Hyperspectral Classification of Kiwiberry Ripeness for Postharvest Sorting Using PLS-DA and SVM: From Baseline Models to Meta-Inspired Stacked SVM

Abstract

The accurate and non-destructive assessment of fruit ripeness is essential for post-harvest sorting and quality management. This study evaluated a meta-inspired classification framework integrating partial least squares discriminant analysis (PLS-DA) with support vector machines (SVMs) trained on latent variables (sSVM) or on class probabilities (pSVM) derived from multiple PLS-DA components. Two kiwiberry varieties, ‘Geneva’ and ‘Weiki’, were analyzed using variety-specific and combined datasets. Performance was assessed in calibration and prediction using accuracy, F₀₅, Cohen’s kappa, precision, sensitivity, specificity, and likelihood ratios. Conventional PLS-DA provided reasonably good classification, but pSVM models, particularly those with an RBF kernel (pSVM_R), consistently outperformed other approaches and ensured higher stability across all datasets. Unlike sSVMs, which were prone to overfitting, pSVM_R models achieved the highest accuracy of 92.4–96.9%, Cohen’s kappa of 84.8–93.9%, and precision of 89.1–94.2%, clearly surpassing both score-based SVM and PLS-DA. Contrasting tendencies were observed between cultivars: ‘Geneva’ models improved during prediction, while ‘Weiki’ models declined, especially in specificity. Combined datasets provided greater stability but slightly reduced peak performance than single-variety models. These findings highlight the value of probability-enriched stacking models for non-invasive ripeness discrimination, suggesting that adaptive or hybrid strategies may further enhance generalization across diverse cultivars.

1. Introduction

Actinidia arguta (Siebold et Zucc.) Planch. ex Miq., commonly referred to as kiwiberry or minikiwi, is a climacteric fruit valued for its exceptional sensory and nutritional properties [1,2,3,4]. It is particularly rich in vitamin C, polyphenols, carotenoids, and other bioactive compounds that contribute to its classification as a “superfruit” [5,6]. In Poland, commercial cultivation has developed over the last 25 years, with varieties ‘Weiki’, ‘Geneva’, and ‘Bingo’ gaining the greatest commercial importance [4,7].

Despite its potential, kiwiberry remains a commercially challenging fruit. Its short storage life, high susceptibility to uneven ripening, and seasonal availability limit market expansion and profitability [8,9]. In the Polish climate, fruits are typically harvested at the stage of harvest maturity, when they are still firm and sour, and then undergo rapid and heterogeneous ripening after harvest [10]. This variability makes post-harvest sorting according to ripeness a key step for extending market supply, reducing losses, and ensuring uniform fruit quality.

Traditional assessment of soluble solids content (SSC), the main ripeness indicator in Actinidia spp., relies on destructive refractometric measurements [11,12,13]. Therefore, non-destructive technologies are required for practical implementation on sorting lines. This requirement can be successfully met by solutions based on vision techniques, particularly hyperspectral imaging. This approach has proven to be an effective quality control tool, integrating spectral and spatial information for the rapid and non-invasive assessment of examined objects. Recent studies have demonstrated the effectiveness of HSI in predicting the quality of Actinidia ssp., including their health [14,15,16], texture [17,18], maturity [19,20,21,22], soluble solid content and acidity [23,24,25], or storage condition [26,27].

For spectral data modeling, partial least squares (PLS) and its variant PLS-DA are widely used, providing dimensionality reduction and interpretable class separation [20,28]. These methods, increasingly applied to hyperspectral images with high collinearity and redundancy [29]. They facilitate spectral interpretation and create a solid basis for integrating advanced machine learning techniques.

PLS-DA extends PLS regression to classification tasks, offering dimensionality reduction and class separation in high-dimensional and collinear datasets. Unlike principal component regression (PCR), PLS explicitly incorporates information from predictors and responses, improving class discrimination. The optimal number of latent components is typically chosen by cross-validation [30,31,32,33].

On the other hand, the high dimensionality and collinearity of spectral data encourage the use of advanced machine learning algorithms and automated variable selection methods such as CARS, BoSS, GA, VIP, or SPA [19,20,21,24,34,35,36]. Such approaches reduce redundancy and capture nonlinear relationships beyond the reach of traditional models. Support vector machines (SVMs) are particularly effective in this context. This method offers strong generalization in high-dimensional feature spaces and robustness to nonlinear class boundaries through kernel functions such as linear or radial basis function (RBF) [37,38,39]. Their ability to define complex decision margins makes them well-suited for hyperspectral applications [25,40,41]. Moreover, stacking or meta-learning strategies that combine SVM with feature extraction may further enhance performance, providing a promising direction for ripeness-based kiwiberry sorting.

While hyperspectral imaging has been widely explored for fruit quality evaluation, the choice and design of classification strategies remain critical for practical implementation. Previous studies often rely on single classifiers or raw spectral inputs, overlooking the potential of combining feature extraction with robust machine learning models. This study addresses this gap by systematically comparing PLS-DA, SVM with different kernels, and hybrid SVM-PLS approaches, providing insights into their relative performance and suitability for automated kiwiberry sorting. The novelty of this work lies in its integrated evaluation of multiple strategies, applied for the first time to Actinidia arguta, a fruit species still rarely studied in post-harvest management. Notably, PLS-based methods are particularly well-suited for high-dimensional data typical of hyperspectral analyses, where the number of predictors substantially exceeds the number of observations. In this context, hyperspectral imaging does not precisely quantify individual chemical constituents, but captures spectral patterns that enable object classification, grouping, and similarity assessment in post-harvest sorting processes.

Therefore, this study aimed to evaluate a meta-inspired classification framework combining multiple modeling strategies to classify kiwiberry fruits according to their ripeness. Specifically, we compared classical PLS-DA, SVM models trained on latent variables extracted from PLS-DA (sSVM), and SVM models leveraging class probabilities (pSVM) derived from multiple PLS-DA models with varying numbers of components, each validated by random 10-fold cross-validation. This approach enabled a systematic assessment of the potential of integrating hyperspectral imaging with advanced classification techniques for optimized post-harvest fruit sorting. pSVMs may be considered stuck models, since they rely directly on the original feature space and are constrained by the geometry of the raw predictors, while sSVMs operate in a transformed latent space derived from PLS-DA scores, which changes the numerical nature of the optimization problem and alters the generalization behavior. The study underscores the importance of model architecture and feature extraction in advancing non-invasive post-harvest fruit sorting systems.

2. Materials and Methods

2.1. Collection and Preparation of Kiwiberry Samples

The fruits of A. arguta represent two varieties, ‘Weiki’ and ‘Geneva’, which were used to develop classifiers. Fruits originated from a commercial plantation in Bodzew, along Grójec, Mazowieckie voivodship, Poland (51°47′50″ N, 20°48′43″ E; USDA hardiness zone 6b). Harvests were carried out in 2021 and 2022 in early September. Fruits were collected randomly at the harvest maturity stage, with an SSC of 6.5–7% [8,9]. Fruits were kept at 1 °C and 90% relative humidity and monitored for up to several days to capture ripeness variations, beginning on the day of harvest. From this batch 5938 samples were chosen randomly and employed for model development, comprising 3810 ‘Weiki’ fruits and 2128 ‘Geneva’ fruits, the latter of which were collected exclusively in 2022. The experimental protocol, including fruit storage, initial sample preparation, acquisition and preliminary processing of hyperspectral images, and measurement of fruit extract, was identical in both study years and similar to those described by Janaszek-Mańkowska and Ratajski [42].

2.2. Hyperspectral Imaging Setup

The hyperspectral imaging setup (HIS) employed a push-broom line scanning technique and consisted of two hyperspectral cameras, FX10 (CMOS detector) and FX17 (cooled InGaAs detector) from SPECIM Ltd. (Oulu, Finland), a 250 W halogen lamp, and an external PC. The FX10 camera operated in the 400–1000 nm range (visible near-infrared, VNIR) with 448 spectral bands, while the FX17 covered 900–1700 nm (short-wave infrared, SWIR) with 224 bands. To reduce external light interference, all HIS components except the PC were enclosed in a vision chamber, which was the core of the prototype kiwiberry sorting device and also consisted of forced air exhaust system and four transducers (F&F, model MB-AHT-1) connected to a PLC (Siemens, S7-1214C) to monitor and prevent overheating from the halogen lamp (Figure 1). Both cameras were placed above a conveyor belt, moving randomly positioned fruits at 40 mm/s. The optical axis of FX10 was perpendicular to the conveyor, while FX17 was tilted ~12° to capture the same sample area. The halogen lamp was positioned at a 45° angle. Both cameras acquired 12-bit images with mean spectral resolutions of 1.32 nm (FX10) and 3.57 nm (FX17).

