Comprehensive Quality Analysis of Common Vetch (Vicia sativa L.) Varieties Using Image Processing Techniques and Artificial Intelligence

Abstract

Common vetch (Vicia sativa L.) is a cool-season annual legume cultivated for grain and forage, valued for its high nutrient content, broad edaphoclimatic adaptability, and suitability for crop rotations. Physical seed attributes are critical for variety classification, quality evaluation, and breeding selection. This study aimed to characterize the nutritional composition, mineral contents, and physical attributes of nine common vetch varieties and to assess the feasibility of binary variety classification using supervised machine learning (ML). Proximate analyses (e.g., crude protein, tannin), macro/micro minerals, and morpho-physical seed descriptors were determined. Multivariate and discriminant analyses were conducted. Binary classifiers were developed with a multilayer perceptron (MLP) and random forest (RF) under stratified 10-fold cross-validation. The highest values were observed for crude protein (22.66%, Alper), ADF (11.36%, Alınoğlu), NDF (16.47%, Alperen), and tannin (3.12%, Alınoğlu). For mineral profiles, Alper, Ankara Moru, and Doruk emerged as prominent varieties. In pairwise discrimination, Ankara Moru vs. Ayaz achieved 89% (MLP) and 90% (RF) accuracy, followed by Ankara Moru vs. Özveren with 88% (MLP) and 90.50% (RF). These results demonstrate that MLP and RF can classify common vetch varieties from physical attributes with high reliability. Integrating compositional, mineral, and morpho-physical data with supervised learning provides an objective, low-cost pathway for variety identification. The approach has direct implications for quality assessment, planting system design, and breeding. Future work should expand datasets, incorporate color-rich/hyperspectral cues, and compare feature-based models with domain-adapted deep learning on larger, multi-site collections.

1. Introduction

The genus Vicia L. is a member of the tribe Fabeae of the subfamily Papilionoideae, comprising approximately 150–210 species distributed primarily across North America, Asia, Europe, tropical Africa, and the temperate regions of South America [1,2]. The greatest specific diversity occurs in Northwest Asia and Türkiye, where fifty-nine vetch species have been identified [3]. Among these, common vetch (Vicia sativa L.) is a self-pollinated annual herbaceous plant cultivated for hay, silage, green manure, cover crops, and pasture [4]. It also serves as a valuable feed source in animal production due to its high protein and mineral content [5], with seeds containing 15–33.7% crude protein [6]. Numerous studies have investigated the mineral and protein composition of common vetch [6,7,8], though these traits are influenced by genetic factors and environmental conditions [9,10].

Morphological and dimensional characteristics are significant quality indicators of seeds, essential for engineering equipment design and operations such as seed categorization and quality evaluation. Maintaining seed quality, detecting impurities, and optimizing production are vital for agricultural research and quality control [11,12]. Computer vision (CV) and machine learning (ML) technologies have largely supplanted traditional visual inspection methods, which are often prone to errors and inefficiencies. These modern approaches significantly enhance accuracy and efficiency [13]. By leveraging color, texture, and shape features, classification outcomes have improved markedly, enabling more effective agricultural quality control and streamlined production processes [14,15].

ML provides effective tools for developing precise and reliable classifiers for seed classification. Techniques include artificial neural networks (ANNs), decision trees (DT), logistic regression (LR), fuzzy logic (FL), and genetic algorithms (GAs). ML offers non-linear models capable of predicting feature relationships between input and output layers, typically employed for accurate identification of descriptive attributes in quality assessments [16]. Physical properties can be measured and analyzed using image processing techniques to extract relevant features for classification. Algorithms such as support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF), and multilayer perceptron (MLP) have shown promising results in seed classification tasks [17].

Recent advances in integrating physical measurements with ML algorithms have significantly enhanced data analysis capabilities in agricultural science, food technology, and material characterization. ML models have been used to process spectral and morphological data for classification and prediction tasks with high accuracy [18]. Studies have classified seeds of pepper [19,20], watermelon [21], sunflower [22,23], and wheat [24]. Additionally, ref. [25] employed MLP, RF, Bayes Net (BN), and Logit Boost (LB) algorithms to distinguish six maize seed varieties, with MLP achieving 98.6–99.8% accuracy. Ref. [26] used RF and MLP for binary classification of tomato varieties, obtaining accuracies of 74.50–97.00% for MLP and 79.50–96.25% for RF. Ref. [27] classified soybean seeds using NB, MLP, RF, and SVM, discriminating Traksoy and Ceyhan varieties with 90% accuracy (RF) and 89% (MLP). Ref. [28] classified five rice varieties using KNN, MLP, RF, SVM, LR, and DT, with RF achieving the highest average accuracy (98.04%) using shape and morphological features. However, no studies have combined ML and image processing for vetch seed classification using size and shape characteristics. This study aims to develop binary classification models using MLP and RF algorithms to classify nine common vetch varieties based on size and shape characteristics. The innovative aspect lies in the binary classification of common vetch seeds using comparable physical characteristics through classical analytical methods and ML approaches.

2. Materials and Methods

Nine registered varieties of common vetch seeds were used in this study.

Table 1. Information on the varieties used in the experiments.

Varieties	Institute	Coordinates
Alınoğlu	Field Crops Central Research Institute, Turkiye	39°57′15.0″ N 32°48′45.8″ E
Ankara Moru		39°57′14.9″ N 32°48′46.4″ E
Ayaz		39°57′15.0″ N 32°48′46.3″ E
Alper	Aegean Agricultural Research Institute, Turkiye	38°33′59.8″ N 27°03′09.7″ E
Doruk		38°33′57.2″ N 27°03′09.8″ E
Ürkmez		38°33′57.9″ N 27°03′11.0″ E
Alperen	Directorate of Trakya Agricultural Research Institute, Turkiye	41°38′48.5″ N 26°35′47.7″ E
Kristal 2020		41°38′48.5″ N 26°35′48.4″ E
Özveren	Eastern Mediterranean Agricultural Research Institute, Turkiye	36°51′17.5″ N 35°20′31.3″ E

details the institutions and locations where these varieties were cultivated. Varieties sourced from the Aegean and Mediterranean regions experience mild winters and hot, dry summers, with rainfall mainly during winter months (‘Alper’, ‘Ayaz’, ‘Doruk’, ‘Özveren’). The Tekirdağ and Ankara regions have a continental climate with hot summers and cold winters (‘Alınolu’, ‘Ankara Moru’, ‘Ayaz’, Alperen’, ‘Kristal 2020’). The selected varieties are widely used in Turkey and worldwide and possess similar morphological characteristics, enabling testing of subtle variety-specific distinctions in a realistic manner.

2.1. Determination of Nutritional Properties and Macro-Micro Mineral Content

Seeds were ground using a mill with a 1 mm sieve diameter, and nutritional analyses were conducted in the laboratories of Field Crops and Biosystems Engineering at Erciyes University. Samples were dried at 70 °C for 120 h before grinding. Crude protein analysis followed [29], NDF (%) according to [30], and ADF (%) according to [31]. For NDF, 0.5 g samples were weighed into filter bags, processed in an ANKOM 200 Fiber Analyzer, washed, soaked in acetone for 15 min, dried at 105 °C for 3 h, and weighed. ADF analysis followed a similar procedure. For both, the bag tare, adjusted by a blank factor representing weight change in empty bags under identical conditions, was subtracted from the final bag + residue weight; the residue mass was divided by the initial sample mass (dry matter basis) and multiplied by 100. For tannin analysis, 0.01 g samples were placed in screw-capped tubes, mixed with 6 mL tannin solution, boiled for 1 h, cooled on ice, and absorbance read at 550 nm using a spectrophotometer (UV-vis 1800 Shimadzu, Japan). Absorbance values at 500 nm were recorded, and results calculated as mg catechin/g dry sample using a catechin calibration curve (y = 0.0056x + 0.0193) [32]. Ca, K, P, Mg, Mn, Na, Zn, Fe, B, and Cu contents were determined using ICP-OES (Perkin-Elmer Optima 2100 DV, ICP/OES, Shelton, Connecticut CT 06484-4794, USA) after wet digestion with nitric acid–hydrogen peroxide.

