Smart non-destructive prediction of antioxidant, mineral, and essential oil traits in Persian parsley (Petroselinum crispum Mill.) landraces

Abstract

An integrative principal component analysis–artificial neural network (PCA–ANN) framework was developed to characterize ecotypic variation among parsley landraces and predict quality-related traits in medicinal and aromatic plants. Fifteen Iranian parsley landraces collected from diverse agro-ecological regions, together with two commercial cultivars as reference genotypes, were analyzed to establish predictive links between easily measurable morphological traits and key biochemical, mineral, and essential-oil (EO) characteristics. Twenty-one independent morphological variables were recorded and used as model inputs. To minimize redundancy and multicollinearity, PCA was applied exclusively to the morphological dataset, reducing it to a smaller set of uncorrelated components that preserved most of the variance. These components served as input features for optimized ANN architectures developed to predict antioxidant properties, EO yield and composition, and mineral nutrient content. The resulting PCA–ANN framework achieved strong predictive performance, with R² up to 0.94. It accurately predicted antioxidant, mineral, and compositional profiles from morphological traits alone, demonstrating the potential of morphological phenotyping as a rapid, non-destructive proxy for complex chemical analyses. This integrative modeling approach reduces reliance on time-consuming and costly procedures such as GC–MS and offers a practical decision-support tool for genotype selection, breeding, and quality evaluation in medicinal and aromatic crops. The proposed framework provides a scalable, data-driven strategy for advancing precision agriculture and sustainable management of herbal plant resources.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12870-026-09360-6.

Introduction

Agriculture, as a cornerstone of sustainable development, plays a vital role in safeguarding global food security and maintaining economic stability. Its importance has become increasingly evident as humanity faces pressing challenges such as climate change, rapid population growth, and the depletion of natural resources [1]. Within this context, medicinal and aromatic plants such as parsley (Petroselinum crispum Mill.) occupy a prominent niche because they are rich in phytochemical compounds with diverse applications in the pharmaceutical, nutraceutical, and food industries [2].

Parsley is characterized by a complex phytochemical profile that underpins its recognized biological activities. Its tissues accumulate flavonoids (e.g., apigenin, kaempferol), phenolic acids, carotenoids, ascorbic acid, and essential oils rich in apiol and myristicin. These metabolites act synergistically in redox regulation, membrane stabilization, and signaling modulation, conferring strong antioxidant capacity and contributing to anti-inflammatory and antimicrobial effects [3–5]. Evidence further indicates reno-protective, hepatoprotective, and hypoglycemic effects, reinforcing parsley’s relevance as both a functional food and a medicinal resource [6–8].

From a physiological perspective, the biosynthesis and accumulation of these secondary metabolites are tightly linked to carbon allocation, photosynthetic performance, nutrient availability, and stress signaling pathways. Morphological vigor (e.g., canopy development, leaf architecture, biomass accumulation) influences light interception and assimilate partitioning, which in turn regulate the phenylpropanoid and terpenoid biosynthetic pathways governing phenolic and essential-oil production. Consequently, genotype-driven variation in structural and growth traits may serve as a biologically meaningful proxy for metabolic capacity.

Because these bioactive compounds are strongly modulated by genetic background, agro-ecological origin, and local adaptation history, substantial variability exists among parsley landraces in both concentration and composition. Capturing this variability through integrative, data-driven approaches is therefore essential. Linking easily measurable morphological descriptors with complex biochemical and compositional traits provides a mechanistically informed and scalable framework for rapid quality assessment, precision breeding, and intelligent management of this economically important aromatic crop.

In addition to its phytochemical richness, parsley exhibits remarkable genetic diversity across Mediterranean and semi-arid environments, particularly in Iran, where indigenous landraces represent a valuable reservoir for breeding programs targeting yield, quality, and stress resilience [8]. Similar ecotypic variation associated with agro-ecological origin has been reported in other medicinal and aromatic crops, including saffron (Crocus sativus L.), where multivariate analyses revealed substantial differences in vegetative performance among populations originating from diverse agro-climatic regions [9]. Furthermore, the evaluation of genetic resources across multiple environments has proven valuable for identifying superior genotypes and supporting breeding programs in medicinal and aromatic crops, as demonstrated in saffron through multi-site assessments of clonal productivity and adaptation [10].

Accurate assessment of biochemical traits, including phenolics, flavonoids, ascorbic acid (vitamin C), and carotenoids, together with mineral content and essential-oil (EO) yield and composition, remains a complex task. This complexity arises from the intricate and dynamic interactions among genetic, environmental, and agronomic factors that collectively influence plant performance and the biosynthetic pathways of secondary metabolites [4]. Traditional methods for evaluating plant traits, such as field experimentation and biochemical assays, are empirically valuable but often time-consuming, labor-intensive, and costly. They also exhibit limited scalability and analytical precision when applied to large, heterogeneous populations [11]. Moreover, conventional statistical approaches frequently fail to capture the nonlinear relationships and multifactorial interactions that characterize plant responses to stressors such as drought, salinity, and nutrient deficiency, conditions that are increasingly prevalent in arid and semi-arid regions [12]. Morphological variability in parsley leads to substantial fluctuations in biomass production, secondary-metabolite accumulation, and overall market value [13]. Consequently, identifying superior genotypes requires advanced analytical and predictive tools capable of integrating multisource data encompassing morphological, biochemical, and environmental parameters. Such integrative frameworks can enhance the accuracy of plant-performance predictions under variable climatic and management conditions and support the advancement of precision agriculture [14].

Computational and machine-learning approaches have recently emerged as powerful tools for modeling agricultural systems. These algorithms can process large, multivariate datasets and capture nonlinear interactions among variables that traditional statistical models often overlook [11, 15]. Recent studies have further demonstrated the potential of artificial intelligence and machine-learning approaches, including hybrid neural-network and optimization algorithms, to predict antioxidant, mineral, and growth-related traits and to decode complex relationships among plant phenotypic and nutritional characteristics, highlighting their value for precision phenotyping, quality assessment, and biological interpretation in horticultural crops [16, 17]. Among them, artificial neural networks (ANNs) are particularly effective because they can approximate complex relationships between inputs and outputs through adaptive learning. Their multilayer architecture enables the modeling of nonlinear, high-dimensional relationships between predictors and response variables. Training occurs through iterative error minimization, allowing the model to self-adjust and optimize prediction accuracy [18]. Such flexibility makes ANNs especially suitable for predicting plant traits that are influenced by multiple, interdependent factors, where linear models such as regression or ANOVA are constrained by statistical assumptions [11].

The use of ANNs in agriculture has expanded rapidly for yield prediction, stress detection, and quality assessment. Owing to their ability to process complex datasets, they have been widely used for predictive modeling in agricultural systems. For example, ANNs have successfully predicted maize and soybean yields by integrating soil, climatic, and historical data, demonstrating strong predictive performance in complex agricultural datasets [19]. In medicinal plants, ANNs have been applied to predict the EO composition of Satureja rechingeri Jamzad under stress conditions, integrating environmental and biochemical parameters to enhance precision [20]. Similar applications have been reported for aromatic species such as lemon balm and mint, where ANN models achieved R² values of approximately 0.90 when predicting phytochemical traits [21]. Despite these advances, most ANN studies have focused on staple crops, with limited attention to medicinal herbs, particularly native populations in biodiversity-rich regions such as Iran. Furthermore, few studies have integrated morphological, biochemical, and mineral traits within a single predictive framework, which limits the identification of superior genotypes for sustainable production.

Parsley’s biochemical complexity, reflected in its wide variation in EO composition and secondary metabolites such as apiol, myristicin, and linalool, directly influences its aroma, pharmacological efficacy, and market value. These quality-determining compounds are highly sensitive to genotype and environmental factors, leading to substantial variability in yield and composition among landraces. Consequently, accurate prediction of biochemical profiles is essential for cultivar improvement and quality standardization. Artificial intelligence (AI)-based predictive analytics, particularly ANN modeling, offer an effective means of interpreting such multifactorial variability and enabling smart phenotyping for precision cultivation, breeding, and value-chain optimization.