2.3. Processing Workflow for Hyperspectral Reflectance Data

The OpenCV custom application supported the automatic acquisition of hyperspectral cubes of each fruit and further image pre-processing. In order to correct for spectral light variations, sensor sensitivity, and dark current, radiometric calibration was performed using dark and white reference images (0% and 99% reflectance, respectively) as proposed by Xiong et al. [43]:

$$I_C={\displaystyle \frac{I_O-I_D}{I_W-I_D}},$$

(1)

where I_C and I_O are the calibrated and original images, I_D is the dark reference (camera shutter closed, light off), and I_W is the white reference (PTFE tile imaged before each cycle under the same light conditions as I_O).

Each hypercube was sliced into 16-bit 2D images, representing reflectance in individual spectral bands. Although hyperspectral images initially covered the full ranges recorded by both cameras, bands with low signal-to-noise ratio (SNR), poor discriminative capacity, observed mainly at the ends of the SWIR range, or redundancy due to spectral overlap were discarded after preliminary inspection. Only bands within ~490–909 nm and ~977–1379 nm were retained for further analyses. In the sequence of selected 2D images corresponding to each fruit, a region of interest (ROI) was defined to include only the fruit, and subsequent image processing steps were applied exclusively within this ROI. At the same time, the remaining background was excluded from analysis. ROI location, which covered the entire fruit surface, was specified separately for VNIR and SWIR image sequences using the IsoData binary threshold [44] on hypercube slices with the highest contrast between background and sample reflectance. For each waveband λ, the mean reflectance within the ROI was calculated according to Formula (2):

$${\overline{R}}_{\lambda }={\displaystyle \frac{{\displaystyle {\sum }_{p=1}^n{R}_p}}{n}},$$

(2)

where ${\overline{R}}_{\lambda }$ is the mean reflectance within the ROI for a given wavelength λ, R_p is the reflectance of the n-th pixel, and n is the total number of pixels.

Although calibration reduces many artifacts, spectral data can still be affected by baseline drift, nonlinearity, or scattering effects, which introduce noise unrelated to the actual signal. Several pre-processing techniques were applied separately to the VNIR and SWIR ranges to address these disadvantages. These included Multiplicative Scatter Correction (MSC), Standard Normal Variate (SNV), Savitzky–Golay (SG) smoothing, mean centering (MC), and spectral derivatives of the first (D1) and the second order (D2). For SG filtering, a frame size of 15 was used, corresponding to bandwidths of 18.56 nm (VNIR) and 48.52 nm (SWIR). Derivatives, calculated after SG smoothing with low-order polynomial fitting, were used to reduce baseline and slope effects while improving the resolution of overlapping features. MSC compensated for scatter by aligning spectra with a reference [45,46], whereas SNV normalized each spectrum relative to its mean and standard deviation [47]. MC emphasized spectral variation by shifting each spectrum relative to the overall average. SG filtering smoothed the signal through polynomial convolution [48]. Derivatives further corrected baseline and slope shifts, though smoothing was essential to limit noise amplification [46,49]. All these corrections were implemented in R version 4.5.1 using the mdatools package [50,51].

2.4. Determination of SSC as a Ripeness Index

SSC served as the reference index of kiwiberry ripeness. Following hyperspectral image acquisition, fruit juice was extracted and analyzed with a digital refractometer (ATAGO, PAL-1, Tokyo, Japan), providing ±0.1% accuracy within the 0–53° Brix range. Measured SSC values spanned from 4.4% to 17.1% (mean 8.02 ± 2.5%) for ‘Weiki’ fruits and from 4.3% to 13.6% (mean 7.82 ± 1.7%) for ‘Geneva’ fruits, thus covering a broad maturity spectrum, including levels exceeding commercial standards.

2.5. Data Preparation

Data obtained from hyperspectral images were organized into datasets corresponding to the kiwiberry variety. Class of fruit ripeness was defined based on SSC so that samples with SSC ≤ 7% were assigned to the unripe class (A—positive class), whereas the remaining fruits were classified as ripe (B—negative class). The class label was treated as the dependent variable, while the predictor set comprised 425 spectral variables, including 309 wavebands from the VNIR range and 116 from the SWIR range. Three datasets were prepared: (i) ‘Weiki’ variety data (W) from two harvest years, (ii) ‘Geneva’ variety data (G), and (iii) a combined dataset including both ‘Geneva’ and ‘Weiki’ varieties (WG). Each dataset was split into training and test subsets in a 4:1 ratio (80% training, 20% test). The training part was used to calibrate classification models, while the independent test set served as external validation of model performance. Stratified random sampling was performed with the createDataPartition function from the caret package [52] of R version 4.5.1, ensuring balanced class proportions in training and test subsets. Model calibration relied on training datasets (with internal cross-validation applied during training), whereas test data were used for prediction. In this context, ‘calibration’ refers to model optimization and evaluation based on cross-validated training results, while ‘prediction’ denotes the assessment of the final model on the independent test set. Detailed information on the datasets’ structure is summarized in

Table 1. Dataset sizes and class proportions after train/test split.

Dataset	Training		Test
	Class A	Class B	Class A	Class B
‘Weiki’ (W)	1526	1524	380	380
‘Geneva’ (G)	852	852	212	212
‘Weiki’ + ‘Geneva’ (WG)	2378	2376	592	592

.

Following the data partition, spectral preprocessing (MSC, SNV, SG, MC, D1, D2) was applied, yielding 18 training—test set pairs for subsequent analyses. Preprocessing was performed separately for training and test subsets to avoid data leakage.

2.6. PLS-DA and SVM Classification

Classification models were implemented in R version 4.5.1 using caret package, which applied the pls package for PLS-DA [53] and the kernlab library for SVM [54,55]. Model calibration was consistently performed using random 10-fold cross-validation on the training set to prevent overfitting. The modeling procedure consisted of two main stages. First, PLS-DA was employed to obtain independent classifiers, which also served as a Level-0 models to provide a set of latent variables (scores) as well as winning class probabilities for subsequent Level-1 SVM models developed in the second stage. The idea of Level-0 and Level-1 generalizers is given in [56]. The optimal number of PLS components in each model was determined based on the kappa statistic [57], defined as follows:

$$\kappa \left(\%\right)=100\times {\displaystyle \frac{p_o-p_e}{1-p_e}},$$

(3)

where p_o, which denotes the observed agreement (proportion of correctly classified samples), is given as:

$$p_o={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^{k}T{P}_i},$$

(4)

and p_e denotes the expected agreement by chance, calculated from the marginal distributions of the confusion matrix as follows:

$$p_e={\displaystyle \sum _{i=1}^k\frac{\left(T{P}_i+F{N}_i\right)\cdot \left(T{P}_i+F{P}_i\right)}{n^2}},$$

(5)

where TP_i, FP_i, and FN_i denote the elements of the confusion matrix shown in

Table 2. Confusion matrix for binary classification.

	Predicted Positive (A)	Predicted Negative (B)	Overall Observed
Observed positive (A)	TP–true positive	FN–false negative	OP = TP + FN
Observed negative (B)	FP–false positive	TN–true negative	ON = FP + TN
Overall predicted	PP = TP + FP	PN = FN + TN	n = TP + FN + FP + TN

for class i, and n is the total number of observations.

Table 2 presents a confusion matrix for binary classification with positive class A (unripe fruits) and negative class B (ripe fruits), where TP and TN denote the number of correctly classified samples, whereas FP and FN indicate the number of misclassifications [58].

The optimal number of PLS components typically amounted to 18 or 20, indicating relatively high model complexity. Thus, for subsequent analyses with SVM classifiers based on scores (sSVM), the maximum number of latent variables extracted from the training set was capped at 20, and test set scores were calculated using the corresponding loading weights, ensuring that both datasets were represented in the same latent space. Thereby, the dimensionality of the input space was reduced from 425 original spectral variables to 20 scaled and transformed latent predictors. Moreover, considering the most common optimal number of latent variables (20), an additional 18 alternative and independent PLS-DA models were fitted, with the number of components increased stepwise by one, starting from 3. For each PLS-DA model, class probabilities for the winning class (class A) were computed separately for the training and test sets. The probabilities derived from the training set were subsequently used as predictors in SVM models (pSVM), whereas those obtained from the test set were retained exclusively for their independent validation. Therefore, pSVMs may be considered stacked models similar to description by Ting and Witten [56], as they operated on meta-level inputs rather than raw or transformed predictors. By contrast, sSVMs relied directly on latent variables, linear combinations of the original spectral features.