2.2. Determination of the Physical Attributes

Sample mass was determined using a precision electronic scale (Ohaus Adventurer AX224, Australia Eastern Creek NSW, ±0.001 g). The image acquisition system included a digital CCD camera (Nikon D5600, Japan), lens (Nikon AF-S Nikkor, Japan), tripod (Slik Able 300HC Pro, Japan), and lighting apparatus (Life of photo Led 770, China). Seeds were affixed to white cardboard using double-sided adhesive tape and photographed under sufficient light to avoid shadows (Figure 1). The camera was positioned vertically 50 cm high. A ruler on a fiberglass plate established a dimensional calibration (pixels per millimeter). Images were stored as *.tiff files. Using SigmaScan^®Pro 5.0 and Matlab R2024a software, physical characteristics were measured in horizontal and vertical orientations: length (L, mm), width (W, mm), thickness (T, mm), projected area (PA, mm²), equivalent diameter (ED, mm), perimeter (P, mm), shape factor (SF, mm), and compactness (C, %). Formulas were used to determine volume (V, mm³, volume of ellipse), density (ρ, g cm⁻³, m/V), shape index (SI) [33], roundness (R) [34], geometric mean diameter (Dg, mm) [35], surface area (S, mm²) [34], elongation (E) [36], and sphericity (φ, %) [35]. Figure 2 illustrates the segmentation process: background removal, grayscale conversion, inversion, and final cleanup using morphological operations and removal of objects <100 pixels (threshold determined experimentally).

2.3. Binary Classification by ML

Classification was performed using Weka^® v3.9 software on a laptop (Core i7-12650H, 4.70 GHz, 16 GB DDR4 RAM). Binary classification utilized mass, shape, and size features. To develop a low-cost, interpretable model, RF (providing variable importance) and MLP (capturing non-linear decision boundaries) were selected. SVM was also included. Given the limited data size, the study focused on an evidence-based comparison of these paradigms, considering overfitting risks and computational cost, rather than claiming universality. The proposed method offers an economically viable solution implementable with common equipment and software.

Min-max normalization scaled values between 0 and 1. No additional feature selection or reduction was applied. A total of 11,700 data points were used. Model performance was evaluated using 10-fold cross-validation, confusion matrices, and performance metrics. The MLP model employed a 13-11-9 structure. RF models used 100 iterations. Hyperparameters were as follows:

⮚

MLP: batchSize: 100; debug: False; doNotCheckCapabilities: False; decay: False; hiddenLayers: ((attribs + classes)/2); normalizeNumericClass: True; momentum: 0.2; learningRate: 0.3; normalizeAttributes: True; NominalToBinaryFilter: True; validationThreshold: 20; trainingTime: 500; activation function: sigmoid; seed: 0.

⮚

RF: batchSize: 100; number of trees: 100; tree depth: none; breakTiesRandomly: False; doNotCheckCapabilities: False; debug: False; numExecutionSlots: 1; numIterations: 100; seed: 1.

⮚

SVM: batchSize: 100; buildCalibrationModels: False; calibrator: Logistic; doNotCheckCapabilities: False; epsilon: 1.0 × 10⁻¹²; filterType: Normalize training data; kernel: PolyKernel; numFolds: −1; randomSeed: 1; toleranceParameter: 0.001.

Inputs were mass, volume, length, width, thickness, geometric mean diameter, projected area, surface area, sphericity, SI, roundness, aspect ratio, and elongation. Outputs represented the nine different varieties. Inputs were derived from physical/morphometric features, not raw images. Given the limited data size and computational cost, no feature selection or dimensionality reduction was applied to avoid overfitting.

2.4. Evaluation of Models

Classifier performance was evaluated using precision, F-measure (F₁), PRC area, ROC area, true positive rate (TPR), and accuracy, calculated via Equations (1)–(6) [37].

$$Precision=\mathrm{T}\mathrm{P}/(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P})$$

(1)

$$F-Measure\left(or{F}_1\right)=2\mathrm{T}\mathrm{P}/(2\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N})$$

(2)

$$PRCArea=\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{U}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{v}\mathrm{s}.\mathrm{T}\mathrm{P}\mathrm{R}\mathrm{C}\mathrm{u}\mathrm{r}\mathrm{v}\mathrm{e}$$

(3)

$$ROCArea=\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{U}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{T}\mathrm{P}\mathrm{R}\mathrm{v}\mathrm{s}.\mathrm{F}\mathrm{P}\mathrm{R}\mathrm{C}\mathrm{u}\mathrm{r}\mathrm{v}\mathrm{e}$$

(4)

$$Accuracy=(\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N})/(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N})$$

(5)

$$TPR=\mathrm{T}\mathrm{P}/(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N})$$

(6)

2.5. Multivariate Analysis

One-way ANOVA and Tukey’s test (p < 0.05) determined significant differences. Linear discriminant analysis centroids were used for scatter plots. MANOVA and Hotelling’s pairwise comparisons with Bonferroni correction and squared Mahalanobis distances assessed similarity/dissimilarity across varieties. Analyses used PAST v3.20 and SPSS v20.0. ROC curve data were consolidated into a single response variable to compare classification methods statistically [38,39]. The area under the ROC curve (0–1) served as a performance metric [40], with values near 1 indicating optimal discrimination.

3. Results and Discussion

3.1. Nutritional Properties and Macro–Micro Mineral Content of Different Varieties of Vetch

Crude protein content among common vetch varieties is shown in

Table 2. Nutritional properties of common vetch genotypes (%).

Varieties	Crude Protein	ADF	NDF	Tannin
Alper	22.66a	9.06bc	13.65bc	3.06abc
Alperen	16.93e	10.79ab	16.47a	2.76e
Alınoğlu	20.91bc	11.36a	14.69ab	3.12a
Ankara Moru	19.71cd	8.57bc	11.92cd	2.97bcd
Ayaz	20.61bc	9.03bc	12.04cd	0.48f
Doruk	18.80d	8.94bc	15.25ab	3.10ab
Kristal 2020	14.69f	9.44abc	14.68ab	2.96cd
Özveren	21.05bc	8.19c	11.25de	3.07abc
Ürkmez	21.96ab	8.74bc	9.57e	2.90d
Mean	19.70	9.35	13.28	2.71

Note: Means indicated with different letters in the same column are significantly different (p < 0.05).

. Alper had the highest protein content (22.66%), while Kristal 2020 had the lowest (14.69%). Significant variations were observed. Common vetch, as a legume, contains substantial crude protein. Ref. [6] reported 15–33.7% crude protein across 388 accessions from Turkey and Middle Eastern countries. Ref. [9] found an average of 342 g/kg in 44 European accessions, with NDF and ADF averages of 121 and 69.8 g/kg dry matter, respectively. Ref. [41] reported 347–374 g/kg on the Tibetan Plateau. Ref. [42] reported 299–323 g/kg in 14 accessions from Serbia. In this study, NDF ranged from 9.57% to 16.47%, and ADF from 8.19% to 11.36%. Ref. [41] reported NDF and ADF ranges of 207–229 g/kg and 59.5–68.8 g/kg dry matter, respectively.

Tannins can have beneficial or detrimental effects depending on animal species, physiological state, and consumption levels [43]. Tannin concentrations ranged from 0.48% to 3.12%, exceeding the 0.13–1.07% reported by [44] for various vetch species. High tannin levels can reduce feed intake and protein/fiber digestion, affecting animal production [45], though low to moderate condensed tannin (20–45 g/kg) can improve feed utilization efficiency [46].

Macro elements ranged as follows: Ca (1220.49–2267.34 mg/kg), K (5061–7674.08 mg/kg), Mg (971.09–1448.73 mg/kg), P (4163.81–5697.81 mg/kg), and S (1258.39–1678.47 mg/kg) (

Table 3. Macro–micro mineral content of common vetch genotypes (mg kg⁻¹).