The present study addresses this gap by establishing a PCA–ANN framework to predict mineral content, antioxidant traits, and EO yield and composition in fifteen Iranian landraces of P. crispum, collected from ecologically diverse environments, together with two commercial cultivars as reference genotypes. These landraces, originating from contrasting agro-climatic zones ranging from lowland semi-arid plains to high-elevation temperate regions, represent a valuable genetic resource for investigating ecotypic variation associated with agro-ecological origin under a common experimental environment. By integrating multisource datasets encompassing morphological, biochemical, and compositional attributes, the proposed PCA–ANN model enhances predictive accuracy and supports precision breeding. It further guides data-driven agronomic management for sustainable and value-added cultivation of this economically and pharmacologically important herb. Therefore, this study developed and validated an integrative PCA–ANN framework to accurately predict antioxidant capacity, essential-oil composition, and mineral traits of parsley landraces using non-destructive morphological descriptors, thereby providing a data-driven approach for characterizing ecotypic variation among parsley landraces.

Materials and methods

Plant material and seed collection

Seeds of fifteen Iranian parsley (P. crispum Mill.) landraces were collected in March 2021 from diverse agro-ecological regions distributed across Iran (Table 1). The collection was carried out by trained personnel of Lorestan University in collaboration with local agricultural extension offices. Each landrace was sampled from smallholder farms that had cultivated parsley for several generations, thereby preserving the characteristics of locally adapted genotypes. Mature and healthy seeds were hand-harvested from at least twenty randomly selected plants per location to capture intra-population genetic diversity. All seeds were manually cleaned, air-dried at room temperature (25 ± 2 °C), and stored in paper envelopes under cool and dry conditions (4 °C, 35% relative humidity) until sowing.

Table 1.

Geographic origin, coordinates, and elevation of Iranian parsley (Petroselinum crispum Mill.) landraces collected for the present study. Two commercial cultivars were also included as reference genotypes for comparative analysis

Code	Collection site / Province	Latitude (° N)	Longitude (° E)	Elevation (m a.s.l.)	Agro-climatic characteristics
1	Dorud (Lorestan)	33.49	49.08	1 600	Mountainous, semi-humid, fertile loam soils
2	Saman (Chaharmahal-Bakhtiari)	32.45	50.91	2 100	High-elevation, cold winter, irrigated valleys
3	Tabriz (East Azerbaijan)	38.08	46.29	1 350	Semi-arid plateau, calcareous soils
4	Khuzestan (Ahvaz region)	31.32	48.69	90	Hot lowland plain, saline soils, desert margins
5	Pol Dokhtar (Lorestan)	33.15	47.72	400	Warm valley, low rainfall, alluvial soils
6	Kurdistan (Sanandaj area)	35.53	46.99	1 600	Temperate mountainous, clay-loam soils
7	Fasa (Fars)	28.93	53.65	1 250	Semi-arid basin, moderate temperature range
8	Tehran	35.68	51.38	1 200	Temperate-semi-arid, urban agricultural zone
9	Shahr-e Rey (Tehran Prov.)	35.60	51.45	1 500	Semi-arid plain, historical cultivation area
10	Shiraz (Fars)	29.62	52.53	1 500	Mild semi-arid climate, calcareous soils
11	Shahrekord (Chaharmahal-Bakhtiari)	32.32	50.85	2 100	Cold mountainous climate, high rainfall
12	Esfahan	32.65	51.68	1 590	Central plateau, arid climate, loamy soils
13	Hamadan	34.80	48.52	1 850	Cool-temperate upland, fertile volcanic soils
14	Arak (Markazi)	34.09	49.70	1 750	Continental climate, clay-loam soils
15	Bojnurd (North Khorasan)	37.47	57.33	1 070	Semi-humid steppe, foothill environment

The collection sites represent a wide range of environmental conditions, including lowland semi-arid plains, temperate valleys, and high-elevation mountain zones. Sampling across altitudes from 90 m in Khuzestan to more than 2100 m in Shahrekord and Saman ensured the inclusion of genotypes potentially adapted to contrasting temperature, moisture, and soil fertility regimes. This environmental diversity provides a valuable basis for evaluating ecotypic variation associated with agro-ecological origin under common field conditions and provides a diverse genetic background for development of the predictive ANN models.

No wild or protected populations were sampled, and no endangered species were involved. Consequently, special collection permits were not required under Iranian biodiversity or agricultural legislation. All procedures complied with the guidelines for plant genetic resource management issued by the Iranian Ministry of Agriculture (AREEO) and followed the principles of the Convention on Biological Diversity (CBD). Geographic coordinates, elevation, and local climatic data for each collection site were recorded using GPS and verified through the Iranian Meteorological Organization’s database.

In addition, seeds of two commercial cultivars, P. crispum var. crispum (flat-leaf, stress-tolerant) and P. crispum var. neapolitanum (curly-leaf, aromatic), were obtained from Richters Herbs (Goodwood, Ontario, Canada). The first represents a high-yielding cultivar commonly used for fresh-market production and known for its resilience under suboptimal growing conditions, while the second is characterized by high aromatic intensity and ornamental foliage. These commercial cultivars were included as reference genotypes to benchmark the agronomic, biochemical, and EO performance of the Iranian landraces within the predictive ANN framework.

Experimental site and design

The experiment was conducted from spring 2021 to autumn 2022 at the Medicinal Plants Research Field of the Department of Horticultural Sciences, Lorestan University, Khorramabad, Iran (33°29′ N, 48°21′ E; 1 170 m above sea level). The area is located in the western Zagros foothills and is characterized by a semi-arid Mediterranean climate, with hot, dry summers and cool, wet winters. The long-term average annual precipitation is approximately 450 mm, the mean annual temperature is 16.5 °C, and relative humidity ranges from 30% in midsummer to 75% in winter.

The experimental soil had a loam–clay texture, with pH 7.5, electrical conductivity (EC) of 0.8 dS m⁻¹, and 1.2% organic matter, representing typical conditions of irrigated fields in the region. Prior to sowing, soil samples (0–30 cm depth) were collected and analyzed to confirm uniform soil fertility and texture throughout the experimental area. The crop was grown under open-field conditions without shading structures or plastic covering.

A randomized complete block design (RCBD) with three replications was implemented to minimize environmental variability and enable robust statistical comparison among landraces. Each experimental plot measured 2 × 3 m and consisted of four rows spaced 50 cm apart. Plants within rows were spaced at 40 cm, resulting in an approximate density of 30 plants per plot. To minimize border effects, plots were separated by 1 m and blocks by 2 m. Data collection was restricted to the central rows of each plot, and border plants were excluded from all measurements. For each replication (plot), five centrally positioned plants were randomly selected. Morphological measurements were recorded on these plants and averaged to obtain one replicate value. Biochemical, antioxidant, mineral, and essential-oil analyses were subsequently performed at the biological-replicate level.

Agronomic practices

Seeds were sown directly in the field in spring 2021, when soil temperatures exceeded 10 °C and relative humidity was moderate. Germination occurred within 10 to 14 days under natural photoperiod and ambient temperature conditions. At the four-leaf stage, plants were manually thinned to achieve an intra-row spacing of 40 cm, resulting in approximately 30 plants per plot. This spacing ensured uniform light interception, aeration, and root development among plants.

Irrigation was applied through a drip system designed to maintain soil moisture close to field capacity. Watering frequency was adjusted weekly based on local evapotranspiration (ET₀) data and visual soil-moisture assessment using tensiometers. Each irrigation event provided approximately 25 to 30 mm of water, depending on weather conditions. The system ensured uniform distribution across all plots and minimized losses through evaporation and runoff.

Manual weeding was carried out at regular intervals throughout the vegetative phase to prevent competition for light, nutrients, and water. No herbicides, fungicides, or insecticides were applied during the experiment, in order to maintain natural plant growth and biochemical integrity.

Fertilization consisted of 150 kg ha⁻¹ nitrogen (urea), 100 kg ha⁻¹ phosphorus (triple superphosphate), and 100 kg ha⁻¹ potassium (potassium sulfate). Phosphorus and potassium were incorporated into the soil before sowing, while nitrogen was applied in two equal portions: one as a basal dose and the second as a topdressing at the early vegetative stage. Fertilizer rates were determined according to regional recommendations for aromatic and medicinal crops.