In the second stage, SVM classifiers with linear (SVM_L) and radial (RBF) kernels (SVM_R) were developed, but the latter were used only with pSVM variants. This decision was based on the linear nature of the latent variables, which are constructed to maximize class separation. Introducing a nonlinear kernel would likely add complexity without improving separability, as linear boundaries are already sufficient in the latent space [59]. While the behavior of RBF kernels in this context is of academic interest, their inclusion was deemed unnecessary for practical model development. For SVM_L models, the regularization parameter C was optimized, whereas for SVM_R models, both C and γ (the kernel width parameter) were tuned. Parameter optimization was done by exhaustive grid search to maximize the κ statistic in a random 10-fold cross-validation. Based on the provided training data, candidate grids were generated automatically using 20 values for each parameter. The selection of components in sSVM_L models was constrained to sequential combinations (e.g., 1–3 or 1–5), while irregular patterns (1, 4, 8) were excluded. Our decision was motivated by the inherent structure of PLS components, where successive latent variables explain progressively smaller portions of the shared variance, and skipping intermediate components would lead to less coherent models.

2.7. Evaluation of Classification Performance

Standard classification metrics, such as accuracy (ACC), sensitivity (TPR), specificity (TNR), precision (PPV), the F₀₅ score, and likelihood ratios (LR₊ and LR₋) were applied to evaluate model classification capacity [60,61]. While ACC provides a general indication of classifier performance, it may miss systematic class imbalances. Therefore, a broader set of complementary measures was employed better to capture the trade-offs between false positives and false negatives. In the context of fruit maturity assessment in our study, misclassifying ripe fruits as unripe is generally considered more detrimental than the opposite case. This is because the presence of ripe fruits accelerates postharvest ripening of the entire batch, potentially compromising storage stability and marketability. Although explicit cost-sensitive learning was not introduced, this asymmetry in error perception motivated the use of metrics that emphasize precision and penalize false positives more strongly, such as the F₀₅ score. Similarly, sensitivity and specificity were reported to provide a balanced view of correct recognition in both classes, and kappa was included as a robust, prevalence-independent summary statistic. Finally, likelihood ratios were calculated, as they directly relate classifier output to practical decision-making in diagnostic-style settings. Based on confusion matrix from Table 1, evaluation metrics for classifiers were calculated as follows:

$$ACC\hspace{0.33em}\left(\%\right)=100\times {\displaystyle \frac{TP+TN}{n}},$$

(6)

$$TPR\hspace{0.33em}\left(\%\right)=100\times {\displaystyle \frac{TP}{OP}},$$

(7)

$$TNR\hspace{0.33em}\left(\%\right)=100\times {\displaystyle \frac{TN}{ON}},$$

(8)

$$PPV\hspace{0.33em}\left(\%\right)=100\times {\displaystyle \frac{TP}{PP}},$$

(9)

$$F_{05}={\displaystyle \frac{\left(1+{\beta }^2\right)\times PPV\times TPR}{{\beta }^2\times PPV+TPR}},\hspace{0.33em}\beta =0.5,$$

(10)

$$L{R}_{+}={\displaystyle \frac{TPR}{1-TNR}},$$

(11)

$$L{R}_{-}={\displaystyle \frac{1-TPR}{TNR}}$$

(12)

where β controls the relative weight assigned to precision and sensitivity, which are indirectly linked to misclassification costs. If recognition of the winning class and its purity are equally important (β = 1), the measure is denoted as F₁. A higher emphasis on sensitivity (β = 2) results in F₂, whereas a stronger weight on precision (β = 0.5) yields F₀₅.

For additional statistical validation of the best-performing models, 95% bootstrap confidence intervals for accuracy, κ, PPV, TPR, and TNR, as well as McNemar’s tests for paired model comparisons, were computed and are reported in the Appendix A (

Table A3. Confidence intervals ⁽¹⁾ of classification performance metrics for the best models.

Model	Confidence Interval	Classification Performance Metrics
		ACC	κ	F05	PPV	TPR *	TNR
G_PLS-DA (D2)	Lower limit	0.8939	0.7886	0.8600	0.8354	0.9466	0.8169
	Upper limit	0.9434	0.8902	0.9278	0.9163	0.9912	0.9072
W_PLS-DA (D2)	Lower limit	0.8961	0.7951	0.8743	0.8564	0.9303	0.8443
	Upper limit	0.9342	0.8706	0.9249	0.9163	0.9725	0.9093
WG_PLS-DA (D2)	Lower limit	0.8695	0.7401	0.8479	0.8297	0.9058	0.8174
	Upper limit	0.9054	0.8112	0.8943	0.8843	0.9490	0.8775
G_pSVM_R (SG)	Lower limit	0.9481	0.9003	0.9230	0.9056	NA	0.8994
	Upper limit	0.9811	0.9668	0.9734	0.9669	NA	0.9647
W_pSVM_R (D2)	Lower limit	0.9092	0.8226	0.8852	0.8655	0.9467	0.8557
	Upper limit	0.9447	0.8937	0.9331	0.9243	0.9830	0.9175
WG_pSVM_R (D1)	Lower limit	0.9071	0.8162	0.8835	0.8650	0.9495	0.8550
	Upper limit	0.9375	0.8767	0.9239	0.9137	0.9786	0.9060

⁽¹⁾ To estimate confidence intervals for classification metrics, we applied the bootstrap bias-corrected and accelerated (BCa) method using the boot package of R version 4.5.1 [64]. BCa accounts for both bias and skewness in the resampling distribution, providing more reliable interval estimates than the simple percentile approach. Abbreviations: Cultivar: W–Weiki, G–Geneva, WG–Weiki + Geneva; Correction: SG-Savitzky–Golay smoothing, D1–the first derivative, D2–the second derivative; Model: PLS-DA–partial least square discriminant analysis, SVM–support vector machine, pSVM_R–probability-based SVM with RBF kernel; Classification performance metrics: ACC–accuracy, κ– Cohen’s kappa, F05–F0.5-score, PPV–precision, TPR–true positive rate (sensitivity, recall), TNR–true negative rate; * NA: confidence interval not available (no false negatives observed in the test set).

and

Table A4. McNemar’s test ⁽¹⁾ results.

Comparable Models	Test Statistics
	Chi-Square	p-Value
G_PLS-DA (D2) vs. G_pSVM_R (SG)	12.0333	0.00052
W_PLS-DA (D2) vs. W_pSVM_R (D2)	1.8846	0.16981
WG_PLS-DA (D2) vs. WG_pSVM_R (D1)	20.2532	0.00001

⁽¹⁾ McNemar’s tests were computed using mcnemar.test function available in the stats library of R version 4.5.1 [51]. Abbreviations: Cultivar: W–Weiki, G–Geneva, WG–Weiki + Geneva; Correction: SG-Savitzky–Golay smoothing, D1–the first derivative, D2–the second derivative; Model: PLS-DA–partial least square discriminant analysis, SVM–support vector machine, pSVM_R-probability-based SVM with RBF kernel.

, respectively).

We measured the time required to generate predictions on the test samples to evaluate whether the classification performance improvement justifies the meta-models’ additional complexity. Median times from 100 repetitions are reported for each stage and per sample in

Table A5. Prediction times ⁽¹⁾ of PLS-DA and PLS-DA–pSVM_R meta-models.

Model	Median [s]		Median per 1 Obs. [s]		nsam	neval
	Level-0	Level-1	Level-0	Level-1
G_PLS-DA (D2)	0.0701149		0.0001654		424	100
G_pSVM_R (SG)	0.0682671	0.0311898	0.0001610	0.0000736	424	100
W_PLS-DA (D2)	0.0857587		0.0001128		760	100
W_pSVM_R (D2)	0.0857587	0.0554228	0.0001128	0.0000729	760	100
WG_PLS-DA (D2)	0.1127689		0.0000952		1184	100
WG_pSVM_R (D1)	0.1113516	0.1150936	0.0000940	0.0000972	1184	100

⁽¹⁾ Inference time measurements were performed with the microbenchmark package (100 repetitions) using R version 4.5.1 [65]. Times for the PLS-DA and SVM stages were measured separately, because the evaluation was executed with two independent scripts (research-grade, not specifically optimized for runtime). Median times are reported for each stage (level-0, level-1) and per sample. Hardware: Lenovo, Morrisville, NC, USA, Intel i5-10400F CPU 2.90 GHz, 16 GB RAM, no GPU. Abbreviations: n_sam–number of samples in test dataset; n_ewal–number of repetitions; Cultivar: W–Weiki, G–Geneva, WG–Weiki + Geneva; Correction: SG-Savitzky–Golay smoothing, D1–the first derivative, D2–the second derivative; Model: PLS-DA–partial least square discriminant analysis, SVM–support vector machine, pSVM_R-probability-based SVM with RBF kernel.

(Appendix A).

All computational analyses were performed on a standard personal computer (Lenovo, Morrisville, NC, USA, Intel i5-10400F CPU 2.90 GHz, 16 GB RAM, no GPU).