Varieties	B	Ca	Cu	Fe	K	Mg	Mn	Na	P	S	Zn
Alper	5.00a	1639.28c	93.49a	52.65e	7674.08a	1327.83c	16.32c	335.92a	5393.60b	1678.47a	43.52c
Alperen	3.22i	1220.49i	62.60i	39.81h	5061.00i	971.09i	22.56a	246.39i	4163.81i	1258.39i	40.64d
Alınoğlu	4.08e	1884.35b	87.35c	56.45c	6627.64c	1342.85b	14.99d	319.86c	5265.08d	1567.99e	37.78f
Ankara Moru	3.67f	2267.34a	79.52f	78.17a	5452.79h	1232.32e	16.65b	270.44h	4380.71h	1301.24h	31.20h
Ayaz	4.14d	1458.75h	83.80d	37.61i	5778.38g	1125.28g	13.12f	278.92f	4449.36g	1644.61b	30.67i
Doruk	4.42b	1605.06d	91.01b	46.81f	7225.46b	1448.73a	14.46e	334.67b	5697.81a	1602.28c	44.40b
Kristal 2020	3.27h	1543.21g	80.87e	60.65b	6230.40d	1110.02h	12.75h	274.15g	4960.85f	1578.85d	49.31a
Özveren	4.34c	1583.38f	68.70h	43.00g	6070.57f	1229.74f	12.43i	306.14d	5268.80c	1404.84g	38.81e
Ürkmez	3.61g	1583.66e	77.03g	53.05d	6093.19e	1264.69d	12.94g	298.77e	5034.80e	1434.51f	37.37g
Mean	3.97	1642.84	80.49	52.02	6245.95	1228.06	15.14	296.14	4957.20	1496.80	39.30

Note: Means indicated with different letters in the same column are significantly different (p < 0.05).

). Ref. [5] reported Ca, K, Mg, and P as 0.94–1.20, 9.69–9.98, 2.20–2.66, and 1.63–3.96 g/kg, respectively; ref. [6] reported 610–1973 mg/kg, 0.98–1.51%, 1139–2129 mg/kg, and 3657–6010 mg/kg. The Ca:P ratio in feed should be 1:1 to 2:1 [47]; here, it varied between 0.28 and 0.52, similar to the 0.24 and 0.73 reported by [5]. Micro elements ranged: B (3.22–5 mg/kg), Cu (62.6–93.49 mg/kg), Fe (37.61–78.17 mg/kg), Mn (12.43–22.56 mg/kg), Zn (30.67–49.31 mg/kg). Cu concentrations exceeded those reported by [9] (5.16 mg/kg), [5] (3.37–8.77 mg/kg), and [6] (1.3–16.7 mg/kg). Fe values were similar to [9] (59.90 mg/kg). Mn and Zn contents aligned with [5]. Variations may arise from plant type, grain maturity at harvest, and environmental factors [41,48].

3.2. Seed Physical Attributes of the Varieties

Physical properties are reported in

Table 4. Physical properties of common vetch genotypes.

Varieties	Mass (M, g)	Volume (V, mm3)	Density (g cm−3)	Length (L, mm)	Width (W, mm)	Thickness (T, mm)	Geometric Mean Diam. (Dg, mm)	Projected Area (PA, mm2)	Surface Area (SA, mm2)
Alper	0.07 ± 0.01c	48.38 ± 8.45e	1.45 ± 0.29b	5.18 ± 0.40efg	4.66 ± 0.37c	3.80 ± 0.28e	4.50 ± 0.27e	16.00 ± 1.89e	63.98 ± 7.56e
Alperen	0.08 ± 0.02ab	50.71 ± 9.08de	1.58 ± 0.39a	5.16 ± 0.40fg	4.60 ± 0.28c	4.05 ± 0.28bc	4.58 ± 0.26de	16.51 ± 1.93de	66.03 ± 7.73de
Alınoğlu	0.08 ± 0.02b	60.75 ± 10.58b	1.32 ± 0.32c	5.67 ± 0.46ab	4.86 ± 0.38b	4.19 ± 0.33ab	4.86 ± 0.29ab	18.62 ± 2.18b	74.48 ± 8.74b
Ankara Moru	0.06 ± 0.01e	52.07 ± 7.57cde	1.15 ± 0.21d	5.29 ± 0.32def	4.58 ± 0.28c	4.08 ± 0.24bc	4.62 ± 0.22cde	16.82 ± 1.62cde	67.28 ± 6.50cde
Ayaz	0.08 ± 0.02b	68.80 ± 22.42a	1.16 ± 0.42d	5.73 ± 0.74a	5.13 ± 0.71a	4.27 ± 0.69a	5.00 ± 0.68a	19.99 ± 4.83a	79.97 ± 19.33a
Doruk	0.07 ± 0.02cd	54.80 ± 8.61cd	1.28 ± 0.34cd	5.45 ± 0.37cd	4.90 ± 0.31b	3.90 ± 0.22de	4.70 ± 0.24cd	17.40 ± 1.81cd	69.59 ± 7.26cd
Kristal 2020	0.07 ± 0.01cd	48.96 ± 8.85e	1.49 ± 0.33b	5.02 ± 0.41g	4.58 ± 0.32c	4.03 ± 0.29cd	4.52 ± 0.27e	16.12 ± 1.93e	64.49 ± 7.73e
Özveren	0.08 ± 0.01a	56.89 ± 6.59bc	1.41 ± 0.21bc	5.36 ± 0.25cde	5.01 ± 0.26ab	4.03 ± 0.23cd	4.76 ± 0.19bc	17.86 ± 1.39bc	71.43 ± 5.55bc
Ürkmez	0.06 ± 0.01de	56.16 ± 8.42bc	1.07 ± 0.23e	5.49 ± 0.32bc	4.88 ± 0.28b	3.98 ± 0.30cd	4.74 ± 0.24bc	17.68 ± 1.79bc	70.74 ± 7.15bc
Mean	0.07 ± 0.02	55.28 ± 12.55	1.32 ± 0.34	5.37 ± 0.48	4.80 ± 0.42	4.04 ± 0.37	4.70 ± 0.36	17.44 ± 2.65	69.78 ± 10.58
F-values	36.18 **	34.50 **	32.64 **	30.57 **	28.56 **	16.19 **	24.74 **	29.60 **	29.60 **

Note: Means indicated with different letters in the same column are significantly different (p < 0.05), and **: significant at p < 0.01.

. Ayaz had the greatest values for length (5.73 mm), width (5.13 mm), thickness (4.27 mm), volume (68.80 mm³), geometric mean diameter (5.00 mm), projection area (19.99 mm²), and surface area (79.97 mm²). Alınoğlu followed with high values for length (5.67 mm), volume (60.75 mm³), thickness (4.19 mm), surface area (74.48 mm²), geometric mean diameter (4.86 mm), and projection area (18.62 mm²). Alperen had the highest density (1.58 g cm⁻³), while Ürkmez had the lowest (1.07 g cm⁻³). Özveren had the second-highest width. Single seed mass was 0.08 g for Alperen, Alınoğlu, Ayaz, and Özveren; 0.07 g for Alper, Doruk, and Kristal; and 0.06 g for Ürkmez. Width varied between 5.73 and 5.02 mm. Alper had the lowest projection and surface areas (16.00 mm² and 63.98 mm², respectively).

Shape attributes are presented in

Table 5. Shape characteristics of different varieties of vetch.

Varieties	Sphericity (S, %)	Shape Index (SI)	Roundness (R)	Aspect Ratio (AR)	Elongation (E)
Alper	87.10 ± 3.47d	1.23 ± 0.07ab	0.76 ± 0.06c	0.74 ± 0.07de	1.37 ± 0.12ab
Alperen	88.98 ± 3.92ab	1.19 ± 0.08bc	0.79 ± 0.07ab	0.79 ± 0.07ab	1.28 ± 0.11d
Alınoğlu	86.06 ± 5.23d	1.26 ± 0.12a	0.74 ± 0.09c	0.74 ± 0.08de	1.36 ± 0.15ab
Ankara Moru	87.45 ± 3.68bcd	1.22 ± 0.08ab	0.77 ± 0.06bc	0.77 ± 0.05bc	1.30 ± 0.10cd
Ayaz	87.25 ± 4.96cd	1.23 ± 0.11ab	0.76 ± 0.08bc	0.74 ± 0.08cde	1.36 ± 0.16ab
Doruk	86.44 ± 2.86d	1.24 ± 0.06a	0.75 ± 0.05c	0.72 ± 0.05e	1.40 ± 0.09a
Kristal 2020	90.28 ± 4.15a	1.17 ± 0.08c	0.82 ± 0.07a	0.81 ± 0.07a	1.25 ± 0.12d
Özveren	88.92 ± 2.04abc	1.19 ± 0.04c	0.79 ± 0.04ab	0.75 ± 0.05cd	1.33 ± 0.08bc
Ürkmez	86.38 ± 3.66d	1.24 ± 0.08a	0.75 ± 0.06c	0.73 ± 0.06de	1.39 ± 0.11a
Mean	87.65 ± 4.10	1.22 ± 0.09	0.77 ± 0.07	0.75 ± 0.07	1.34 ± 0.13
F-values	13.57 **	12.62 **	13.80 **	20.79 **	18.96 **

Note: Means with different letters within a column differ significantly (p < 0.05), and **: significant at p < 0.01.