Irrigation frequency was gradually reduced during the final growth stage to encourage secondary metabolite accumulation and facilitate seed maturation. No growth regulators or synthetic stimulants were used at any stage of the crop cycle.

Trait measurements

At the late vegetative stage (140 days after sowing), five representative plants were randomly selected from the central rows of each plot, excluding border plants, for detailed morphological evaluation. Measurements were conducted between 09:00 and 11:00 h under uniform light conditions to minimize diurnal variation. The following vegetative traits were recorded. Plant height was measured from the junction between the root and shoot (crown base) to the apical meristem of the tallest shoot. Crown diameter was determined as the maximum horizontal spread of the plant canopy measured at soil level. The number of lateral stems was recorded as the total count of secondary shoots arising from the main stem. The number of leaves per plant was determined by counting all fully expanded leaves, while the number of leaflets per leaf was calculated as the average of three representative compound leaves. Leaf length was measured as the distance from the base of the petiole to the tip of the terminal leaflet, and leaf width as the maximum lateral distance across the broadest part of the same leaf blade [18]. Fresh weight (FW) was recorded immediately after harvest, and dry weight (DW) was obtained after oven-drying at 70 °C until constant mass was reached. Internode length was measured as the average distance between consecutive nodes along the main stem. Petiole diameter was measured at the midpoint of the petiole, and petiole length as the linear distance from the leaf base (point of attachment to the stem) to the insertion of the first leaflet [22].

During the second growing season (approximately 300 days after sowing, corresponding to full seed maturity), reproductive and yield-related traits were evaluated. The number of flowering stems per plant was recorded as the total count of reproductive shoots bearing inflorescences. Inflorescence diameter was determined as the average of two perpendicular measurements across the outer edge of the primary umbel. The number of umbels per plant represented the total number of primary inflorescences, while the number of umbellets per plant was the sum of all secondary umbels within the primary inflorescences. The number of fruits per umbellet was calculated as the mean count of fruits from three representative secondary umbels. Seed yield per plant corresponded to the total seed mass after threshing and cleaning, and the thousand-seed weight was estimated as the average weight of one thousand air-dried seeds obtained from three subsamples. Flowering stem diameter was measured at the midpoint of the main reproductive stem, and flowering stem length as the distance from the stem base at the crown to the apex of the primary inflorescence.

Morphological traits were recorded following standardized descriptors for P. crispum [23]. All data represent mean ± SD of three biological replicates per genotype (n = 3).

Photosynthetic pigments

Chlorophylls are essential pigments for photosynthesis, responsible for capturing and transferring light energy during photochemical reactions. Carotenoids act as auxiliary pigments that assist in light harvesting and also serve as major non-enzymatic antioxidants, protecting chloroplasts from oxidative stress. Assessing their concentrations provides insight into photosynthetic efficiency and the physiological condition of plants [24].

Leaf pigment contents were quantified following the method of Lichtenthaler et al. [25]. Fresh leaf samples were extracted in 100% acetone, and absorbance was measured spectrophotometrically at 470, 645, and 662 nm. Total chlorophyll and carotenoid concentrations were calculated using the corresponding equations based on specific absorption coefficients.

Total phenolic and flavonoid contents and antioxidant activity

Phenolic and flavonoid compounds are key secondary metabolites that contribute to the antioxidant capacity and overall functional quality of medicinal and aromatic plants. They play important roles in protecting plant tissues from oxidative stress, enhancing stress tolerance, and improving the nutritional and pharmacological value of the crop. Evaluating these compounds, together with antioxidant activity, provides an integrated assessment of the biochemical potential of parsley landraces [26].

Total phenolic content was determined using the Folin–Ciocalteu method [27]. Absorbance was measured at 765 nm, and results were expressed as milligrams of gallic acid equivalents per gram of dry weight (mg GAE g⁻¹ DW).

Total flavonoid content was quantified using the aluminum chloride colorimetric method described by Quettier-Deleu et al. [28]. A 50 µL aliquot of plant extract was mixed with 400 µL of 2% aluminum chloride and 1.2 mL of 5% potassium acetate. After incubation for 40 min at 37 °C, absorbance was recorded at 415 nm using a UV–Vis spectrophotometer. Flavonoid concentration was calculated from a rutin standard curve (0–500 mg L⁻¹) and expressed as milligrams of rutin equivalents per gram of dry weight (mg RE g⁻¹ DW).

Antioxidant activity was estimated using the ferric reducing antioxidant power (FRAP) assay following Benzie and Strain [29]. The method is based on the reduction of Fe³⁺–TPTZ to the blue-colored Fe²⁺–TPTZ complex under acidic conditions. A 20 µL aliquot of methanolic extract was combined with 180 µL of FRAP reagent, incubated at 37 °C for 8 min, and the absorbance was read at 593 nm. Results were expressed as micromoles of Fe²⁺ per gram of dry extract.

Ascorbic acid content

Ascorbic acid, commonly known as vitamin C, is a vital antioxidant that contributes to redox regulation, enzymatic cofactor activity, and the protection of cellular components from oxidative damage. It plays a key role in maintaining plant stress tolerance and significantly influences the nutritional quality of edible plant tissues [30].

Ascorbic acid content was determined using the 2,6-dichlorophenol-indophenol (DCPIP) titration method described by Ting and Rouseff [31]. Fresh leaf samples were homogenized in a metaphosphoric acid solution and titrated against DCPIP dye until a persistent pink endpoint was observed. The concentration was calculated from a standard ascorbic acid calibration curve, and results were expressed as milligrams per 100 g of dry weight (mg 100 g⁻¹ DW).

Essential oil content

EOs are among the most valuable secondary metabolites in aromatic plants, contributing to their characteristic aroma, flavor, and pharmacological properties [32]. In parsley, EO content is a key quality determinant with strong relevance to both industrial and medicinal applications.

Essential oil (EO) yield was quantified by hydro-distillation using a Clevenger-type apparatus for three hours, following standard pharmacopoeial procedures. The extracted oil was separated, dried over anhydrous sodium sulfate to remove residual moisture, and stored in sealed amber vials at 4 °C until analysis. Yields were expressed as grams of EO per plant (g plant⁻¹). Each biological replicate was analyzed in triplicate, and the mean value was used for statistical analysis [33].

Essential oil composition

The separation and analysis of essential oil constituents were performed using gas chromatography–mass spectrometry (GC–MS) and gas chromatography–flame ionization detection (GC–FID), both equipped with a BP-5 fused-silica capillary column. In GC–MS analysis, helium was used as the carrier gas, with injector and interface temperatures set at 280 and 260 °C, respectively. Electron impact (EI) ionization was conducted at 70 eV, with a split ratio of 1:20. Mass spectra were recorded over the m/z range of 35–450 (amu). The oven temperature was programmed from 55 to 285 °C at a rate of 4 °C min⁻¹ and held isothermally for 15 min [34].

GC–FID analysis was performed under identical oven temperature conditions. Detector and injector temperatures were maintained at 300 and 250 °C, respectively, with helium as the carrier gas at a flow rate of 1 mL min⁻¹. Retention indices (RI) were calculated using a homologous series of n-alkanes (C6–C40) injected under the same chromatographic conditions.

Identification of individual constituents was based on (1) comparison of mass spectra with the WILEY 275 and NIST 17 libraries, (2) comparison of calculated RI values with those reported in Adams [35], and (3) comparison of retention times with authentic standards when available [24]. Quantification was performed by area normalization using GC–FID data, following the method described by Jadidi et al. [36].

Leaf nutrient content

Mineral nutrients play crucial roles in plant growth, photosynthesis, and the biosynthesis of primary and secondary metabolites. Assessing their concentrations provides essential information on the nutritional status of plants and their adaptive responses to environmental conditions [37]. In parsley, macronutrients such as nitrogen, phosphorus, and potassium are closely associated with biomass production and pigment synthesis, whereas micronutrients including iron, zinc, and manganese influence enzymatic activity and antioxidant metabolism.

Fully developed young leaves were collected, oven-dried at 70 °C to constant weight, finely ground, and ashed in a muffle furnace for mineral analysis. Phosphorus was quantified by spectrophotometric determination of the molybdenum blue complex, and potassium, calcium, magnesium, iron, copper, zinc, and manganese by atomic absorption spectroscopy, following standard procedures described by Latimer [38].