3. Results

Classification outcomes presented in this section were obtained for kiwiberry samples (5938 fruits in total: 3810 ‘Weiki’ and 2128 ‘Geneva’) using PLS-DA and SVM models. Analyses were performed separately for each dataset, and multiple preprocessing variants were considered to assess their impact on model performance. Models developed with PLS-DA were treated as the basis, whereas SVMs were assessed for potential improvements in classification performance. The following subsections present calibration and prediction results for the best-performing models in terms of classification metrics, organized according to the individual datasets. This approach allows a clear comparison of the predictive capabilities of the different modeling strategies and highlights the impact of using latent variables and model stacking on classification performance.

Calibration (cross-validated model tuning) and independent test set prediction results for all combinations of model and preprocessing variant are provided in

Table A1. Calibration and prediction results for PLS-DA and SVM classifiers.

Cultivar	Correction	Model	Classification Performance Metrics
			ACC	κ	F05	PPV	TPR	TNR	LR+	LR−	nC	C	γ
Calibration (10-fold cross-validation)
G	D1	PLS-DA	0.8803	0.7606	0.8633	0.8491	0.9249	0.8357	5.6286	0.0899	20
G	D2	PLS-DA	0.8967	0.7934	0.8738	0.8558	0.9542	0.8392	5.9343	0.0545	20
G	MC	PLS-DA	0.8785	0.7570	0.8614	0.8471	0.9237	0.8333	5.5423	0.0915	20
G	MSC	PLS-DA	0.8803	0.7606	0.8629	0.8484	0.9261	0.8345	5.5957	0.0886	20
G	SG	PLS-DA	0.8797	0.7594	0.8609	0.8453	0.9296	0.8298	5.4621	0.0849	17
G	SNV	PLS-DA	0.8838	0.7676	0.8653	0.8501	0.9319	0.8357	5.6714	0.0815	20
W	D1	PLS-DA	0.9433	0.8866	0.9303	0.9210	0.9699	0.9167	11.6535	0.0329	20
W	D2	PLS-DA	0.9439	0.8879	0.9344	0.9274	0.9633	0.9245	12.7826	0.0397	18
W	MC	PLS-DA	0.9397	0.8793	0.9292	0.9215	0.9613	0.9180	11.7360	0.0422	19
W	MSC	PLS-DA	0.9413	0.8826	0.9303	0.9223	0.9640	0.9186	11.8629	0.0393	19
W	SG	PLS-DA	0.9400	0.8800	0.9290	0.9210	0.9626	0.9173	11.6587	0.0408	19
W	SNV	PLS-DA	0.9413	0.8826	0.9294	0.9207	0.9659	0.9167	11.6063	0.0372	19
WG	D1	PLS-DA	0.8973	0.7947	0.8851	0.8753	0.9268	0.8678	7.0191	0.0844	20
WG	D2	PLS-DA	0.9030	0.8061	0.8908	0.8811	0.9319	0.8742	7.4114	0.0780	20
WG	MC	PLS-DA	0.8883	0.7766	0.8797	0.8725	0.9096	0.8670	6.8449	0.1044	20
WG	MSC	PLS-DA	0.9001	0.8002	0.8877	0.8777	0.9298	0.8704	7.1786	0.0808	20
WG	SG	PLS-DA	0.8902	0.7804	0.8818	0.8748	0.9108	0.8695	6.9871	0.1026	20
WG	SNV	PLS-DA	0.8986	0.7972	0.8858	0.8756	0.9294	0.8678	7.0382	0.0815	20
G	D1	sSVM_L	0.8838	0.7676	0.8740	0.8658	0.9085	0.8592	6.4500	0.1066	19	5
G	D2	sSVM_L	0.9237	0.8474	0.9053	0.8915	0.9648	0.8826	8.2200	0.0399	20	9
G	MC	sSVM_L	0.8891	0.7782	0.8766	0.8663	0.9202	0.8580	6.4793	0.0930	20	7
G	MSC	sSVM_L	0.8920	0.7840	0.8796	0.8695	0.9225	0.8615	6.6610	0.0899	20	3
G	SG	sSVM_L	0.8867	0.7735	0.8730	0.8617	0.9214	0.8521	6.2302	0.0923	20	1
G	SNV	sSVM_L	0.8932	0.7864	0.8803	0.8698	0.9249	0.8615	6.6780	0.0872	20	8
W	D1	sSVM_L	0.9443	0.8885	0.9377	0.9330	0.9574	0.9311	13.9143	0.0458	20	1
W	D2	sSVM_L	0.9443	0.8885	0.9368	0.9313	0.9594	0.9291	13.5556	0.0438	18	4
W	MC	sSVM_L	0.9439	0.8879	0.9376	0.9329	0.9567	0.9311	13.9048	0.0465	20	2
W	MSC	sSVM_L	0.9430	0.8859	0.9345	0.9284	0.9600	0.9259	12.9646	0.0432	20	1
W	SG	sSVM_L	0.9433	0.8866	0.9369	0.9323	0.9561	0.9304	13.7642	0.0472	20	10
W	SNV	sSVM_L	0.9436	0.8872	0.9349	0.9285	0.9613	0.9259	12.9823	0.0418	20	1
WG	D1	sSVM_L	0.9009	0.8018	0.8935	0.8875	0.9182	0.8836	7.8916	0.0926	18	2.5
WG	D2	sSVM_L	0.9041	0.8082	0.8950	0.8878	0.9251	0.8830	7.9137	0.0848	19	8
WG	MC	sSVM_L	0.8927	0.7854	0.8856	0.8797	0.9100	0.8754	7.3108	0.1029	20	3
WG	MSC	sSVM_L	0.9000	0.7999	0.8926	0.8866	0.9173	0.8826	7.8207	0.0938	20	10
WG	SG	sSVM_L	0.8908	0.7817	0.8840	0.8783	0.9075	0.8742	7.2174	0.1059	20	2
WG	SNV	sSVM_L	0.8988	0.7976	0.8929	0.8881	0.9126	0.8850	7.9390	0.0988	20	8
G	D1	sSVM_R	0.9818	0.9636	0.9828	0.9835	0.9800	0.9836	59.6429	0.0203	18	2	0.3
G	D2	sSVM_R	0.9777	0.9554	0.9764	0.9755	0.9800	0.9754	39.7619	0.0205	19	2	0.2
G	MC	sSVM_R	0.9842	0.9683	0.9818	0.9802	0.9883	0.9800	49.5294	0.0120	20	2	0.2
G	MSC	sSVM_R	0.9718	0.9437	0.9738	0.9752	0.9683	0.9754	39.2857	0.0325	16	2	0.4
G	SG	sSVM_R	0.9836	0.9671	0.9816	0.9802	0.9871	0.9800	49.4706	0.0132	20	2	0.2
G	SNV	sSVM_R	0.9736	0.9472	0.9733	0.9730	0.9742	0.9730	36.0870	0.0265	18	2	0.3
W	D1	sSVM_R	0.9643	0.9285	0.9623	0.9610	0.9679	0.9606	24.6167	0.0335	20	6	0.2
W	D2	sSVM_R	0.9570	0.9141	0.9636	0.9684	0.9450	0.9692	30.6809	0.0569	20	3	0.4
W	MC	sSVM_R	0.9252	0.8505	0.9636	0.9954	0.8545	0.9961	217.3333	0.1462	20	2	0.9
W	MSC	sSVM_R	0.9587	0.9174	0.9640	0.9679	0.9489	0.9685	30.1667	0.0528	18	3	0.6
W	SG	sSVM_R	0.9541	0.9082	0.9678	0.9779	0.9292	0.9790	44.3125	0.0724	20	2	0.5
W	SNV	sSVM_R	0.9597	0.9193	0.9613	0.9624	0.9567	0.9626	25.6140	0.0450	20	2	0.3
WG	D1	sSVM_R	0.9636	0.9272	0.9608	0.9588	0.9689	0.9583	23.2727	0.0325	20	4	0.2
WG	D2	sSVM_R	0.9531	0.9062	0.9575	0.9607	0.9449	0.9613	24.4239	0.0574	19	6	0.4