. Kristal had the highest sphericity (90.28%), while Alınoğlu had the lowest (86.06%). Roundness was similar for Ayaz-Alper (0.76), Alperen-Özveren (0.79), and Doruk-Ürkmez (0.75). Kristal had the highest roundness (0.82). Values near 1 indicate nearly spherical seeds. Alınoğlu, with SI > 1.26, was defined as oval. Kristal had the highest aspect ratio (0.81). The highest elongation values were for Doruk (1.40), Ürkmez (1.39), and Alper (1.37); the lowest was for Kristal (1.25). Form index values increased as sphericity and roundness decreased. Ref. [49] reported length (5.19 mm) and width (4.33 mm) for vetch seeds at different moisture contents (10.57–20.63%), consistent with this study, though the projection area (23.52–29.05 mm²) was higher. Ref. [50] reported projection areas of 15.33–17.66 mm² for lentil seeds at 15.6–22.5% moisture. Ref. [51] reported the mean length and width of 4.88 mm and 4.40 mm, respectively. Ref. [27] reported mean projection area and volume for soybean seeds as 31.53 mm² and 133.84 mm³, with elongation and SI of 1.30 and 1.20. Differences may reflect species variation.

3.3. Linear Discrimination Analysis, Pairwise Comparison, and Multivariate Tests

Discriminant analysis of physical features is shown in Figure 3. Eigenvalues for functions 1–8 were 1.087, 0.360, 0.302, 0.156, 0.045, 0.035, 0.003, and 0.000, respectively. The first four functions explained 95.8% of the variance (54.6%, 18.1%, 15.2%, and 7.9%). Wilks’ lambda indicated that 21.6% of the between-group differences remained unexplained. Discriminant function coefficients determined predictor importance.

Group centers based on canonical discriminant functions are shown in Figure 4. Length, geometric mean diameter, sphericity, and SI were key discriminative features. For Doruk and Ürkmez, geometric mean diameter and length validated positioning on canonical function 1; for Ankara Moru and Ayaz, sphericity, form index, and roundness corroborated positioning on canonical function 2. Ref. [52] used PCA for dry bean quality traits, with the first three principal components explaining 70% of variance under non-irrigated conditions, and PC1 and PC2 representing 100% under drip irrigation. In this application, PC1 explained 70.16% of the overall variation and 29.84% by PC2.

Wilks’ Lambda, Hotelling Trace, and Pillai Trace indicated significant differences in physical attributes among varieties (p < 0.01).

Table 6. Differences among the varieties.

Eigenvalue Statistics	Function 1		Function 2			Function 3			Function 4		Function 5			Function 6			Function 7			Function 8
Eigenvalues	1.087		0.360			0.302			0.156		0.045			0.035			0.003			0.000
% of variance	54.6		18.1			15.2			7.9		2.3			1.8			0.2			0.0
% of cumulative variance	54.6		72.7			87.9			95.8		98.1			99.8			100.0			100.0
Canonical correlation	0.722		0.514			0.482			0.368		0.208			0.185			0.055			0.012
MANOVA results
Effect	Statistics					Value		Hypothesis DF					Error DF			F			p (sigma)
Variables	Pillai’s trace					1.346		96					7096			14.95			0.000 **
	Wilks’ Lambda					0.187		96					5938			17.48			0.000 **
	Hotelling Trace					2.182		96					7026			19.97			0.000 **
Hotelling’s pairwise comparisons. Bonferroni corrected p values in upper triangle. Mahalanobis distances in lower triangle
Varieties	1	2		3			4			5		6			7			8			9
Alper	-	1.05 × 10−9		9.38 × 10−17			4.84 × 10−11			2.33 × 10−47		2.35 × 10−2			2.66 × 10−8			9.99 × 10−15			1.09 × 10−6
Alperen	1.99	-		7.10 × 10−12			5.90 × 10−13			1.21 × 10−48		1.02 × 10−16			1.33 × 10−4			2.19 × 10−15			1.42 × 10−17
Alınoğlu	3.24	2.35		-			2.14 × 10−17			8.26 × 10−39		5.32 × 10−16			3.55 × 10−19			1.52 × 10−17			1.79 × 10−15
Ankara Moru	2.21	2.54		3.37			-			6.84 × 10−53		1.00 × 10−15			7.23 × 10−10			5.82 × 10−30			3.70 × 10−10
Ayaz	11.87	12.39		8.79			14.24			-		6.88 × 10−48			1.62 × 10−50			5.24 × 10−49			1.20 × 10−47
Doruk	0.81	3.24		3.10			3.05			12.08		-			6.32 × 10−17			2.18 × 10−11			1.09 × 10+00
Kristal 2020	1.76	1.17		3.72			2.01			13.18		3.28			-			1.82 × 10−24			5.28 × 10−16
Özveren	2.86	2.98		3.40			6.21			12.54		2.27			4.86			-			6.65 × 10−20
Ürkmez	1.50	3.40		3.00			2.06			11.98		0.53			3.10			3.87			-

Note: ** Highly significant (p < 0.01).

shows MANOVA, Bonferroni correction, and Mahalanobis distance values. Wilks’ Lambda < 1 indicates increasing between-group differences. Pillai Trace, considered the most accurate multivariate statistic, accounts for all variances explaining independent factors in dependent variables. Here, Pillai trace, Wilks’ Lambda, and Hotelling Trace were 1.346, 0.187, and 2.182, respectively. Varieties with a Mahalanobis distance < 3 had highly similar features. Alper and Alınoğlu, with minimal Mahalanobis distance, shared analogous traits. The greatest distance was between Ankara Moru and Ayaz. Bonferroni-adjusted P-values supported these findings.

3.4. Performance Results of Binary Classification

MLP, RF, and SVM models were developed for pairwise classification based on shape, size, and mass attributes. Confusion matrices and performance metrics for Alper are in

Table 7. The confusion matrices and performance metrics for Alper variety.