Artificial neural network (ANN) modeling

The performance of an ANN depends strongly on its architecture and parameter settings. ANNs are computational models inspired by the structure and function of the human brain and are designed to process complex datasets and generate accurate predictions [39]. In this study, a multilayer perceptron (MLP), which is one of the most common ANN architectures used for predictive modeling, was employed [40].

The input dataset comprised 21 morphological variables (Supplementary Table S1), which served as predictors for the modeling process. Because of the high dimensionality and potential multicollinearity among these morphological traits, Principal Component Analysis (PCA) was applied exclusively to the predictor dataset. PCA transformed the correlated morphological variables into a smaller set of uncorrelated principal components (PCs) that retained most of the variance. All input variables were standardized to zero mean and unit variance prior to PCA and ANN training to ensure balanced feature scaling. The biochemical, pigment, mineral, and essential-oil traits served as output variables and were not subjected to PCA. During model development, PCA was fitted exclusively on the training subset within each cross-validation iteration, and the resulting transformation was subsequently applied to the corresponding validation and test subsets. This procedure prevented information leakage and ensured unbiased model evaluation. The extracted PCs were then used as inputs for ANN model development and validation.

The mean values of five plants in each biological replicate (block) were determined for every genotype. Therefore, the total number of observations in the dataset amounted to 51 (17 genotypes × 3 biological replicates). In order to enhance the use of available data and conserve biological variation, ANN models were developed on the basis of biological replicates rather than means for every genotype. Since the sample size was relatively small, repeated k-fold cross-validation was applied to ensure reliable model development and performance evaluation. Such an approach reduces the risk of overfitting, improves the stability of performance estimates, and provides more reliable information about model generalization. Only morphological descriptors were used as inputs, while biochemical traits, pigments (chlorophyll and carotenoids), antioxidants, essential oils, and minerals served as outputs. The MLP architecture, implemented in MATLAB R2018b software, followed methodological principles previously applied for non-destructive quality prediction in horticultural produce [41]. It consisted of an input layer defined by the selected PCs, one or more hidden layers, and an output layer, with the optimal number of hidden layers and neurons determined empirically using a trial-and-error approach [42]. Network training was performed using the Levenberg–Marquardt (LM) algorithm, and model performance was evaluated using Pearson correlation coefficient (R), coefficient of determination (R²), mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE). The LM algorithm was chosen because it offers faster convergence and higher accuracy for small- to medium-sized datasets compared with standard gradient-descent or quasi-Newton methods. To ensure robust generalization, the repeated k-fold cross-validation procedure was used for model selection and hyperparameter tuning, whereas a separate hold-out test set (20% of the data) was reserved for final model evaluation. The remaining 80% of the dataset was used for repeated k-fold cross-validation during model selection and hyperparameter tuning.

The overall workflow of the ANN modeling procedure, which included data normalization, PCA-based dimensionality reduction, model training, validation, and performance evaluation, is illustrated in Fig. 1. This modeling framework was used to predict EO constituents, antioxidant parameters, and mineral elements in parsley. The detailed configuration of the final ANN models, which had a major influence on predictive accuracy, is summarized in Table 2. ANNs were selected because of their ability to model nonlinear and multidimensional relationships between morphological inputs and complex biochemical outputs, making them a suitable approach for biological datasets characterized by high variability and multicollinearity.

Fig. 1.

Workflow of the artificial neural network (ANN) modeling procedure used for predicting essential-oil constituents, antioxidant traits, and mineral elements in parsley (Petroselinum crispum Mill.). The flowchart outlines the sequential steps of data processing and model development, including data acquisition, normalization, principal component analysis (PCA)–based dimensionality reduction, dataset partitioning (training and testing), repeated k-fold cross-validation, ANN structure optimization using the Levenberg–Marquardt (LM) algorithm, and final model evaluation. For simplicity, PCA is illustrated as a preprocessing step in the workflow. In the actual implementation, however, PCA was fitted exclusively within each training fold of the repeated k-fold cross-validation procedure and the resulting transformation was subsequently applied to the corresponding validation and test data, thereby preventing information leakage

Table 2.

Network architecture and training parameters of the artificial neural network (ANN) models used for predicting antioxidant traits, essential-oil constituents, and mineral elements in parsley (Petroselinum crispum Mill.)

Parameter	Value / Description
Number of input-layer units	4
Number of hidden layers	1
Number of hidden-layer units	5–12 (optimized range)
Number of output-layer units	5 / 6 / 8 (according to each prediction group)
Transfer function (hidden layer)	Sigmoid
Transfer function (output layer)	Purelin
Training function	Levenberg–Marquardt
Learning rate	0.01
Performance goal	1 × 10⁻⁶
Maximum number of epochs	1000

Statistical parameters

A suite of established statistical metrics was used to evaluate model performance and predictive accuracy. These included the Pearson correlation coefficient (R), coefficient of determination (R²), mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE). Together, these indices quantify both the strength of the relationship between predicted and observed values and the magnitude of prediction errors. RMSE, MSE, and MAE values were calculated using the original measurement units of each response variable. Consequently, these error metrics are scale-dependent and should be interpreted within each trait rather than compared directly across variables expressed in different units. All metrics were calculated according to Eqs. (1, 2, 3, 4 and 5), which provide a reliable basis for comparing the accuracy and robustness of the developed models.

1

2

3

4

5

In these equations, and denote the predicted and experimental values for observation (i), respectively. Z_{p, avg} and Z_{e, avg} represent the mean predicted and mean experimental values, respectively. The total number of observations is denoted by (n). Higher values of R and especially R² indicate stronger agreement between predicted and observed data, whereas lower MSE, RMSE, and MAE values correspond to smaller prediction errors. These parameters collectively provide a quantitative assessment of the performance, accuracy, and reliability of each ANN model.

Results

Phenotypic, biochemical, and compositional variability among parsley landraces

Marked variability was observed among the Iranian parsley landraces in morphological, biochemical, EO, and mineral characteristics (Supplementary Tables S1–S4). This diversity likely reflects genetic differences and ecotypic variation associated with the agro-ecological origin of the landraces.

In terms of morphological and yield-related traits (Supplementary Table S1), high-altitude landraces such as Shahrekord–Saman and Kurdistan exhibited greater plant height, larger inflorescences, and higher dry biomass, indicating adaptation to cooler and more humid conditions. Conversely, lowland accessions such as Khuzestan and Bojnurd produced compact morphotypes with thicker stems and heavier seeds, traits often associated with tolerance to warmer and drier environments. Among the reference cultivars, P. crispum var. crispum showed vigorous vegetative growth and high biomass accumulation, while P. crispum var. neapolitanum displayed a shorter stature and finely divided, highly aromatic foliage.

Biochemical analyses (Supplementary Table S2) revealed wide variation in the accumulation of pigments, phenolic and flavonoid compounds, and antioxidant activity. Landraces from cooler regions, including Dorud and Tabriz, recorded the highest total phenolic contents, while those from warmer regions such as Arak and Fasa showed moderate levels. The Kurdistan and Hamedan landraces exhibited superior antioxidant capacity, as indicated by high FRAP values. Overall, most Iranian genotypes contained higher phenolic, flavonoid, and vitamin C levels than the commercial cultivars, suggesting greater metabolic richness and potential for functional food applications.

Significant diversity was also detected in EO yield and composition (Supplementary Table S3). The Crispum cultivar exhibited the highest EO yield, while Shahrekord–Saman and Kurdistan landraces were characterized by high proportions of linalool and apiol. The Arak ecotype showed an exceptionally high myristicin concentration, contributing to a distinct aromatic profile. Although Neapolitanum displayed strong aroma intensity, its oil yield was lower, consistent with its selection for culinary and ornamental uses rather than for industrial extraction. These differences emphasize the influence of genetic background and ecotypic variation associated with agro-ecological origin.