WG	MC	sSVM_R	0.9626	0.9251	0.9653	0.9673	0.9575	0.9676	29.5714	0.0439	20	4	0.3
WG	MSC	sSVM_R	0.9575	0.9150	0.9607	0.9630	0.9516	0.9634	26.0115	0.0502	17	4	0.4
WG	SG	sSVM_R	0.9609	0.9218	0.9639	0.9660	0.9554	0.9663	28.4000	0.0462	20	3	0.3
WG	SNV	sSVM_R	0.9575	0.9150	0.9584	0.9591	0.9558	0.9592	23.4330	0.0461	18	3	0.3
G	D1	pSVM_L	0.8768	0.7535	0.8646	0.8543	0.9085	0.8451	5.8636	0.1083		1
G	D2	pSVM_L	0.8950	0.7899	0.8817	0.8710	0.9272	0.8627	6.7521	0.0844		1
G	MC	pSVM_L	0.8762	0.7523	0.8660	0.8573	0.9026	0.8498	6.0078	0.1146		1
G	MSC	pSVM_L	0.8809	0.7617	0.8687	0.8586	0.9120	0.8498	6.0703	0.1036		1
G	SG	pSVM_L	0.8797	0.7594	0.8663	0.8551	0.9143	0.8451	5.9015	0.1014		1
G	SNV	pSVM_L	0.8826	0.7653	0.8689	0.8575	0.9178	0.8474	6.0154	0.0970		1
W	D1	pSVM_L	0.9433	0.8866	0.9350	0.9290	0.9600	0.9265	13.0804	0.0432		1
W	D2	pSVM_L	0.9433	0.8866	0.9369	0.9323	0.9561	0.9304	13.7642	0.0472		1
W	MC	pSVM_L	0.9403	0.8807	0.9345	0.9302	0.9522	0.9285	13.3303	0.0516		1
W	MSC	pSVM_L	0.9420	0.8839	0.9331	0.9266	0.9600	0.9239	12.6293	0.0433		1
W	SG	pSVM_L	0.9430	0.8859	0.9349	0.9289	0.9594	0.9265	13.0714	0.0439		1
W	SNV	pSVM_L	0.9420	0.8839	0.9341	0.9283	0.9581	0.9259	12.9381	0.0454		1
WG	D1	pSVM_L	0.8969	0.7939	0.8893	0.8831	0.9151	0.8788	7.5556	0.0967		1
WG	D2	pSVM_L	0.9018	0.8035	0.8932	0.8864	0.9218	0.8817	7.8007	0.0888		1
WG	MC	pSVM_L	0.8849	0.7699	0.8793	0.8744	0.8991	0.8708	6.9642	0.1160		1
WG	MSC	pSVM_L	0.8961	0.7922	0.8894	0.8839	0.9121	0.8801	7.6105	0.1000		1
WG	SG	pSVM_L	0.8870	0.7741	0.8815	0.8768	0.9008	0.8733	7.1163	0.1137		1
WG	SNV	pSVM_L	0.8931	0.7863	0.8879	0.8835	0.9058	0.8805	7.5845	0.1071		1
G	D1	pSVM_R	0.9108	0.8216	0.8894	0.8731	0.9613	0.8603	6.8824	0.0450		16	0.1792
G	D2	pSVM_R	0.9155	0.8310	0.8938	0.8774	0.9660	0.8650	7.1565	0.0393		32	0.2686
G	MC	pSVM_R	0.9284	0.8568	0.9132	0.9020	0.9613	0.8955	9.2022	0.0433		16	0.1161
G	MSC	pSVM_R	0.9243	0.8486	0.9031	0.8875	0.9718	0.8768	7.8857	0.0321		128	0.1906
G	SG	pSVM_R	0.9319	0.8638	0.9161	0.9044	0.9660	0.8979	9.4598	0.0379		32	0.1098
G	SNV	pSVM_R	0.9190	0.8380	0.8966	0.8798	0.9707	0.8674	7.3186	0.0338		128	0.2085
W	D1	pSVM_R	0.9492	0.8984	0.9401	0.9336	0.9672	0.9311	14.0571	0.0352		64	0.2233
W	D2	pSVM_R	0.9551	0.9102	0.9475	0.9421	0.9699	0.9403	16.2637	0.0321		16	0.2019
W	MC	pSVM_R	0.9521	0.9043	0.9425	0.9356	0.9712	0.9331	14.5294	0.0309		32	0.1887
W	MSC	pSVM_R	0.9548	0.9095	0.9467	0.9409	0.9705	0.9390	15.9247	0.0314		32	0.2799
W	SG	pSVM_R	0.9462	0.8925	0.9365	0.9294	0.9659	0.9265	13.1607	0.0368		16	0.1930
W	SNV	pSVM_R	0.9508	0.9016	0.9419	0.9354	0.9685	0.9331	14.4902	0.0338		64	0.1908
WG	D1	pSVM_R	0.9331	0.8662	0.9204	0.9110	0.9601	0.9061	10.2377	0.0441		64	0.1479
WG	D2	pSVM_R	0.9253	0.8506	0.9076	0.8943	0.9647	0.8859	8.4649	0.0399		128	0.1632
WG	MC	pSVM_R	0.9316	0.8633	0.9176	0.9072	0.9617	0.9015	9.7735	0.0425		128	0.1282
WG	MSC	pSVM_R	0.9279	0.8557	0.9149	0.9052	0.9558	0.8998	9.5504	0.0491		32	0.1320
WG	SG	pSVM_R	0.9316	0.8633	0.9161	0.9046	0.9651	0.8981	9.4835	0.0389		64	0.1261
WG	SNV	pSVM_R	0.9283	0.8565	0.9140	0.9034	0.9592	0.8973	9.3484	0.0455		32	0.1256
Prediction
G	D1	PLS-DA	0.8962	0.7925	0.8790	0.8652	0.9387	0.8538	6.4194	0.0718	20
G	D2	PLS-DA	0.9222	0.8443	0.8984	0.8809	0.9764	0.8679	7.3929	0.0272	20
G	MC	PLS-DA	0.8892	0.7783	0.8715	0.8571	0.9340	0.8443	6.0000	0.0782	20
G	MSC	PLS-DA	0.8939	0.7877	0.8563	0.8275	0.9953	0.7925	4.7955	0.0060	20
G	SG	PLS-DA	0.8986	0.7972	0.8767	0.8596	0.9528	0.8443	6.1212	0.0559	17
G	SNV	PLS-DA	0.9127	0.8255	0.8973	0.8855	0.9481	0.8774	7.7308	0.0591	20
W	D1	PLS-DA	0.9224	0.8447	0.9031	0.8886	0.9658	0.8789	7.9783	0.0389	20
W	D2	PLS-DA	0.9184	0.8368	0.9021	0.8897	0.9553	0.8816	8.0667	0.0507	18
W	MC	PLS-DA	0.9079	0.8158	0.8943	0.8837	0.9395	0.8763	7.5957	0.0691	19
W	MSC	PLS-DA	0.7434	0.4868	0.8138	0.9557	0.5105	0.9763	21.5556	0.5013	19
W	SG	PLS-DA	0.9066	0.8132	0.8925	0.8815	0.9395	0.8737	7.4375	0.0693	19
W	SNV	PLS-DA	0.9184	0.8368	0.9021	0.8897	0.9553	0.8816	8.0667	0.0507	19
WG	D1	PLS-DA	0.8851	0.7703	0.8681	0.8540	0.9291	0.8412	5.8511	0.0843	20
WG	D2	PLS-DA	0.8894	0.7787	0.8736	0.8607	0.9291	0.8497	6.1798	0.0835	20
WG	MC	PLS-DA	0.8775	0.7551	0.8668	0.8576	0.9054	0.8497	6.0225	0.1113	20
WG	MSC	PLS-DA	0.6816	0.3632	0.7332	0.9498	0.3834	0.9797	18.9167	0.6293	20
WG	SG	PLS-DA	0.8792	0.7584	0.8677	0.8581	0.9088	0.8497	6.0449	0.1074	20
WG	SNV	PLS-DA	0.8936	0.7872	0.8759	0.8618	0.9375	0.8497	6.2360	0.0736	20
G	D1	sSVM_L	0.6769	0.3538	0.6818	0.6984	0.6226	0.7311	2.3158	0.5161	19	5
G	D2	sSVM_L	0.6910	0.3821	0.7060	0.7485	0.5755	0.8066	2.9756	0.5263	20	9
G	MC	sSVM_L	0.5259	0.0519	0.5138	0.5301	0.4575	0.5943	1.1279	0.9127	20	7
G	MSC	sSVM_L	0.6439	0.2877	0.6416	0.6225	0.7311	0.5566	1.6489	0.4831	20	3
G	SG	sSVM_L	0.5259	0.0519	0.5156	0.5294	0.4670	0.5849	1.1250	0.9113	20	1
G	SNV	sSVM_L	0.6462	0.2925	0.6481	0.6632	0.5943	0.6981	1.9688	0.5811	20	8
W	D1	sSVM_L	0.7395	0.4789	0.7404	0.7420	0.7342	0.7447	2.8763	0.3569	20	1
