high peaks and outliers. This can be explained by the occurrence of smaller maize plants in which sudden maturation occurs in specific years. A high R 2 is suggestive that the variation was estimated, and that the data fitted satisfactorily to the suggested model [74,75]. Marcillo and Miguez [15] in a meta-analysis of maize yield dependence on intercrops, concluded that CC from the Poaceae family neither positively nor negatively affected maize yield, while leguminous plants and their mixture showed a significant impact. Similar results were observed in our experiment as CC and years showed positive effects on the yield of subsequent maize crops.
In our study, the smaller number of ears of maize can be explained by the soil's lower moisture content on the plots with the combined sowing of winter peas and triticale. However, this is contrary to the study of [76], who determined higher soil moisture after legume CCs compared with the mixture with cereal, while the lowest moisture content was in the soil after triticale. As shown in Figure 6, nitrogen fertilization had the lowest impact on the plant height, number of leaves and internodes, plant density, number of ears, and grain yield, while years were the most influential. The assumption is that fresh biomass mineralization can be sufficient for the N-demand of latter-sown maize. Refs. [77,78] found that plant density (the number of plants per unit area) is a decisive factor in maize production, but the optimal plant density is not always the same, but rather varies depending on the year. In recent decades, maize grain yield has become increasingly dependent on climatic conditions during plant growth and development [79]. Therefore, predicting optimal plant density represents a challenging task. The maize yield in the subsequent sowing directly depends on the water spent by the winter CCs and precipitation distribution during subsequent crop cultivation [80,81]. According to [82], CCs may have a more significant potential to reduce drought stress in maize after longterm use in systems with less soil disturbance. Rosa et al. [83] found that CCs in semi-arid conditions did not contribute to the increase in maize grain yield, and most CC types are associated with a reduction in maize yield. The same authors relate this consequence to the lack of soil moisture and nitrogen availability during the maize growing season. Despite many benefits, ref. [84] points out that the benefits of CCs in practice are still limited because the effects on productivity and economic return are variable. In addition to that, for conclusive CCs assessment and compliance with main crops, it requires multi-location trials and ideotype selection of both CC and crop sequence. In the work of [85], ANN was used to predict the role of individual nutrients on various parameters of rice plants. Contrary to our results, the authors determined that nitrogen is the most limiting nutrient among the studied nutrients in terms of obtaining the maximum grain yield of rice, regardless of the season and variety. They observed an improvement in plant height, number of tillers, and dry matter, indicating an important role of N in vegetative growth. According to Figure 2, there are differences in the amount of precipitation and their distribution during the growing period of maize, with 2018 and 2020 being higher in precipitation for the period June-September. The authors [85,86] point out that environmental variables, such as air temperature, total precipitation, insolation, and soil properties, significantly influence plant parameters and crop performance prediction using ANN models.
The main limitations of this and similar studies are that, although each of the analyzed treatments (factors) in ANN models has a specific role in the growth and development of plants, complex interactions among them often make it difficult to understand the separate actions of individual factors. Moreover, it is almost impossible to fully assess the individual effects of input parameters on the final performance of plants and crops. This is particularly pronounced in field conditions, where crop performance is subjected to a very complex environment, with sudden variations in climatic parameters such as temperature, precipitation, relative air humidity, etc. In addition, statistical models are often criticized for failing to provide a scientific understanding of the processes being studied.

Conclusions
This study indicates that empirical artificial neural network models could be successfully utilized to predict plant-examined parameters. The prediction of those variables using parameters such as cover crops, way of using cover crops, N fertilization, and year showed high accuracy in estimating with the artificial neural network model. The obtained models provided a good fit to experimental data and were adequate to predict the output variables successfully, showing a reasonably good predictive capability (overall R 2 for plant height, number of leaves, number of internodes, plant density, number of ears, grain yield, 1000-grain weight, hectolitre weight, dry matter harvest residue, harvest index, leaf percentage, stem percentage, and ear percentage were 0.791; 0.791; 0.878; 0.960; 0.818; 0.700; 0.843; 0.691, accordingly). The computed maximums of plant height were 232.3 cm, number of leaves 9.7, number of internodes 10.2, plant density 54,340 plant ha −1 , number of ears 0.9, grain yield 9.8 t ha −1 , 1000-grain weight 272.4 g, hectolitre weight 67.0 kg HL −1 , dry matter harvest residue 9.2 t ha −1 , harvest index 0.52, percentage of leaf 18.9%, stem 36%, and ear 45.1%, respectively. Our study implies that subsequent maize production in rainfed conditions can be performed with winter pea cover crops used for green manure, supplementary nitrogen addition of 50 kg ha −1 , and under a rainfall amount of >330 mm (June-September). The developed mathematical models give satisfactory accuracy for potential practical application in maize production and enable the proper selection of cropping management for the farmers. Given the scope of this work, it would be essential to suggest cover crops as necessary agroecological measures that would amplify the positive outcomes of other field management practices in field crop production.