Classifiers	Predicted		Actual	Accuracy (%)	TPR	Precision	F1	ROC	PRC
Alper vs. Alperen
MLP	Alper	Alperen	-	-	-	-	-	-	-
	76	24	Alper	77.00	0.760	0.776	0.768	0.788	0.773
	22	78	Alperen	-	0.780	0.765	0.772	0.788	0.732
RF	Alper	Alperen	-	-	-	-	-	-	-
	74	26	Alper	77.50	0.740	0.796	0.767	0.824	0.842
	19	81	Alperen	-	0.810	0.757	0.775	0.823	0.804
SVM	Alper	Alperen	-	-	-	-	-	-	-
	80	20	Alper	74.50	0.800	0.721	0.758	0.745	0.677
	31	69	Alperen	-	0.690	0.775	0.730	0.745	0.690
Alper vs. Alınoğlu
MLP	Alper	Alınoğlu	-	-	-	-	-	-	-
	81	19	Alper	79.50	0.810	0.786	0.798	0.851	0.849
	22	78	Alınoğlu	-	0.780	0.804	0.792	0.851	0.831
RF	Alper	Alınoğlu	-	-	-	-	-	-	-
	78	22	Alper	77.00	0.780	0.765	0.772	0.858	0.873
	24	76	Alınoğlu	-	0.760	0.776	0.768	0.858	0.802
SVM	Alper	Alınoğlu	-	-	-	-	-	-	-
	83	17	Alper	78.50	0.830	0.761	0.794	0.785	0.717
	26	74	Alınoğlu	-	0.740	0.813	0.775	0.785	0.732
Alper vs. A. Moru
MLP	Alper	A. Moru	-	-	-	-	-	-	-
	74	26	Alper	73.50	0.740	0.733	0.736	0.814	0.822
	27	73	A. Moru	-	0.730	0.737	0.734	0.814	0.756
RF	Alper	A. Moru	-	-	-	-	-	-	-
	77	23	Alper	78.50	0.770	0.794	0.782	0.863	0.886
	20	80	A. Moru	-	0.800	0.777	0.788	0.863	0.812
SVM	Alper	A. Moru	-	-	-	-	-	-	-
	65	35	Alper	72.50	0.650	0.765	0.703	0.725	0.672
	20	80	A. Moru	-	0.800	0.696	0.744	0.725	0.657
Alper vs. Ayaz
MLP	Alper	Ayaz	-	-	-	-	-	-	-
	91	9	Alper	85.00	0.910	0.813	0.858	0.909	0.887
	21	79	Ayaz	-	0.790	0.898	0.840	0.909	0.886
RF	Alper	Ayaz	-	-	-	-	-	-	-
	89	11	Alper	87.00	0.890	0.856	0.879	0.925	0.891
	15	85	Ayaz	-	0.850	0.885	0.867	0.928	0.939
SVM	Alper	Ayaz	-	-	-	-	-	-	-
	92	8	Alper	82.00	0.920	0.767	0.836	0.820	0.745
	28	72	Ayaz	-	0.720	0.900	0.800	0.820	0.788
Alper vs. Doruk
MLP	Alper	Doruk	-	-	-	-	-	-	-
	69	31	Alper	65.00	0.690	0.639	0.663	0.685	0.676
	39	61	Doruk	-	0.610	0.651	0.649	0.685	0.686
RF	Alper	Doruk	-	-	-	-	-	-	-
	57	43	Alper	60.00	0.570	0.606	0.588	0.636	0.621
	37	63	Doruk	-	0.630	0.594	0.612	0.636	0.643
SVM	Alper	Doruk	-	-	-	-	-	-	-
	60	40	Alper	60.00	0.600	0.600	0.597	0.595	0.557
	40	60	Doruk	-	0.600	0.600	0.597	0.595	0.557
Alper vs. Kristal
MLP	Alper	Kristal	-	-	-	-	-	-	-
	78	22	Alper	72.00	0.780	0.696	0.736	0.738	0.716
	34	66	Kristal	-	0.660	0.750	0.702	0.738	0.721
RF	Alper	Kristal	-	-	-	-	-	-	-
	71	29	Alper	72.00	0.710	0.724	0.717	0.747	0.695
	27	73	Kristal	-	0.730	0.716	0.723	0.747	0.713
SVM	Alper	Kristal	-	-	-	-	-	-	-
	74	26	Alper	68.50	0.740	0.667	0.701	0.685	0.623
	37	63	Kristal	-	0.630	0.708	0.667	0.685	0.631
Alper vs. Özveren
MLP	Alper	Özveren	-	-	-	-	-	-	-
	78	22	Alper	78.00	0.780	0.780	0.780	0.823	0.824
	22	78	Özveren	-	0.780	0.780	0.780	0.823	0.756
RF	Alper	Özveren	-	-	-	-	-	-	-
	75	25	Alper	75.50	0.750	0.758	0.754	0.831	0.821
	24	76	Özveren	-	0.760	0.752	0.756	0.831	0.774
SVM	Alper	Özveren	-	-	-	-	-	-	-
	79	21	Alper	80.50	0.790	0.814	0.802	0.805	0.748
	18	82	Özveren	-	0.820	0.796	0.808	0.805	0.743
Alper vs. Ürkmez
MLP	Alper	Ürkmez	-	-	-	-	-	-	-
	65	35	Alper	69.00	0.650	0.707	0.677	0.759	0.745
	27	73	Ürkmez	-	0.730	0.676	0.702	0.759	0.753
RF	Alper	Ürkmez	-	-	-	-	-	-	-
	67	33	Alper	68.50	0.670	0.691	0.680	0.736	0.713
	30	70	Ürkmez	-	0.700	0.680	0.690	0.736	0.705
SVM	Alper	Ürkmez	-	-	-	-	-	-	-
	67	33	Alper	69.50	0.670	0.705	0.687	0.695	0.638
	28	72	Ürkmez	-	0.720	0.686	0.702	0.695	0.634

Note: As blues darken, true values increase; as grays darken, false values increase.

. RF achieved 77.50% accuracy for Alper vs. Alperen, MLP 77.00%. Other metrics confirmed these results. For RF, TPR, precision, F₁, ROC, and PRC were 0.740, 0.796, 0.767, 0.824, 0.842 for Alper and 0.810, 0.757, 0.775, 0.823, 0.804 for Alperen. SVM had the lowest accuracy (74.50%). For Alper vs. Alınoğlu, MLP achieved 79.50% accuracy (TPR: 0.810 Alper, 0.780 Alınoğlu; precision: 0.786, 0.804; F₁: 0.798, 0.792; PRC: 0.849, 0.831; ROC: 0.851, both). RF achieved 78.50% for Alper vs. Ankara Moru, while MLP and SVM were lower (73.50%, 72.50%). For Alper vs. Ayaz, RF achieved 89.00%, MLP 85.00%, SVM 82.00%. For Alper vs. Doruk, MLP achieved 65.00%, RF and SVM 60.00%. For Alper vs. Kristal, MLP and RF achieved 72.00%, SVM 68.50%. For Alper vs. Özveren, SVM achieved 80.50%, MLP 78.00%, and RF 75.50%. For Alper vs. Ürkmez, SVM achieved 69.50%, MLP 69.00%, and RF 68.50%.

For Alperen vs. Ayaz, RF achieved 88.00% accuracy, MLP achieved 86.00% accuracy, and SVM achieved 81.00% accuracy. For Alperen vs. Kristal, MLP had the lowest accuracy (60.50%), RF achieved 62.00% accuracy, SVM achieved 67.50% accuracy. In MLP, TPR was 0.620 Alperen and 0.590 Kristal; precision 0.602, 0.608; F₁ 0.611, 0.599; PRC 0.718, 0.695; ROC 0.698 both. MLP correctly classified 62% of Alperen, misclassifying 38% as Kristal (see

Table 8. The confusion matrices and performance metrics for Alperen variety.

Classifiers	Predicted		Actual	Accuracy (%)	TPR	Precision	F1	ROC	PRC
Alperen vs. Alınoğlu
MLP	Alperen	Alınoğlu	-	-	-	-	-	-	-
	74	26	Alperen	72.50	0.740	0.718	0.729	0.772	0.741
	29	71	Alınoğlu	-	0.710	0.732	0.721	0.772	0.771
RF	Alperen	Alınoğlu	-	-	-	-	-	-	-
	74	26	Alperen	75.00	0.740	0.755	0.747	0.767	0.735
	24	76	Alınoğlu	-	0.760	0.750	0.750	0.767	0.730
SVM	Alperen	Alınoğlu	-	-	-	-	-	-	-
	79	21	Alperen	76.00	0.790	0.745	0.767	0.760	0.694
	27	73	Alınoğlu	-	0.790	0.777	0.753	0.760	0.702
Alperen vs. A. Moru
MLP	Alperen	A. Moru	-	-	-	-	-	-	-
	74	26	Alperen	77.00	0.740	0.787	0.763	0.831	0.853
	20	80	A. Moru	-	0.800	0.755	0.777	0.831	0.779
RF	Alperen	A. Moru	-	-	-	-	-	-	-
	75	25	Alperen	77.00	0.750	0.781	0.765	0.859	0.869
	21	79	A. Moru	-	0.790	0.760	0.775	0.859	0.835
SVM	Alperen	A. Moru	-	-	-	-	-	-	-
	65	35	Alperen	79.50	0.650	0.915	0.760	0.795	0.770
	6	94	A. Moru	-	0.940	0.729	0.821	0.795	0.715
Alperen vs. Ayaz
MLP	Alperen	Ayaz	-	-	-	-	-	-	-
	88	12	Alperen	86.00	0.880	0.846	0.863	0.942	0.948
	16	84	Ayaz	-	0.840	0.875	0.857	0.942	0.941
RF	Alperen	Ayaz	-	-	-	-	-	-	-
	89	11	Alperen	88.00	0.890	0.873	0.881	0.925	0.905
	13	87	Ayaz	-	0.870	0.880	0.880	0.925	0.935
SVM	Alperen	Ayaz	-	-	-	-	-	-	-
	89	11	Alperen	81.00	0.890	0.767	0.824	0.810	0.738
	27	73	Ayaz	-	0.730	0.869	0.793	0.810	0.769
Alperen vs. Doruk
MLP	Alperen	Doruk	-	-	-	-	-	-	-
	79	21	Alperen	77.50	0.790	0.767	0.778	0.879	0.874
	24	76	Doruk	-	0.760	0.784	0.771	0.879	0.895
RF	Alperen	Doruk	-	-	-	-	-	-	-
	85	15	Alperen	82.00	0.850	0.802	0.825	0.900	0.903
	21	79	Doruk	-	0.790	0.840	0.814	0.900	0.896
SVM	Alperen	Doruk	-	-	-	-	-	-	-
	79	21	Alperen	80.50	0.790	0.814	0.802	0.805	0.748
	18	82	Doruk	-	0.820	0.796	0.808	0.805	0.743
Alperen vs. Kristal
MLP	Alperen	Kristal	-	-	-	-	-	-	-
	62	38	Alperen	60.50	0.620	0.602	0.611	0.698	0.718
	41	59	Kristal		0.590	0.608	0.599	0.698	0.695
RF	Alperen	Kristal	-	-	-	-	-	-	-
	62	38	Alperen	62.00	0.620	0.620	0.620	0.678	0.646
	38	62	Kristal	-	0.620	0.620	0.620	0.678	0.682
SVM	Alperen	Kristal	-	-	-	-	-	-	-
	58	42	Alperen	67.50	0.580	0.716	0.641	0.675	0.625
	23	77	Kristal	-	0.770	0.647	0.703	0.675	0.613
Alperen vs. Özveren
MLP	Alperen	Özveren	-	-	-	-	-	-	-
	78	22	Alperen	76.50	0.780	0.757	0.768	0.849	0.842
	25	75	Özveren	-	0.750	0.773	0.761	0.849	0.870
RF	Alperen	Özveren	-	-	-	-	-	-	-
	74	26	Alperen	74.00	0.740	0.740	0.740	0.828	0.806
	26	74	Özveren	-	0.740	0.740	0.740	0.828	0.845
SVM	Alperen	Özveren	-	-	-	-	-	-	-
	81	19	Alperen	79.00	0.810	0.779	0.794	0.790	0.726
	23	77	Özveren	-	0.770	0.802	0.786	0.790	0.733
Alperen vs. Ürkmez
MLP	Alperen	Ürkmez	-	-	-	-	-	-	-
	81	19	Alperen	80.00	0.810	0.794	0.802	0.904	0.913
	21	79	Ürkmez	-	0.790	0.806	0.798	0.904	0.910
RF	Alperen	Ürkmez	-	-	-	-	-	-	-
	80	20	Alperen	81.50	0.800	0.825	0.812	0.901	0.914
	17	83	Ürkmez	-	0.830	0.806	0.818	0.901	0.883
SVM	Alperen	Ürkmez	-	-	-	-	-	-	-
	76	24	Alperen	82.00	0.760	0.864	0.809	0.820	0.776
	12	88	Ürkmez	-	0.880	0.786	0.830	0.820	0.751