W	D2	sSVM_L	0.6868	0.3737	0.6898	0.6983	0.6579	0.7158	2.3148	0.4779	18	4
W	MC	sSVM_L	0.5776	0.1553	0.5714	0.5908	0.5053	0.6500	1.4436	0.7611	20	2
W	MSC	sSVM_L	0.7224	0.4447	0.7346	0.7584	0.6526	0.7921	3.1392	0.4385	20	1
W	SG	sSVM_L	0.5763	0.1526	0.5721	0.5848	0.5263	0.6263	1.4085	0.7563	20	10
W	SNV	sSVM_L	0.6553	0.3105	0.6576	0.6705	0.6105	0.7000	2.0351	0.5564	20	1
WG	D1	sSVM_L	0.6005	0.2010	0.6004	0.6010	0.5980	0.6030	1.5064	0.6667	18	2.5
WG	D2	sSVM_L	0.6216	0.2432	0.6215	0.6254	0.6064	0.6368	1.6698	0.6180	19	8
WG	MC	sSVM_L	0.5853	0.1706	0.5799	0.6004	0.5101	0.6605	1.5025	0.7417	20	3
WG	MSC	sSVM_L	0.5524	0.1047	0.5366	0.5674	0.4409	0.6639	1.3116	0.8422	20	10
WG	SG	sSVM_L	0.5870	0.1740	0.5833	0.5977	0.5321	0.6419	1.4858	0.7289	20	2
WG	SNV	sSVM_L	0.5676	0.1351	0.5627	0.5749	0.5186	0.6166	1.3524	0.7808	20	8
G	D1	sSVM_R	0.5165	0.0330	0.1458	1.0000	0.0330	1.0000	70,000.00	0.9670	18	2	0.3
G	D2	sSVM_R	0.5259	0.0519	0.2148	1.0000	0.0519	1.0000	110,000.0	0.9481	19	2	0.2
G	MC	sSVM_R	0.5071	0.0142	0.0862	0.8000	0.0189	0.9953	4.0000	0.9858	20	2	0.2
G	MSC	sSVM_R	0.5259	0.0519	0.2500	0.8235	0.0660	0.9858	4.6667	0.9474	16	2	0.4
G	SG	sSVM_R	0.5071	0.0142	0.0862	0.8000	0.0189	0.9953	4.0000	0.9858	20	2	0.2
G	SNV	sSVM_R	0.5307	0.0613	0.2679	0.8824	0.0708	0.9906	7.5000	0.9381	18	2	0.3
W	D1	sSVM_R	0.5750	0.1500	0.4708	0.9831	0.1526	0.9974	58.0000	0.8496	20	6	0.2
W	D2	sSVM_R	0.5474	0.0947	0.3477	0.9737	0.0974	0.9974	37.0000	0.9050	20	3	0.4
W	MC	sSVM_R	0.5026	0.0053	0.0258	1.0000	0.0053	1.0000	20,000.00	0.9947	20	2	0.9
W	MSC	sSVM_R	0.5000	0.0000			0.0000	1.0000	0.0000	1.0000	18	3	0.6
W	SG	sSVM_R	0.5092	0.0184	0.0962	0.8889	0.0211	0.9974	8.0000	0.9815	20	2	0.5
W	SNV	sSVM_R	0.5013	0.0026	0.0130	1.0000	0.0026	1.0000	10,000.00	0.9974	20	2	0.3
WG	D1	sSVM_R	0.5541	0.1081	0.3885	0.9324	0.1166	0.9916	13.8000	0.8910	20	4	0.2
WG	D2	sSVM_R	0.5127	0.0253	0.1212	0.9412	0.0270	0.9983	16.0000	0.9746	19	6	0.4
WG	MC	sSVM_R	0.5093	0.0186	0.0865	1.0000	0.0186	1.0000	110,000.0	0.9814	20	4	0.3
WG	MSC	sSVM_R	0.5008	0.0017	0.0084	1.0000	0.0017	1.0000	10,000.00	0.9983	17	4	0.4
WG	SG	sSVM_R	0.5093	0.0186	0.0865	1.0000	0.0186	1.0000	110,000.0	0.9814	20	3	0.3
WG	SNV	sSVM_R	0.5025	0.0051	0.0403	0.7143	0.0084	0.9966	2.5000	0.9949	18	3	0.3
G	D1	pSVM_L	0.9033	0.8066	0.8922	0.8834	0.9292	0.8774	7.5769	0.0806		1
G	D2	pSVM_L	0.9292	0.8585	0.9182	0.9099	0.9528	0.9057	10.1000	0.0521		1
G	MC	pSVM_L	0.8844	0.7689	0.8688	0.8559	0.9245	0.8443	5.9394	0.0894		1
G	MSC	pSVM_L	0.8962	0.7925	0.8591	0.8307	0.9953	0.7972	4.9070	0.0059		1
G	SG	pSVM_L	0.8939	0.7877	0.8759	0.8615	0.9387	0.8491	6.2188	0.0722		1
G	SNV	pSVM_L	0.9080	0.8160	0.8948	0.8844	0.9387	0.8774	7.6538	0.0699		1
W	D1	pSVM_L	0.9224	0.8447	0.9075	0.8963	0.9553	0.8895	8.6429	0.0503		1
W	D2	pSVM_L	0.9224	0.8447	0.9075	0.8963	0.9553	0.8895	8.6429	0.0503		1
W	MC	pSVM_L	0.9158	0.8316	0.9086	0.9031	0.9316	0.9000	9.3158	0.0760		1
W	MSC	pSVM_L	0.6461	0.2921	0.6722	0.9744	0.3000	0.9921	38.0000	0.7056		1
W	SG	pSVM_L	0.9158	0.8316	0.9075	0.9010	0.9342	0.8974	9.1026	0.0733		1
W	SNV	pSVM_L	0.9211	0.8421	0.9079	0.8980	0.9500	0.8921	8.8049	0.0560		1
WG	D1	pSVM_L	0.8801	0.7601	0.8689	0.8594	0.9088	0.8514	6.1136	0.1071		1
WG	D2	pSVM_L	0.8995	0.7990	0.8864	0.8760	0.9307	0.8682	7.0641	0.0798		1
WG	MC	pSVM_L	0.8742	0.7483	0.8660	0.8590	0.8953	0.8530	6.0920	0.1228		1
WG	MSC	pSVM_L	0.6748	0.3497	0.7223	0.9481	0.3699	0.9797	18.2500	0.6431		1
WG	SG	pSVM_L	0.8801	0.7601	0.8682	0.8583	0.9105	0.8497	6.0562	0.1054		1
WG	SNV	pSVM_L	0.8953	0.7905	0.8814	0.8703	0.9291	0.8615	6.7073	0.0824		1
G	D1	pSVM_R	0.9387	0.8774	0.9187	0.9043	0.9811	0.8962	9.4545	0.0211		16	0.1792
G	D2	pSVM_R	0.9481	0.8962	0.9319	0.9204	0.9811	0.9151	11.5556	0.0206		32	0.2686
G	MC	pSVM_R	0.9599	0.9198	0.9420	0.9295	0.9953	0.9245	13.1875	0.0051		16	0.1161
G	MSC	pSVM_R	0.8844	0.7689	0.8671	0.8528	0.9292	0.8396	5.7941	0.0843		128	0.1906
G	SG	pSVM_R	0.9693	0.9387	0.9532	0.9422	1.0000	0.9387	16.3077	0.0000		32	0.1098
G	SNV	pSVM_R	0.9222	0.8443	0.8984	0.8809	0.9764	0.8679	7.3929	0.0272		128	0.2085
W	D1	pSVM_R	0.9434	0.8868	0.9227	0.9080	0.9868	0.9000	9.8684	0.0146		64	0.2233
W	D2	pSVM_R	0.9289	0.8579	0.9109	0.8976	0.9684	0.8895	8.7619	0.0355		32	0.2799
W	MC	pSVM_R	0.9447	0.8895	0.9257	0.9122	0.9842	0.9053	10.3889	0.0174		32	0.1887
W	MSC	pSVM_R	0.7605	0.5211	0.8281	0.9459	0.5526	0.9684	17.5000	0.4620		16	0.2019
W	SG	pSVM_R	0.9447	0.8895	0.9269	0.9142	0.9816	0.9079	10.6571	0.0203		16	0.1930
W	SNV	pSVM_R	0.9447	0.8895	0.9257	0.9122	0.9842	0.9053	10.3889	0.0174		64	0.1908
WG	D1	pSVM_R	0.9240	0.8480	0.9051	0.8910	0.9662	0.8818	8.1714	0.0383		64	0.1479
WG	D2	pSVM_R	0.9113	0.8226	0.8869	0.8684	0.9696	0.8530	6.5977	0.0356		128	0.1632
WG	MC	pSVM_R	0.9316	0.8632	0.9133	0.8998	0.9713	0.8919	8.9844	0.0322		128	0.1282
WG	MSC	pSVM_R	0.7154	0.4307	0.7824	0.9603	0.4493	0.9814	24.1818	0.5611		32	0.1320
WG	SG	pSVM_R	0.9350	0.8699	0.9172	0.9042	0.9730	0.8970	9.4426	0.0301		64	0.1261
WG	SNV	pSVM_R	0.9164	0.8328	0.8976	0.8834	0.9595	0.8733	7.5733	0.0464		32	0.1256