Analysis of leaf mineral content (Supplementary Table S4) further confirmed considerable variability among genotypes. The Crispum cultivar had the highest concentrations of calcium, magnesium, and potassium, whereas Neapolitanum was distinguished by elevated iron and zinc levels. Landraces from Hamedan, Arak, and Kurdistan exhibited balanced macro- and micronutrient profiles, while lowland populations such as Khuzestan and Bojnurd contained lower levels of several micronutrients, possibly due to soil or climatic constraints. The observed mineral variability provides a useful foundation for selecting nutritionally superior genotypes and optimizing fertilizer management in parsley cultivation.

Collectively, these findings highlight the extensive phenotypic and biochemical diversity among Iranian parsley landraces and the distinct performance of the two reference cultivars. This dataset forms the basis for subsequent multivariate analysis and predictive modeling aimed at identifying key traits driving variation in antioxidant capacity, EO composition, and mineral content.

Principal component analysis (PCA)

Principal Component Analysis (PCA) was applied to the 21 morphological variables listed in Supplementary Table S1 to reduce data dimensionality and identify the main sources of morphological variation among parsley landraces and cultivars. The first four principal components (PC₁–PC₄) together explained more than 98% of the total variance (Fig. 2). PCA was applied exclusively to the morphological dataset, whereas biochemical, mineral, and EO traits were retained in their original form and used directly as model outputs.

Fig. 2.

Variance contribution and corresponding eigenvalues of the first four principal components (PCs) derived from principal component analysis (PCA) of the 21 morphological input variables used to train the artificial neural network (ANN) models. PC₁–PC₄ collectively explained more than 98% of the total variance, capturing the major patterns of morphological variability among parsley (Petroselinum crispum Mill.) landraces. These components were subsequently used as ANN inputs to minimize multicollinearity and improve model training efficiency

The variance contributions and corresponding eigenvalues of the principal components are presented in Fig. 2. These components effectively summarized the major sources of morphological variation and provided a reduced set of informative and non-collinear inputs for ANN modeling. The high proportion of explained variance indicates that these four components retained most of the information contained in the original morphological dataset while substantially reducing dimensionality and multicollinearity, thereby providing suitable inputs for subsequent ANN analysis.

ANN model architecture and training

The PCA-transformed morphological dataset was used as input for the ANN models, which were trained using the LM algorithm. The ANN models were trained and validated using morphological variables as input features and biochemical, EO, and mineral traits as output variables. The corresponding datasets are summarized in Supplementary Tables S1–S4. The main network parameters are summarized in Table 2, and the detailed training statistics for each model are provided in Supplementary Table S5.

Several MLP architectures were evaluated by varying the number of hidden-layer neurons between five and twelve. The best-performing topologies were identified as 4–9–6 for antioxidant traits, 4–8–5 for EO constituents, and 4–10–8 for mineral elements. Each model achieved convergence before the maximum epoch limit (1000) with a learning rate of 0.01 and a performance goal of 1 × 10⁻⁶.

The complete ANN workflow, including preprocessing, training, validation, and testing, is illustrated in Fig. 1. Model training converged successfully for all optimized network architectures.

Prediction of antioxidant traits

The ANN model trained to estimate antioxidant-related parameters achieved high predictive accuracy across the examined variables. The best-fit network topology for this trait group was 4–9–6, consisting of four PCs as inputs, nine neurons in the hidden layer, and six outputs representing the antioxidant attributes (Fig. 3). The model performance for total chlorophyll, carotenoid, and total phenolic contents is illustrated in Fig. 4, while additional predictions for vitamin C, total flavonoid content, and FRAP are shown in Fig. 5.

Fig. 3.

Optimal artificial neural network (ANN) topology (4–9–6) used for predicting antioxidant traits in parsley (Petroselinum crispum Mill.). The network consists of an input layer with four principal components (PC₁–PC₄), one hidden layer with nine neurons, and an output layer with six antioxidant-related variables (total chlorophyll, carotenoids, total phenolics, flavonoids, vitamin C, and FRAP). Connection lines indicate weighted links between neurons, and weight adjustments during training were optimized using the Levenberg–Marquardt (LM) algorithm to minimize prediction error

Fig. 4.

Performance of the artificial neural network (ANN) model for predicting total chlorophyll (A–B), carotenoid content (C–D), and total phenolic content (E–F) in parsley (Petroselinum crispum Mill.). Panels (A), (C), and (E) show concordance plots comparing predicted ANN outputs with experimentally measured values, while panels (B), (D), and (F) depict the corresponding correlations between observed and predicted data

Fig. 5.

Performance of the artificial neural network (ANN) model for predicting vitamin C (A–B), total flavonoid content (C–D), and ferric-reducing antioxidant power (FRAP) (E–F) in parsley (Petroselinum crispum Mill.). Panels (A), (C), and (E) display concordance plots comparing predicted ANN outputs with experimentally measured values, while panels (B), (D), and (F) depict the corresponding correlations between observed and predicted data

The strong correlation between experimental and predicted data (R = 0.81–0.93; R² = 0.66–0.86) indicates that the PCA-ANN framework effectively captured nonlinear relationships between morphological descriptors and antioxidant-related traits (Table 3). Among the evaluated traits, flavonoid content exhibited the highest predictive accuracy (R² = 0.86; RMSE = 1.50), followed by total phenolic content and total chlorophyll. The high concordance between observed and predicted values in Figs. 4 and 5 confirms that morphological descriptors can be effectively used to predict antioxidant-related traits in parsley.

Table 3.

Statistical performance of the artificial neural network (ANN) model for predicting antioxidant traits in parsley (Petroselinum crispum Mill.). R, Pearson’s correlation coefficient; R², coefficient of determination; MSE, mean squared error; RMSE, root mean squared error; MAE, mean absolute error. Higher R and especially R² values indicate stronger agreement between observed and predicted data, whereas lower MSE, RMSE, and MAE values reflect smaller prediction errors and greater model accuracy

Response	R	R²	MSE	RMSE	MAE
Total chlorophyll	0.8188	0.6704	3.0193	1.7376	1.3696
Carotenoids	0.8541	0.7295	0.2532	0.5032	0.3841
Phenolics	0.8131	0.6611	245.026	15.6533	13.7569
Flavonoids	0.9297	0.8643	2.2349	1.4950	1.1690
Vitamin C	0.8331	0.6940	17.3102	4.1605	3.4345
Antioxidant activity (FRAP)	0.8881	0.7888	5741.67	75.773	69.965

These results demonstrate that the ANN model can serve as a reliable computational tool for rapid prediction of antioxidant traits in parsley germplasm, offering a non-destructive and cost-effective alternative to traditional laboratory assays.

Prediction of essential-oil yield and composition

The ANN model developed to predict EO yield and composition demonstrated strong predictive capability across the five major EO constituents analyzed: β-phellandrene, linalool, myristicin, apiol, and total oil yield (Table 4). The optimal network topology for EO traits was 4-8-5, consisting of four PCs as inputs, eight neurons in the hidden layer, and five output neurons corresponding to the key EO parameters (Fig. 6).

Table 4.

Statistical performance of the artificial neural network (ANN) model for predicting essential-oil (EO) constituents in parsley (Petroselinum crispum Mill.). R, Pearson’s correlation coefficient; R², coefficient of determination; MSE, mean squared error; RMSE, root mean squared error; MAE, mean absolute error. Higher R and R² values indicate stronger agreement between observed and predicted data, whereas lower MSE, RMSE, and MAE values reflect smaller prediction errors and higher model accuracy

Response	R	R²	MSE	RMSE	MAE
Essential-oil yield	0.8417	0.7085	0.0051	0.0714	0.0585
β-phellandrene	0.8776	0.7702	11.5886	3.4042	2.9862
Linalool	0.9708	0.9424	8.1683	2.8580	2.5437
Myristicin	0.9335	0.8714	31.0348	5.5709	4.7377
Apiol	0.9124	0.8325	5.5673	2.3595	2.0417

Fig. 6.