Note: As blues darken, true values increase; as grays darken, false values increase.

). SVM achieved the highest accuracy for Alperen vs. Ürkmez (82.00%) and Alperen vs. Doruk (80.50%) but the lowest for Alperen vs. Alınoğlu (76.00%) and Alperen vs. Özveren (79.00%).

Confusion matrices and performance metrics for Alınoğlu are in

Table 9. The confusion matrices and performance metrics for Alınoğlu varieties.

Classifiers	Predicted		Actual	Accuracy (%)	TPR	Precision	F1	ROC	PRC
Alınoğlu vs. A. Moru
MLP	Alınoğlu	A. Moru	-	-	-	-	-	-
	80	20	Alınoğlu	80.50	0.800	0.808	0.804	0.886	0.865
	19	81	A. Moru	-	0.810	0.802	0.806	0.886	0.880
RF	Alınoğlu	A. Moru	-	-	-	-	-	-	-
	82	18	Alınoğlu	80.00	0.820	0.788	0.804	0.867	0.848
	22	78	A. Moru	-	0.780	0.813	0.796	0.867	0.839
SVM	Alınoğlu	A. Moru	-	-	-	-	-	-	-
	83	17	Alınoğlu	83.00	0.830	0.830	0.830	0.830	0.774
	17	83	A. Moru	-	0.830	0.830	0.830	0.830	0.774
Alınoğlu vs. Ayaz
MLP	Alınoğlu	Ayaz	-	-	-	-	-	-	-
	76	24	Alınoğlu	73.50	0.760	0.724	0.741	0.826	0.813
	29	71	Ayaz	-	0.710	0.747	0.728	0.826	0.834
RF	Alınoğlu	Ayaz	-	-	-	-	-	-	-
	78	22	Alınoğlu	75.00	0.780	0.736	0.757	0.832	0.800
	28	72	Ayaz	-	0.720	0.766	0.742	0.832	0.829
SVM	Alınoğlu	Ayaz	-	-	-	-	-	-	-
	78	22	Alınoğlu	71.00	0.780	0.684	0.729	0.710	0.644
	36	64	Ayaz	-	0.640	0.744	0.688	0.710	0.656
Alınoğlu vs. Doruk
MLP	Alınoğlu	Doruk	-	-	-	-	-	-	-
	76	24	Alınoğlu	78.50	0.760	0.800	0.779	0.853	0.859
	19	81	Doruk	-	0.810	0.771	0.790	0.853	0.833
RF	Alınoğlu	Doruk	-	-	-	-	-	-	-
	75	25	Alınoğlu	77.00	0.750	0.781	0.765	0.840	0.833
	21	79	Doruk	-	0.790	0.760	0.775	0.840	0.832
SVM	Alınoğlu	Doruk	-	-	-	-	-	-	-
	71	29	Alınoğlu	77.50	0.710	0.816	0.759	0.775	0.724
	16	84	Doruk	-	0.840	0.743	0.789	0.775	0.704
Alınoğlu vs. Kristal
MLP	Alınoğlu	Kristal	-	-	-	-	-	-	-
	84	16	Alınoğlu	79.50	0.840	0.771	0.804	0.860	0.846
	25	75	Kristal	-	0.750	0.824	0.785	0.860	0.854
RF	Alınoğlu	Kristal	-	-	-	-	-	-	-
	76	24	Alınoğlu	76.50	0.760	0.768	0.764	0.830	0.836
	23	77	Kristal	-	0.770	0.762	0.766	0.830	0.801
SVM	Alınoğlu	Kristal	-	-	-	-	-	-	-
	88	12	Alınoğlu	83.50	0.880	0.807	0.842	0.835	0.770
	21	79	Kristal	-	0.790	0.868	0.827	0.835	0.791
Alınoğlu vs. Özveren
MLP	Alınoğlu	Özveren	-	-	-	-	-	-	-
	73	27	Alınoğlu	73.50	0.730	0.737	0.734	0.832	0.807
	26	74	Özveren	-	0.740	0.733	0.736	0.832	0.833
RF	Alınoğlu	Özveren	-	-	-	-	-	-	-
	75	25	Alınoğlu	76.00	0.750	0.765	0.758	0.832	0.823
	23	77	Özveren	-	0.770	0.755	0.762	0.832	0.836
SVM	Alınoğlu	Özveren	-	-	-	-	-	-	-
	69	31	Alınoğlu	73.00	0.690	0.750	0.719	0.730	0.673
	23	77	Özveren	-	0.770	0.713	0.740	0.730	0.664
Alınoğlu vs. Ürkmez
MLP	Alınoğlu	Ürkmez	-	-	-	-	-	-	-
	73	27	Alınoğlu	75.50	0.730	0.768	0.749	0.843	0.865
	22	78	Ürkmez	-	0.780	0.743	0.761	0.843	0.821
RF	Alınoğlu	Ürkmez	-	-	-	-	-	-	-
	71	29	Alınoğlu	73.00	0.710	0.740	0.724	0.818	0.830
	25	75	Ürkmez	-	0.750	0.721	0.735	0.818	0.804
SVM	Alınoğlu	Ürkmez	-	-	-	-	-	-	-
	74	26	Alınoğlu	76.50	0.740	0.779	0.759	0.765	0.706
	21	79	Ürkmez	-	0.790	0.752	0.771	0.765	0.699

Note: As blues darken, true values increase; as grays darken, false values increase.

. For Alınoğlu vs. Ankara Moru, MLP achieved 80.50% accuracy, RF achieved 80.00% accuracy, SVM achieved 83.00% accuracy. In MLP, TPR was 0.800 Alınoğlu, 0.810 Ankara Moru; precision 0.808, 0.802; F₁ 0.804, 0.806; PRC 0.865, 0.880; ROC 0.886 both. For Alınoğlu vs. other varieties, MLP accuracy ranged from 73.50 to 80.50%, RF ranged from 73.00 to 80.00%, SVM ranged from 71.00 to 83.50%. SVM performed best for Alınoğlu vs. Kristal and Alınoğlu vs. Doruk, worst for Alınoğlu vs. Ayaz and Alınoğlu vs. Özveren.