Abbreviations: Cultivar: W–Weiki, G–Geneva, WG–Weiki + Geneva; Correction: MSC-Multiplicative Scatter Correction, SNV-Standard Normal Variate, SG-Savitzky–Golay smoothing, MC-mean centering, D1–the first derivative, D2–the second derivative; Model: PLS-DA–partial least square discriminant analysis, SVM–support vector machine, pSVM_L–probability-based SVM with linear kernel, pSVM_R-probability-based SVM with RBF kernel, sSVM_L score-based SVM with linear kernel; Classification performance metrics: ACC–accuracy, κ–Cohen’s kappa, F₀₅–F_0.5-score, PPV–precision, TPR–true positive rate (sensitivity, recall), TNR–true negative rate, LR₊ –positive likelihood ratio, LR₋ –negative likelihood ratio, nC–number of PLS components, C–SVM penalty parameter, γ–SVM RBF kernel coefficient.

(Appendix A). Confusion matrices for the best-performing models and Area Under the Curve (AUC) values are reported in

Table A2. Confusion matrices and AUCs for the best models.

Model	TP	FN	FP	TN	AUC
Calibration
G_PLS-DA (D2)	813	39	137	715	0.949
G_sSVM_L (D2)	822	30	100	752	0.960
G_pSVM_L (D2)	790	62	117	735	0.946
G_pSVM_R (SG)	823	29	87	765	0.973
W_PLS-DA (D2)	1470	56	115	1409	0.979
W_sSVM_L (D2)	1464	62	108	1416	0.982
W_pSVM_L (D2)	1459	67	106	1418	0.979
W_pSVM_R (D2)	1481	45	93	1431	0.978
WG_PLS-DA (D2)	2216	162	299	2077	0.952
WG_sSVM_L (D2)	2200	178	278	2098	0.954
WG_pSVM_L (D2)	2192	186	281	2095	0.952
WG_pSVM_R (D1)	2283	95	223	2153	0.971
Prediction
G_PLS-DA (D2)	207	5	28	184	0.97
G_sSVM_L (D2)	122	90	41	171	0.781
G_pSVM_L (D2)	202	10	20	192	0.970
G_pSVM_R (SG)	212	0	13	199	0.991
W_PLS-DA (D2)	363	17	45	335	0.965
W_sSVM_L (D2)	250	130	108	272	0.736
W_pSVM_L (D2)	363	17	42	338	0.967
W_pSVM_R (D2)	368	12	42	338	0.959
WG_PLS-DA (D2)	550	42	89	503	0.941
WG_sSVM_L (D2)	359	233	215	377	0.656
WG_pSVM_L (D2)	551	41	78	514	0.942
WG_pSVM_R (D1)	572	20	70	522	0.969

Abbreviations: TP–true-positive, FN–false-negative, FP–false-positive, TN–true-negative, AUC–area under ROC curve; Cultivar: W–Weiki, G–Geneva, WG–Weiki + Geneva; Correction: SG-Savitzky–Golay smoothing, D1–the first derivative, D2–the second derivative; Model: PLS-DA–partial least square discriminant analysis, SVM–support vector machine, pSVM_L–probability-based SVM with linear kernel, pSVM_R-probability-based SVM with RBF kernel, sSVM_L score-based SVM with linear kernel.

(Appendix A). In addition, Receiver Operating Characteristic (ROC) curves were calculated to further assess classifier discrimination performance (Appendix A, Figure A1). Finally, representative spectral data, averaged across SSC intervals on the Brix scale (4–6, 6–8, …, 16–18), illustrate typical differences between unripe and ripe fruits (Appendix A, Figure A2).

3.1. Classification Results of ‘Geneva’ Kiwiberry

The classification models for ‘Geneva’ fruits were developed using single-year data, which limited their ability to account for seasonal variability. Moreover, the training set accounted for approximately 55% of the total cases available in the W dataset.

Table 3. ‘Geneva’ kiwiberry classification results.

	ACC (%)	κ (%)	F05 (–)	PPV (%)	TPR (%)	TNR (%)	LR+ (–)	LR− (–)	nC (2) (–)
Calibration
G_PLS-DA (D2) (1)	89.67	79.34	87.38	85.58	95.42	83.92	5.93	0.054	20
G_sSVM_L (D2)	92.37	84.74	90.53	89.15	96.48	88.26	8.22	0.040	20
G_pSVM_L (D2)	89.50	78.99	88.17	87.10	92.72	86.27	6.75	0.084
G_pSVM_R (SG)	93.19	86.38	91.61	90.44	96.60	89.79	9.46	0.038
Prediction
G_PLS-DA (D2)	92.22	84.43	89.84	88.09	97.64	86.79	7.39	0.027	20
G_sSVM_L (D2)	69.10	38.21	70.60	74.85	57.55	80.66	2.98	0.526	20
G_pSVM_L (D2)	92.92	85.85	91.82	90.99	95.28	90.57	10.10	0.052
G_pSVM_R (SG)	96.93	93.87	95.32	94.22	100.00	93.87	16.31	0.000

⁽¹⁾ G denotes ‘Geneva’ variety and parentheses include the optimal spectral data correction ⁽²⁾ nC–optimal number of PLS components selected based on κ (Cohen’s kappa) in PLS-DA and score-based models.

summarizes the performance of the best classifiers during calibration and prediction on the independent test set.

Among all preprocessing variants, the baseline G_PLS-DA model developed on D2 spectra yielded the best calibration results, with ACC = 89.67%, κ = 79.34%, and F₀₅ = 87.38. The PPV was 85.6%, indicating the probability that a positive result corresponds to unripe fruit. TPR was 95.42% and exceeded TNR by 11.5%, reflecting a tendency of the model to produce false positives (FPs). This was also confirmed by the values of the model likelihood ratios, especially the LR₊ of 5.93, indicating that for every 100 correctly classified unripe fruits, approximately 16.9 ripe ones (100/LR₊ if classes are balanced) were misclassified. This corresponds to a 16.9% risk of incorrectly confirming unripe fruits upon a positive result. At the same time, the LR₋ of 0.05 corresponds to a 94.5% chance of correctly excluding unripe fruits upon a negative result. During prediction, the G_PLS-DA model achieved slightly better performance with ACC of 92.22% and F₀₅ of 89.84, while TPR and TNR reached 97.64% and 86.79%, representing an average increase of approximately 2.5%. The balanced improvement in TPR and TNR also increased κ to 84.43% with a gain of 5.09%. Consequently, the risk of misclassifying ripe fruits decreased by 3.3%, while the chance of correctly excluding unripe fruits with a negative result increased by 2.7%.

The best G_sSVM_L classifier also used D2 data. It outperformed the baseline G_PLS-DA model during calibration, primarily due to improved TNR of 88.26%, which increased by 4.34%, and a slight increase in TPR by 1.06%. While the tendency to produce FPs by this model remained, the risk of incorrectly confirming unripe fruits upon a positive result decreased by 4.7%, and the chance of correctly excluding unripe fruits increased marginally by 1.5%. As a result, ACC, κ, and F₀₅ increased by 2.7%, 5.4%, and 3.15, respectively. However, evaluation of this model on the test set turned out to be surprisingly worse, because TPR dropped to 57.55% and TNR achieved only 80.66%, leading to ACC of 69.1%, F₀₅ of 70.6, and κ of 38.21%. Such a result suggests model overfitting despite parameter tuning during calibration.

In contrast, probability-based meta-classifiers did not exhibit such sensitivity to overfitting, maintaining consistent performance across calibration and test sets. The best G_pSVM_L model, which also utilized the D2 dataset, showed calibration performance similar to the baseline model, maintaining high accuracy in recognizing unripe fruits while showing a greater tendency to generate FNs simultaneously. The ACC, κ, and F₀₅ values of 89.5%, 78.99%, and 88.17, respectively, were slightly lower, but the differences did not exceed 1%. Although TPR decreased by 2.7%, TNR improved by 2.35% to 86.27%, favorable for reducing FPs. Accordingly, the chance of correctly excluding class A fruit on a negative result declined to 91.6%, whereas the risk of incorrectly confirming class A fruit on a positive result decreased to 14.8%. Compared with G_sSVM_L, the G_pSVM_L classifier performed slightly worse in calibration, particularly in TPR, which decreased by 3.76%, but its prediction results revealed greater robustness. On the test set, the model achieved TPR of 95.28% and TNR of 90.57%, which translated into ACC of 92.92%, F₀₅ of 91.82, and κ of 85.85%. These outcomes confirm a slightly lower imbalance between class recognition, although model precision at prediction reached only 90.99%, reflecting residual asymmetry in classification performance.