Optimal artificial neural network (ANN) topology (4–8–5) constructed for predicting essential-oil (EO) constituents in parsley (Petroselinum crispum Mill.). The diagram illustrates the multilayer perceptron (MLP) architecture comprising four input neurons (principal components obtained from PCA), one hidden layer with eight neurons, and five output neurons corresponding to the major EO parameters (β-phellandrene, linalool, myristicin, apiol, and total EO yield). Weighted interconnections between neurons represent synaptic links optimized through the Levenberg–Marquardt (LM) algorithm to minimize mean squared error and enhance predictive accuracy

The model exhibited high accuracy, with R ranging from 0.84 to 0.97 and R² up to 0.94 (Table 4). Similar to the antioxidant model, the EO network exhibited consistently high correlation coefficients (R = 0.84–0.97), confirming strong predictive performance across EO traits. Among the EO constituents, linalool and myristicin achieved the strongest predictive performance (R² = 0.94 and 0.87, respectively), followed by apiol and β-phellandrene. The performance plots presented in Fig. 7 confirm the high degree of concordance between observed and predicted values, with most data points closely aligned along the 1:1 line.

Fig. 7.

Performance of the artificial neural network (ANN) model for predicting essential-oil (EO) yield (A–B), β-phellandrene (C–D), linalool (E–F), myristicin (G–H), and apiol I–J in parsley (Petroselinum crispum Mill.). Panels (A), (C), (E), (G), and I display concordance plots comparing predicted ANN outputs with experimentally measured values, while panels (B), (D), (F), (H), and J depict the corresponding correlations between observed and predicted data

These findings indicate that the PCA–ANN framework effectively modeled complex nonlinear relationships between morphological descriptors and EO yield and composition. The ability to estimate EO yield and composition using easily measurable morphological descriptors offers a practical alternative to conventional GC–MS analysis, significantly reducing analytical time and cost. Moreover, the accurate prediction of aroma-defining compounds such as linalool and apiol underscores the model’s potential utility in quality control and cultivar selection for pharmaceutical, cosmetic, and culinary markets.

The high agreement between observed and predicted concentrations across EO constituents (Fig. 7) is consistent with the strong predictive performance of the EO model summarized in Table 4.

Prediction of mineral elements

The ANN model trained to estimate the mineral composition of parsley leaves achieved excellent predictive performance across the eight analyzed elements: calcium (Ca), magnesium (Mg), potassium (K), phosphorus (P), iron (Fe), manganese (Mn), zinc (Zn), and copper (Cu) (Table 5). The optimal network topology for this dataset was 4–10–8, consisting of four PCs as inputs, ten neurons in the hidden layer, and eight output neurons corresponding to the mineral elements (Fig. 8).

Table 5.

Statistical performance of the artificial neural network (ANN) model for predicting mineral elements in parsley (Petroselinum crispum Mill.). R, Pearson’s correlation coefficient; R², coefficient of determination; MSE, mean squared error; RMSE, root mean squared error; MAE, mean absolute error. Higher R and R² values indicate stronger agreement between observed and predicted data, whereas lower MSE, RMSE, and MAE values reflect smaller prediction errors and greater model accuracy

Response	R	R²	MSE	RMSE	MAE
Ca	0.9717	0.9441	197 518.7	444.43	345.96
Cu	0.8728	0.7617	0.7838	0.8853	0.7757
K	0.9063	0.8213	2 459 517.7	1 568.28	1 304.99
P	0.9287	0.8625	979 968.9	989.93	912.67
Mg	0.8980	0.8064	314 149.4	560.49	498.58
Mn	0.9222	0.8504	5.9371	2.4366	2.1571
Fe	0.9150	0.8372	319.9730	17.8878	13.9299
Zn	0.8890	0.7904	1 122.79	33.51	28.36

Fig. 8.

Optimal artificial neural network (ANN) topology (4–10–8) designed for predicting mineral elements in parsley (Petroselinum crispum Mill.). The schematic illustrates the multilayer perceptron (MLP) architecture comprising four input neurons (principal components derived from PCA), one hidden layer with ten neurons, and eight output neurons corresponding to the analyzed mineral nutrients (Ca, K, Mg, P, Fe, Zn, Cu, and Mn). Interconnecting lines represent synaptic weights optimized through the Levenberg–Marquardt (LM) algorithm to minimize mean squared error and improve model generalization across parsley landraces

The model exhibited strong R (0.87–0.97) and R² (0.76–0.94), confirming its high predictive reliability (Table 5). Consistent with the antioxidant and EO networks, the mineral model exhibited high correlation coefficients (R = 0.87–0.97), confirming the robustness of the PCA–ANN framework across different trait categories. Among the minerals, calcium displayed the highest prediction accuracy (R² = 0.94), followed by potassium and phosphorus. The performance plots presented in Figs. 9 and 10 show close alignment between observed and predicted values, with minimal deviation from the 1:1 reference line across all elements. These graphical diagnostics confirm the robustness and reliability of the predictions, indicating that the PCA–ANN framework effectively captured complex nonlinear relationships between morphological descriptors and mineral nutrient concentrations.

Fig. 9.

Performance of the artificial neural network (ANN) model for predicting manganese (Mn) (A–B), iron (Fe) (C–D), copper (Cu) (E–F), and magnesium (Mg) G–H in parsley (Petroselinum crispum Mill.). Panels (A), (C), (E), and G show concordance plots comparing predicted ANN outputs with experimentally measured values, while panels (B), (D), (F), and H depict the corresponding correlations between observed and predicted data

Fig. 10.

Performance of the artificial neural network (ANN) model for predicting calcium (Ca) (A–B), phosphorus (P) (C–D), potassium (K) (E–F), and zinc (Zn) G–H in parsley (Petroselinum crispum Mill.). Panels (A), (C), (E), and G display concordance plots comparing predicted ANN outputs with experimentally measured values, while panels (B), (D), (F), and H show the corresponding correlations between observed and predicted data

These findings demonstrate that morphological descriptors contain substantial predictive information for the estimation of mineral nutrient concentrations. The strong predictive capacity of the ANN framework highlights its suitability for rapid nutrient diagnostics and precision-fertilization management in parsley cultivation. Furthermore, the ability to estimate micronutrient status (e.g., Fe, Zn, and Mn) non-destructively offers practical advantages for evaluating nutritional quality in breeding and commercial production programs.

The close clustering of observed and predicted mineral concentrations along the 1:1 line (Fig. 9 and 10) further corroborates the comparative statistics presented in Table 5, confirming the strong overall accuracy and generalization ability of the mineral prediction model. Representative comparisons of observed and predicted mean values for top-performing traits, including linalool, calcium, and flavonoids, are provided in Supplementary Table S6. These results illustrate the close agreement and low prediction errors achieved by the optimized ANN models. Together, they validate the overall predictive reliability of the PCA–ANN framework and confirm its capacity to generate accurate and consistent estimations across independent biochemical, EO, and mineral datasets.

Model interpretation

The successful prediction of antioxidant, essential-oil (EO), and mineral traits using morphological descriptors indicates that measurable differences in plant structure and growth were associated with variation in biochemical and compositional characteristics. These findings suggest that morphological vigor and leaf architecture are closely associated with variation in antioxidant capacity and EO-related traits. The successful prediction of linalool, total phenolics, and other quality attributes further indicates that morphological descriptors contain useful information related to variation in these traits, highlighting an association between canopy morphology and secondary-metabolite accumulation.

This relationship supports the biological plausibility of the PCA–ANN model and suggests that the predictive performance is linked to meaningful morphological variation. It also underscores the potential of non-destructive morphological assessment as a practical surrogate for complex biochemical analyses.

Model performance comparison

The overall predictive accuracy of all ANN models is summarized in Table 6. Across antioxidant, EO, and mineral datasets, correlation coefficients (R) ranged from 0.81 to 0.97, with corresponding R² values between 0.66 and 0.94. These results demonstrate consistently high predictive performance across the different trait groups. Detailed training and validation statistics for each model are presented in Supplementary Table S5.

Table 6.