Ankara Moru and Ayaz were among the most distinguishable pairs: RF 90.00%, MLP 89.00%, SVM 82.50%. RF achieved its highest accuracy (90.50%) for Ankara Moru vs. Özveren; MLP and SVM both achieved 88.00%. For RF, TPR was 0.910 Ankara Moru, 0.900 Özveren; precision 0.901, 0.909; F₁ 0.905 both; PRC 0.965, 0.968; ROC 0.966 both (

Table 10. The confusion matrices and performance metrics for Ankara Moru varieties.

Classifiers	Predicted		Actual	Accuracy (%)	TPR	Precision	F1	ROC	PRC
A. Moru vs. Ayaz
MLP	A. Moru	Ayaz	-	-	-	-	-	-	-
	93	7	A. Moru	89.00	0.930	0.861	0.894	0.909	0.891
	15	85	Ayaz	-	0.850	0.924	0.885	0.909	0.908
RF	A. Moru	Ayaz	-	-	-	-	-	-	-
	92	8	A. Moru	90.00	0.920	0.885	0.902	0.965	0.951
	12	88	Ayaz	-	0.880	0.917	0.898	0.965	0.969
SVM	A. Moru	Ayaz	-	-	-	-	-	-	-
	96	4	A. Moru	82.50	0.960	0.756	0.846	0.825	0.746
	31	69	Ayaz	-	0.690	0.945	0.798	0.825	0.807
A. Moru vs. Doruk
MLP	A. Moru	Doruk	-	-	-	-	-	-	-
	93	7	A. Moru	87.50	0.930	0.838	0.882	0.911	0.875
	18	82	Doruk	-	0.820	0.921	0.868	0.911	0.918
RF	A. Moru	Doruk	-	-	-	-	-	-	-
	89	11	A. Moru	87.00	0.890	0.856	0.873	0.909	0.879
	15	85	Doruk	-	0.850	0.885	0.867	0.909	0.922
SVM	A. Moru	Doruk	-	-	-	-	-	-	-
	79	21	A. Moru	77.00	0.790	0.760	0.775	0.770	0.705
	25	75	Doruk	-	0.750	0.781	0.765	0.770	0.711
A. Moru vs. Kristal
MLP	A. Moru	Kristal	-	-	-	-	-	-	-
	84	16	A. Moru	70.00	0.840	0.656	0.737	0.733	0.659
	44	56	Kristal	-	0.560	0.778	0.651	0.733	0.771
RF	A. Moru	Kristal	-	-	-	-	-	-	-
	79	21	A. Moru	73.50	0.790	0.712	0.749	0.799	0.769
	32	68	Kristal	-	0.680	0.764	0.720	0.799	0.810
SVM	A. Moru	Kristal	-	-	-	-	-	-	-
	83	17	A. Moru	71.50	0.830	0.675	0.744	0.715	0.645
	40	60	Kristal	-	0.600	0.779	0.678	0.715	0.668
A. Moru vs. Özveren
MLP	A. Moru	Özveren	-	-	-	-	-	-	-
	89	11	A. Moru	88.00	0.890	0.873	0.881	0.931	0.912
	11	87	Özveren	-	0.870	0.888	0.879	0.931	0.931
RF	A. Moru	Özveren	-	-	-	-	-	-	-
	91	9	A. Moru	90.50	0.910	0.901	0.905	0.966	0.965
	10	90	Özveren	-	0.900	0.909	0.905	0.966	0.968
SVM	A. Moru	Özveren	-	-	-	-	-	-	-
	90	10	A. Moru	88.00	0.900	0.865	0.882	0.880	0.829
	14	86	Özveren	-	0.860	0.896	0.878	0.880	0.840
A. Moru vs. Ürkmez
MLP	A. Moru	Ürkmez	-	-	-	-	-	-	-
	85	15	A. Moru	80.50	0.850	0.780	0.813	0.873	0.849
	24	76	Ürkmez	-	0.760	0.835	0.796	0.873	0.873
RF	A. Moru	Ürkmez	-	-	-	-	-	-	-
	83	17	A. Moru	80.50	0.830	0.790	0.810	0.878	0.853
	22	78	Ürkmez	-	0.780	0.821	0.800	0.878	0.852
SVM	A. Moru	Ürkmez	-	-	-	-	-	-	-
	67	33	A. Moru	75.00	0.670	0.798	0.728	0.750	0.699
	17	83	Ürkmez	-	0.830	0.716	0.769	0.750	0.679

Note: As blues darken, true values increase; as grays darken, false values increase.

). For Ankara Moru comparisons, SVM accuracy ranged from 71.50% (vs. Kristal) to 88.00% (vs. Özveren), with the lowest performance for Ayaz, Doruk, and Ürkmez. Low accuracy in some pairs (e.g., Alperen–Kristal) relates to overlap in thickness and geometric diameter distributions. Secondary criteria (sphericity, SI, aspect ratio) offered limited improvement, with low canonical distance between class centers. Intra-batch biological variation, segmentation tolerances, and measurement resolution were also influential.

For Ayaz vs. Doruk, RF achieved 89.50% accuracy, MLP achieved 87.50% accuracy, SVM achieved 81.00% accuracy. In RF, TPR was 0.860 Ayaz, 0.930 Doruk; precision 0.925, 0.869; F₁ 0.860, 0.899; PRC 0.954, 0.908; ROC 0.938 both (

Table 11. The confusion matrices and performance metrics for Ayaz varieties.

Classifiers	Predicted		Actual	Accuracy (%)	TPR	Precision	F1	ROC	PRC
Ayaz vs. Doruk
MLP	Ayaz	Doruk	-	-	-	-	-	-	-
	82	18	Ayaz	87.50	0.820	0.921	0.868	0.908	0.941
	7	93	Doruk	-	0.930	0.838	0.882	0.908	0.816
RF	Ayaz	Doruk	-	-	-	-	-	-	-
	86	14	Ayaz	89.50	0.860	0.925	0.860	0.938	0.954
	7	93	Doruk	-	0.930	0.869	0.899	0.938	0.908
SVM	Ayaz	Doruk	-	-	-	-	-	-	-
	69	31	Ayaz	81.00	0.690	0.908	0.784	0.810	0.781
	7	93	Doruk	-	0.930	0.750	0.830	0.810	0.732
Ayaz vs. Kristal
MLP	Ayaz	Kristal	-	-	-	-	-	-	-
	80	20	Ayaz	85.50	0.800	0.899	0.847	0.901	0.885
	9	91	Kristal	-	0.910	0.820	0.863	0.901	0.874
RF	Ayaz	Kristal	-	-	-	-	-	-	-
	88	12	Ayaz	88.50	0.880	0.889	0.884	0.932	0.944
	11	89	Kristal	-	0.890	0.881	0.886	0.932	0.908
SVM	Ayaz	Kristal	-	-	-	-	-	-	-
	70	30	Ayaz	79.50	0.700	0.864	0.773	0.795	0.755
	11	89	Kristal	-	0.890	0.748	0.813	0.795	0.721
Ayaz vs. Özveren
MLP	Ayaz	Özveren	-	-	-	-	-	-	-
	85	15	Ayaz	86.50	0.850	0.876	0.863	0.935	0.913
	12	88	Özveren	-	0.880	0.854	0.867	0.935	0.928
RF	Ayaz	Özveren	-	-	-	-	-	-	-
	87	13	Ayaz	89.50	0.870	0.916	0.892	0.944	0.946
	8	92	Özveren	-	0.920	0.876	0.898	0.944	0.911
SVM	Ayaz	Özveren	-	-	-	-	-	-	-
	61	39	Ayaz	73.00	0.610	0.803	0.693	0.730	0.685
	15	85	Özveren	-	0.850	0.685	0.759	0.730	0.658
Ayaz vs. Ürkmez
MLP	Ayaz	Ürkmez	-	-	-	-	-	-	-
	85	15	Ayaz	88.50	0.850	0.914	0.881	0.916	0.943
	8	92	Ürkmez	-	0.920	0.860	0.889	0.916	0.873
RF	Ayaz	Ürkmez	-	-	-	-	-	-	-
	84	16	Ayaz	88.00	0.840	0.913	0.875	0.930	0.932
	8	92	Ürkmez	-	0.920	0.852	0.885	0.930	0.912
SVM	Ayaz	Ürkmez	-	-	-	-	-	-	-
	63	37	Ayaz	75.00	0.630	0.829	0.716	0.750	0.707
	13	87	Ürkmez	-	0.870	0.702	0.777	0.750	0.675

Note: As blues darken, true values increase; as grays darken, false values increase.

). For Ayaz comparisons, SVM consistently underperformed relative to MLP and RF, with an accuracy of 73.00–81.00%.

Table 12. The confusion matrices and performance metrics for Doruk, Kristal, Özveren, and Ürkmez varieties.

Classifiers	Predicted		Actual	Accuracy (%)	TPR	Precision	F1	ROC	PRC
Doruk vs. Kristal
MLP	Doruk	Kristal	-	-	-	-	-	-	-
	82	18	Doruk	77.00	0.820	0.745	0.781	0.800	0.746
	28	72	Kristal	-	0.720	0.800	0.758	0.800	0.833
RF	Doruk	Kristal	-	-	-	-	-	-	-
	79	21	Doruk	77.00	0.790	0.760	0.775	0.846	0.833
	25	75	Kristal	-	0.750	0.781	0.765	0.846	0.847
SVM	Doruk	Kristal	-	-	-	-	-	-	-
	83	17	Doruk	78.50	0.830	0.761	0.794	0.785	0.717
	26	74	Kristal	-	0.740	0.813	0.775	0.785	0.732
Doruk vs. Özveren
MLP	Doruk	Özveren	-	-	-	-	-	-	-
	83	17	Doruk	76.00	0.830	0.728	0.776	0.831	0.796
	31	69	Özveren	-	0.690	0.802	0.742	0.831	0.823
RF	Doruk	Özveren	-	-	-	-	-	-	-
	76	24	Doruk	76.5	0.760	0.768	0.764	0.835	0.821
	23	77	Özveren	-	0.770	0.762	0.766	0.835	0.802
SVM	Doruk	Özveren	-	-	-	-	-	-	-
	75	25	Doruk	78.50	0.750	0.806	0.77	0.785	0.730
	18	82	Özveren	-	0.820	0.766	0.792	0.785	0.718
Doruk vs. Ürkmez
MLP	Doruk	Ürkmez	-	-	-	-	-	-	-
	64	36	Doruk	64.00	0.640	0.640	0.640	0.645	0.599
	36	64	Ürkmez	-	0.640	0.640	0.640	0.645	0.644
RF	Doruk	Ürkmez	-	-	-	-	-	-	-
	66	34	Doruk	68.00	0.660	0.688	0.673	0.706	0.647
	30	70	Ürkmez	-	0.700	0.673	0.686	0.706	0.690
SVM	Doruk	Ürkmez	-	-	-	-	-	-	-
	60	40	Doruk	64.00	0.600	0.630	0.619	0.660	0.583
	32	68	Ürkmez	-	0.680	0.650	0.661	0.660	0.581
Kristal vs. Özveren
MLP	Kristal	Özveren	-	-	-	-	-	-	-
	80	20	Kristal	76.50	0.800	0.748	0.773	0.841	0.827
	27	73	Özveren	-	0.730	0.785	0.756	0.841	0.848
RF	Kristal	Özveren	-	-	-	-	-	-	-
	83	17	Kristal	80.50	0.830	0.790	0.810	0.911	0.910
	22	78	Özveren	-	0.780	0.821	0.800	0.911	0.917
SVM	Kristal	Özveren	-	-	-	-	-	-	-
	75	25	Kristal	75.50	0.750	0.758	0.754	0.755	0.693
	24	76	Özveren	-	0.760	0.752	0.756	0.755	0.692
Kristal vs. Ürkmez
MLP	Kristal	Ürkmez	-	-	-	-	-	-	-
	66	34	Kristal	74.00	0.660	0.786	0.717	0.786	0.842
	18	82	Ürkmez	-	0.820	0.707	0.759	0.786	0.698
RF	Kristal	Ürkmez	-	-	-	-	-	-	-
	70	30	Kristal	77.00	0.700	0.814	0.753	0.851	0.862
	16	84	Ürkmez	-	0.840	0.737	0.785	0.851	0.832
SVM	Kristal	Ürkmez	-	-	-	-	-	-	-
	71	29	Kristal	78.00	0.710	0.826	0.763	0.780	0.731
	15	85	Ürkmez	-	0.850	0.746	0.794	0.780	0.709
Özveren vs. Ürkmez
MLP	Özveren	Ürkmez	-	-	-	-	-	-	-
	79	21	Özveren	81.50	0.790	0.832	0.810	0.859	0.875
	16	84	Ürkmez	-	0.840	0.800	0.820	0.859	0.818
RF	Özveren	Ürkmez	-	-	-	-	-	-	-
	77	23	Özveren	79.00	0.770	0.802	0.786	0.879	0.866
	19	81	Ürkmez	-	0.810	0.779	0.794	0.879	0.865
SVM	Özveren	Ürkmez	-	-	-	-	-	-	-
	78	22	Özveren	79.50	0.780	0.804	0.792	0.795	0.737
	19	81	Ürkmez	-	0.810	0.786	0.798	0.795	0.732

Note: As blues darken, true values increase; as grays darken, false values increase.

presents results for Doruk, Kristal, Özveren, and Ürkmez. For Özveren vs. Ürkmez, MLP achieved 81.50% accuracy, SVM achieved 79.50% accuracy, RF achieved 79.00% accuracy. In MLP, TPR was 0.790 Özveren, 0.840 Ürkmez; precision 0.832, 0.800; F₁ 0.810, 0.820; PRC 0.875, 0.818; ROC 0.859 both. SVM performed best for Doruk vs. Kristal and Doruk vs. Özveren (78.50%). Doruk vs. Ürkmez had low accuracy across models: RF—68.00%, and MLP and SVM—64.00%. For Kristal vs. Özveren, SVM had the lowest accuracy (75.50%); for Kristal vs. Ürkmez, SVM performed best (78.00% vs. MLP 74.00%, RF 77.00%).

ROC curves are shown in Figure 5. The best-classified pairs were Ayaz-Özveren, Ankara Moru-Özveren, and Ankara Moru-Ayaz. The worst were Alper-Doruk, Alperen-Kristal, and Doruk-Ürkmez.

Consistent with this study, ref. [53] classified 12 alfalfa varieties, reporting 91.53% (LDA) and 93.47% (SVM) accuracy. Ref. [54] reported 84.6–99.3% (RF), 87.4–99.3% (SVM), and 99.8–100% (LDA) for alfalfa seeds. Ref. [55] classified corn seeds using NB (93.37%), DT (94.95%), KNN (95.59%), LDA (95.97%), MLP (96.26%), and SVM (96.46%). Hyperspectral imaging and deep learning often achieve >95% accuracy [56,57,58], but require expensive equipment, large datasets, and computational resources. This study classified vetch seeds using physical characteristics, ML, and RGB imaging, achieving up to 90% accuracy, offering a practical, cost-effective solution for agricultural quality control. The scarcity of vetch classification studies underscores the originality and contribution of this work. Using RF, MLP, and SVM provides a balanced baseline for interpretability and capturing non-linear patterns; with limited data, more complex models risk overfitting for marginal gains.

4. Conclusions

Common vetch varieties exhibit consistent inter-variety differences based on controlled RGB imaging and morpho-physical descriptors. Ayaz was most distinguishable, achieving the highest pairwise discrimination metrics. Nutritionally, Alınoğlu and Alper were notable, while macro- and micronutrient concentrations varied meaningfully across varieties, indicating that phenotypic and compositional traits can jointly inform variety identification and agronomic decisions.

ML algorithms effectively classified common vetch seeds based on physical attributes. RF achieved the strongest pairwise result (90.50% for Ankara Moru vs. Özveren). Overall, supervised models provided reliable discrimination, supporting the feasibility of low-cost, feature-based classification under controlled imaging.

The physical attributes and ML pipeline can support quality assessment, planting system design (e.g., sizing, singulation), and breeding studies (e.g., non-destructive screening). Industrially, the workflow is compatible with routine seed inspection, flagging ambiguous lots for confirmatory analysis, aiding traceable, metric-based decision-making. The small seed size complicates high-throughput measurement. Future work should (i) explore hyperspectral and color-rich cues to enhance separability of visually similar varieties; (ii) improve performance with larger datasets, refined feature selection, and deep learning; and (iii) extend validation across multiple years and locations. These steps will enhance robustness and narrow the gap with end-to-end image-based classifiers for other species.