Overall comparison of artificial neural network (ANN) model performance for predicting antioxidant, essential-oil (EO), and mineral traits in parsley (Petroselinum crispum Mill.). R, Pearson correlation coefficient; R², coefficient of determination; RMSE, root mean squared error; MAE, mean absolute error. Higher R and R² values indicate stronger agreement between predicted and observed data, whereas lower RMSE and MAE values denote smaller prediction errors within each trait

Model type	Output traits (no.)	Optimal topology	R (range)	R² (range)	RMSE (range)	MAE (range)	Best-predicted variable(s)	Reference figures / tables
Antioxidant traits	6	4–9–6	0.81–0.93	0.66–0.86	1.50–75.8	1.17–70.0	Flavonoids (R = 0.93, R² = 0.86)	Figs.  3, 4 and 5
EO composition	5	4–8–5	0.84–0.97	0.71–0.94	0.07–5.6	0.06–4.7	Linalool (R = 0.97, R² = 0.94)	Figs. 6 and 7
Mineral elements	8	4–10–8	0.87–0.97	0.76–0.94	0.89–1568	0.78–1305	Calcium (R = 0.97, R² = 0.94)	Figs. 8, 9 and 10

RMSE and MAE values are reported in the original units of each trait. Because the evaluated variables differ in scale and measurement units, RMSE and MAE values should be interpreted within individual traits and should not be directly compared across variables or trait groups. Consequently, comparisons among model categories should be based primarily on R and R² rather than on absolute error metrics

Together, these statistical assessments indicate that the PCA–ANN framework efficiently captured complex nonlinear relationships between morphological descriptors and antioxidant, EO, and mineral traits, providing a robust and non-destructive method for predicting key physiological and compositional characteristics in parsley.

Discussion

Importance of landrace diversity and biological variability

The considerable diversity observed among the evaluated Iranian parsley landraces (Table 1) formed the biological basis for model development and validation. As shown in Supplementary Tables S1–S4, wide variation was recorded across morphological, biochemical, EO, and mineral traits. This diversity exceeded that of the two commercial cultivars and reflects long-term adaptation to heterogeneous environments, including lowland semi-arid regions and high-altitude temperate areas. Such ecotypic differentiation has resulted in unique combinations of morphological vigor, pigment content, antioxidant capacity, and EO composition.

The breadth of phenotypic and biochemical variation was also evident in the PCA results (Fig. 2), where landraces were clearly separated from the commercial references, confirming high intra- and inter-population heterogeneity. This variability not only provides a rich resource for identifying superior genotypes with desirable agronomic and biochemical profiles but also enhances the representativeness of data-driven models. Incorporating genetically diverse landraces provides a broader biological basis for model development, allowing ANNs to capture relationships between morphological variation and underlying biochemical and compositional traits.

From a breeding perspective, the inclusion of landrace-based datasets in intelligent modeling frameworks supports the dual goals of genetic resource conservation and targeted selection for quality and resilience. Such integration bridges traditional agro-biodiversity with modern computational tools, promoting sustainable crop improvement and precision cultivation.

Predictive performance of the ANN models

The present study demonstrated the high predictive capability of ANNs for estimating antioxidant traits, EO constituents, and mineral elements in parsley based on morphological inputs. The strong correlation coefficients (R = 0.81–0.97) and coefficients of determination (R² up to 0.94) confirm that the developed ANN models effectively captured complex nonlinear relationships between the morphological predictors and the biochemical, EO, and mineral response variables (Tables 2, 3, 4, 5 and 6; Figs. 3, 4, 5, 6, 7, 8, 9 and 10; Supplementary Tables S1–S4). Such R² values are considered high for biological datasets, where inherent variability and noise often limit predictive accuracy to below 0.85 in comparable plant-trait modeling studies [e.g., 42, 43]. These findings are consistent with previous reports showing that ANNs can effectively model complex nonlinear relationships in plant science applications [40, 42].

The inclusion of PCA-derived morphological components (PC₁–PC₄) (Fig. 2) substantially improved network efficiency by minimizing redundancy among correlated morphological traits and leveraging the major sources of phenotypic variability among parsley landraces (Supplementary Table S1). This hybrid PCA–ANN approach reduced model complexity, accelerated model training, and mitigated the effects of multicollinearity, thereby facilitating efficient model development. Comparable performance improvements have been reported for PCA–ANN models developed for other aromatic and medicinal plants, including Satureja rechingeri and Thymus vulgaris [20, 43]. Similar predictive accuracies (R² ≈ 0.90) have also been reported for Random Forest models in basil and mint datasets [e.g., 21], highlighting the potential of non-linear machine-learning approaches for phytochemical prediction. These relationships may help explain the observed predictive associations between structural vigor, leaf morphology, and antioxidant and essential-oil traits. The coherence between morphological descriptors and biochemical outputs supports the biological plausibility of the PCA–ANN framework and suggests that the observed relationships may reflect underlying biological processes.

Antioxidant trait prediction and biochemical interpretation

The ANN model with a 4–9–6 topology (Fig. 3) achieved high predictive accuracy for antioxidant parameters, with the strongest performance observed for flavonoids (R² = 0.86) and FRAP (R² = 0.79) (Table 3; Figs. 4 and 5). These metabolites are central to the antioxidant defense system of parsley and exhibit significant genotypic variability (Supplementary Table S2) [4, 8]. The strong association between predicted and observed values indicates that morphological descriptors can provide useful information for estimating secondary-metabolite accumulation when modeled through an ANN framework. The variability in antioxidant capacity among landraces, particularly in phenolic and flavonoid concentrations (Supplementary Table S2), supports the differentiation patterns revealed by the PCA biplot (Fig. 2) and the strong correlations visualized in the performance plots (Fig. 4). Similar predictive relationships between leaf spectral features and phenolic accumulation have been described in mint and lemon balm using chemometric and machine-learning approaches [44, 45]. The robust prediction of FRAP (Fig. 5) further demonstrates the network’s capacity to integrate multiple biochemical indicators governing total antioxidant activity. Because conventional chemical analyses for antioxidant traits are destructive and time-consuming, the presented PCA–ANN-based method (Tables 2 and 3) provides a non-destructive, rapid, and cost-effective alternative for large-scale phenotyping and screening of parsley germplasm.

Prediction of essential-oil composition

The ANN model with a 4–8–5 configuration (Fig. 6) accurately predicted the five principal EO constituents, achieving R² values up to 0.94 (Table 4; Fig. 7). Among the individual compounds, linalool, myristicin, and apiol showed the highest agreement between predicted and measured concentrations, indicating that morphological descriptors summarized in Supplementary Table S1 contained useful information for predicting these constituents. These metabolites are key aroma and pharmacologically active components that determine both the flavor and medicinal quality of parsley oil [2, 36].

Marked differences in EO yield and constituent ratios among parsley landraces (Supplementary Table S3) were effectively captured by the ANN model (Fig. 7), with detailed performance metrics provided in Table 4. The separation among genotypes visualized in the PCA score plot (Fig. 2) further supports the observed biochemical diversity underlying EO composition.

The strong correlations between observed and predicted data (Fig. 7) indicate that morphological descriptors can be effectively used to estimate EO yield and composition. Comparable applications of ANN and support vector machine (SVM) models have been demonstrated for Ocimum basilicum, including the optimization of essential-oil content and the prediction of rosmarinic-acid concentration, as well as for Coriandrum sativum in spectral quality assessment [46–48].

Overall, the PCA–ANN framework (Tables 2 and 4) provides a rapid, non-destructive, and cost-efficient alternative to conventional GC–MS or GC–FID analysis, enabling accurate biochemical assessment for high-throughput screening and breeding of aromatic parsley landraces.

Prediction of mineral composition and nutritional quality

The ANN topology optimized for mineral prediction (4–10–8; Fig. 8) exhibited strong performance across all eight evaluated elements [i.e., calcium (Ca), magnesium (Mg), potassium (K), phosphorus (P), iron (Fe), manganese (Mn), zinc (Zn), and copper (Cu)], as detailed in Table 5. Calcium and potassium achieved the highest predictive accuracy (R² = 0.94 and 0.86, respectively), reflecting their dominant physiological roles in cell-wall stabilization, osmotic regulation, and enzyme activation [38]. Magnesium and iron were also predicted reliably, indicating that morphological descriptors contain useful information related to variation in mineral nutrient concentrations.

Marked variability in mineral concentrations among parsley landraces (Supplementary Table S4) was efficiently captured by the ANN framework, supported by the PCA-derived differentiation observed in Fig. 2. The results shown in Figs. 9 and 10 further illustrate the close clustering of observed and predicted values along the 1:1 reference line, supporting the reliability of model predictions across the evaluated genotypes.

Comparable ANN-based models have been successfully applied in Lactuca sativa [49] and Spinacia oleracea [50] for nutrient-content prediction from leaf optical data, and a related hyperspectral-based model has been described for Solanum lycopersicum [51]. Such predictive capability has significant implications for precision nutrient management, rapid nutritional diagnosis, and sustainable fertilization practices in herbaceous crops.

Comparative model behavior and methodological implications

The consistent improvement in predictive accuracy achieved through PCA preprocessing underscores the critical role of feature selection in ANN modeling. By transforming correlated predictors into uncorrelated PCs, PCA minimized redundancy and provided a reduced set of non-collinear inputs for ANN modeling. The LM training algorithm provided stable model training and efficient optimization for the relatively small dataset used in this study [40].

Overall, the PCA–ANN framework applied in this study proved effective for modeling nonlinear relationships between morphological descriptors and biochemical, essential-oil, and mineral traits in aromatic and medicinal plants. Its ability to link easily measurable morphological characteristics with complex quality-related attributes offers a scalable foundation for intelligent prediction systems in precision horticulture.

Limitations and future improvements

While the developed PCA–ANN framework demonstrated strong predictive accuracy, several limitations should be acknowledged. Because all landraces were evaluated at a single experimental site under a common set of environmental conditions, the present study does not quantify genotype × environment interactions. Instead, the observed differences should be interpreted as ecotypic variation associated with the agro-ecological origin of the landraces under common field conditions. Although measurements were collected at different developmental stages throughout the crop cycle, all observations originated from a single experimental location. Because the ANN models were developed under a single set of environmental conditions, part of the observed predictive accuracy may reflect environment-specific relationships that are not necessarily stable across years or locations. Consequently, model performance may be overestimated when extrapolated beyond the environmental conditions represented in the present dataset. Future research should therefore incorporate coordinated multi-location and multi-year evaluations encompassing contrasting climatic regions, soil types, and production systems. Such datasets would enable independent external validation and provide a more rigorous assessment of model robustness, transferability, and predictive stability across environments.

The dataset, while comprehensive in trait diversity, remains relatively limited in sample size. Expanding the dataset through larger-scale, multi-location and multi-season sampling would enhance statistical power and improve model generalization across climatic conditions. Moreover, integrating spectral and image-based phenotyping data could provide additional explanatory variables that capture canopy-level variation not directly represented by the morphological descriptors used in the present study.

Another limitation concerns model architecture. The current ANN models were optimized for static datasets and do not yet incorporate dynamic environmental inputs such as temperature, humidity, or soil moisture. Coupling the existing framework with hybrid or ensemble approaches such as CNN–ANN, RF–ANN, or gradient boosting models could improve predictive robustness and adaptability to real-time agricultural decision-making. Benchmarking PCA–ANN performance against alternative modeling approaches such as multiple linear regression, partial least squares regression, random forest, support vector regression, and XGBoost would further clarify performance trade-offs, predictive robustness, model interpretability, and computational efficiency.

An additional limitation is that the PCA–ANN framework was optimized primarily for predictive accuracy and dimensionality reduction rather than for quantifying the contribution of individual morphological variables. Because ANN inputs consisted of principal components derived from multiple correlated traits, direct interpretation of variable importance was beyond the scope of the present study. Future research should incorporate sensitivity analysis, feature-importance metrics, and explainable-AI approaches to quantify the contribution of individual morphological traits and improve the biological interpretability of model predictions.

Finally, external validation using independent datasets is essential to confirm the transferability of the developed models across different parsley genotypes and cultivation systems. Furthermore, genotype-wise validation strategies, in which entire genotypes are excluded during model training and subsequently used for independent testing, would provide a more stringent assessment of model transferability to previously unseen genetic material. Recognizing these limitations provides a roadmap for extending the current modeling framework into a more scalable, robust, and integrative predictive system suitable for precision horticulture, product certification, and agro-industrial standardization.

Practical applications and future perspectives

From an applied perspective, the developed ANN models can serve as decision-support tools for breeders and growers seeking rapid and non-destructive estimation of quality traits. By using easily measurable morphological inputs, the framework enables rapid prediction of antioxidant potential, EO constituents, and mineral content, facilitating genotype selection and precision agronomic management. Such an approach aligns with the principles of data-driven and sustainable agriculture, supporting efficient resource use and improved product quality, in line with the summarized performance trends (Table 6).

Because the framework was developed using datasets encompassing substantial morphological, biochemical, and compositional variability (Supplementary Tables S1–S4), it may provide a useful methodological foundation for adaptation to other aromatic and medicinal crops. This adaptability makes the proposed PCA–ANN framework a valuable reference model for broader applications in plant phenotyping, functional breeding programs, and digital quality monitoring.

Future research should incorporate multi-environment, multi-year, and spectral or image-based datasets into the ANN framework to further enhance robustness and transferability under diverse field conditions. Extending this approach to hybrid architectures that combine PCA–ANN with convolutional neural networks (CNNs), random forest (RF), or gradient boosting models could further improve predictive performance and interpretability across complex agricultural systems. Ultimately, the present study demonstrates the potential of intelligent ANN-based modeling as a rapid, scalable, and environmentally friendly complementary approach to conventional analytical methods for evaluating the biochemical and nutritional profiles of parsley landraces, thereby supporting precision horticulture, certification, and product standardization in medicinal and aromatic crops. Beyond parsley, the proposed PCA–ANN framework could be integrated into AI-assisted field phenotyping platforms and other digital agriculture systems, enabling rapid, non-destructive monitoring of plant quality and physiological status.

Conclusions

The present study successfully developed and validated ANN models for the intelligent prediction of antioxidant traits, EO constituents, and mineral elements in parsley. By integrating PCA with optimized ANN architectures (4–9–6, 4–8–5, and 4–10–8), the models achieved high predictive accuracy, demonstrating that morphological descriptors contain useful information for predicting complex phytochemical and mineral traits and may reduce reliance on destructive, labor-intensive analyses such as GC–MS. The proposed modeling framework provides a rapid, cost-efficient, and non-destructive approach for screening and selecting superior parsley genotypes. The developed PCA–ANN framework has the potential to be adapted to other medicinal and aromatic crops following appropriate validation, thereby supporting precision agriculture and data-driven breeding strategies. To our knowledge, this represents one of the first integrative applications of PCA–ANN modeling for the simultaneous prediction of EO composition, antioxidant capacity, and mineral nutrition in parsley, offering a scalable and intelligent tool for quality assessment and certification in medicinal plant value chains. Future studies should explore hyperspectral and morphometric imaging as additional non-destructive data sources that could be integrated into the PCA–ANN framework, extending its applicability toward rapid, non-destructive phenotyping and advanced decision-support systems in precision horticulture.

Abbreviations

ANN

artificial neural network

DW

dry weight

EO

essential oil

FRAP

ferric reducing antioxidant power

FW

fresh weight

GAE

gallic acid equivalents

GC–FID

gas chromatography–flame ionization detector

GC–MS

gas chromatography–mass spectrometry

LM

Levenberg–Marquardt

MLP

multilayer perceptron

PC

principal component

PCA

principal component analysis

RE

rutin equivalents

RMSE

root mean square error

R²

coefficient of determination

Authors’ contributions

Amin Taheri-Garavand; conceived the study, performed ANN modeling, writing–review–editing, Hasan Mumivand; Investigation, Methodology, Data curation, Formal analysis, Visualization, writing–review–editing. Parisa Khanizadeh; Data curation, performed essential oil extraction and GC-MS analysis. Maryam Fizimanesh conducted antioxidant assays and statistical analysis. Dimitrios Fanourakis; Validation, Visualization, writing original draft–review–editing.

Funding

No funding was received for this research.

Data availability

The data that support the findings of this study are available from the corresponding author upon request.

Declarations

Ethics approval and consent to participate

This article does not involve research on humans, animals, or endangered species. All experimental studies and materials comply with relevant institutional, national, and international guidelines and laws. The plant (herbarium code LU1510) was identified by Dr. Faraj Allah Tarnian at the Dept. of Pasture and Water shed Management, Faculty of Science, Lorestan University, Khorramabad, I.R.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.