Unlike other models, the best G_pSVM_R model was developed on the SG dataset and yielded the best calibration results, with ACC = 93.19%, 3.52% higher than the baseline. Marked improvements were also observed in κ, F₀₅, and PPV, which increased by 7.04%, 4.23, and 4.86%, respectively, driven mainly by enhanced recognition of class B fruit. TNR increased substantially by 5.87%, while TPR rose by 1.17%. As a result, the chance of correctly excluding class A fruit on a negative result reached 96.2%, and the risk of false confirmation of class A on a positive result dropped to 10.6%. Like G_pSVM_L, the G_pSVM_R model proved stable during prediction, with test set metrics exceeding calibration by an average of 3.75%. TNR increased by 4.08% compared to the calibration stage, while TPR was higher by 3.4%, reaching 100%. With LR₊ of 16.31, the risk of incorrectly confirming class A fruit on a positive result was reduced to 6.1%.

3.2. Classification Results of ‘Weiki’ Kiwiberry

In contrast to the classifiers developed for ‘Geneva’ fruits, the ‘Weiki’ samples were modeled using data from two growing seasons. This suggests that they acquired a more universal character, as they accounted not only for individual fruit variability but also for inter-seasonal variation. Classification performance of the best models, at calibration and prediction stages, is shown in

Table 4. ‘Weiki’ kiwiberry classification results.

	ACC (%)	κ (%)	F05 (–)	PPV (%)	TPR (%)	TNR (%)	LR+ (–)	LR– (–)	nC (2) (–)
Calibration
W_PLS-DA (D2) (1)	94.39	88.79	93.44	92.74	96.33	92.45	12.78	0.040	18
W_sSVM_L (D2)	94.43	88.85	93.68	93.13	95.94	92.91	13.56	0.044	18
W_pSVM_L (D2)	94.33	88.66	93.69	93.23	95.61	93.04	13.76	0.047
W_pSVM_R (D2)	95.51	91.02	94.75	94.21	96.99	94.03	16.26	0.032
Prediction
W_PLS-DA (D2)	91.84	83.68	90.21	88.97	95.53	88.16	8.07	0.051	18
W_sSVM_L (D2)	68.68	37.37	68.98	69.83	65.79	71.58	2.31	0.478	18
W_pSVM_L (D2)	92.24	84.47	90.75	89.63	95.53	88.95	8.64	0.050
W_pSVM_R (D2)	92.89	85.79	91.09	89.76	96.84	88.95	8.76	0.036

⁽¹⁾ W denotes ‘Weiki’ variety and parentheses include the optimal spectral data correction ⁽²⁾ nC–optimal number of PLS components selected based on κ (Cohen’s kappa) in PLS-DA and score-based models.

.

Models developed for ‘Weiki’ fruits achieved higher classification performance during calibration than those for ‘Geneva’ fruits. Nonetheless, their performance was lower during prediction, relative to their calibration results and the prediction outcomes for the G dataset. Furthermore, all classifiers achieved their best performance on D2-corrected data, except for the W_pSVM_R model, which is consistent with the observations for ‘Geneva’ fruit classification.

The baseline W_PLS-DA model showed a high TPR of 96.33% and a slightly lower TNR of 92.45%, with a relatively high model complexity as the optimal number of components was 18, two fewer than in G_PLS-DA. The lower specificity indicates a tendency to misclassify ripe fruits as unripe, reflected in a precision below 93% and ACC, κ, and F₀₅ values of 94.39%, 88.79%, and 93.44, respectively. The chance of correctly excluding unripe fruits upon a negative result was 96%, while the risk of incorrectly confirming unripe fruits upon a positive result was 7.8%. Prediction performance of baseline model was slightly lower, with ACC, F₀₅, and κ of 91.84%, 90.21, and 83.68%, respectively. This decrease resulted mainly from a higher incidence of FPs, reducing TNR to 88.16%, while TPR remained high at 95.53%. Consequently, precision dropped by 3.77%, and the risk of falsely confirming unripe fruits on positive classification rose to 12.4%, though the chance of correctly excluding unripe fruits from a negative result remained high at 94.9%.

The score-based model (W_sSVM_L) achieved calibration results nearly identical to those of the baseline model, differing most notably in TNR, though the gap did not exceed 0.5%. However, similar to the models developed for the G dataset, its classification performance deteriorated markedly at the prediction stage. ACC declined to only 68.68%, while TPR and TNR dropped by 30.15% and 21.33%, respectively. The loss of discriminative ability for both classes resulted in PPV and F₀₅ decreasing to 69.83% and 68.98, respectively, while κ fell drastically by 51.48%. Model of such low quality, where LR₊ and LR₋ reached 2.31 and 0.48, respectively, entails as much as 43.2% risk of false confirmation for class A on positive result and only 52.2% chance of its correct exclusion on negative result.

During calibration, the performance metrics of the W_pSVM_L classifier closely matched those of both the baseline and W_sSVM_L models. Sensitivity stabilized at 95.61% and specificity of 93.04%, resulting in a PPV 0.1% higher than W_sSVM_L and 0.48% above the baseline. ACC, κ, and F₀₅ reached 94.33%, 88.66%, and 93.69, indicating no meaningful improvement over the baseline model. Although the LR₊ of 13.76 reflected a 0.5% reduction in the risk of falsely confirming unripe fruit upon a positive result, the LR₋ of 0.05 implied a 0.8% lower probability of correctly excluding unripe fruits upon a negative result. At the prediction stage, the performance of W_pSVM_L declined slightly relative to calibration. Similarly to its counterpart trained on the G dataset, the main factor was a 4.1% drop in TNR, which reduced PPV by 3.6% and lowered ACC by 2.09%. The lower ability to recognize ripe fruit was reflected in κ = 84.47% and a decrease of F₀₅ to 90.75. Consequently, with a 0.3% lower chance of correctly excluding unripe fruit upon a negative result, the risk of a false confirmation of fruit unripeness on a positive result increased by 4.3%, although it remained 0.8% lower than for the baseline model. Compared to the baseline model, the W_pSVM_L classifier generates FPs less frequently, but at the cost of increasing the number of FNs.

Within the same group of discriminative models, the W_pSVM_R classifier achieved slightly better performance than both the baseline and W_pSVM_L model during calibration. With a TPR of 96.99% and a TNR of 94.03%, it combined very high detection of unripe fruit with a good ability to avoid FPs. This balance translated into a PPV of 94.21% and ACC of 95.51%, outperforming W_pSVM_L by 1.46% and 1.11% in both measures. Agreement with the reference classes was also excellent, as reflected in κ = 91.02% and F₀₅ = 94.75. Moreover, the highest LR₊ value within this group indicates that the risk of incorrectly confirming unripe fruit upon a positive result was only 6.1%. At the same time, the lowest LR₋ underscores a 96.8% chance of correctly excluding unripe fruit when the result is negative. On the prediction stage, the advantages of W_pSVM_R diminished but did not disappear. Sensitivity remained almost unchanged at 96.84%, yet specificity fell by about 5%, reducing precision to 89.76% and accuracy to 92.89%. The κ statistic dropped to 85.79% and F₀₅ to 91.09, indicating a noticeable decline in the ability to limit FPs compared with calibration. The LR₊ decreased to 8.76, nearly doubling the risk of incorrectly confirming unripe fruit upon a positive result to 11.4%, while the LR₋ slightly increased, reducing the chance of correctly excluding unripe fruit from a negative result only by 0.3%. Despite decreases in some model capabilities, the classifier maintained TNR and PPV higher than W_pSVM_L in prediction and outperformed the baseline model in balancing sensitivity with FPs control.

3.3. Classification Results for the Combined Data of ‘Geneva’ and ‘Weiki’ Kiwiberry

After analyzing the results obtained separately for the ‘Weiki’ and ‘Geneva’ varieties, another step was to train classifiers on the combined dataset (WG). The pooled data were intended to capture seasonal and varietal variability, thereby allowing an assessment of model generalization capacity and searching for a more universal solution for postharvest fruit sorting. This approach makes it possible to verify whether integrating data from different cultivars enhances the stability and broad applicability of the developed algorithms. Results obtained for the best classifiers, at calibration and prediction stages, are shown in

Table 5. Combined ‘Geneva’ and ‘Weiki’ kiwiberry classification results.

	ACC (%)	κ (%)	F05 (–)	PPV (%)	TPR (%)	TNR (%)	LR+ (–)	LR– (–)	nC (2) (–)
Calibration
WG_PLS-DA (D2) (1)	90.30	80.61	89.08	88.11	93.19	87.42	7.41	0.078	20
WG_sSVM_L (D2)	90.41	80.82	89.50	88.78	92.51	88.30	7.91	0.085	19
WG_pSVM_L (D2)	90.18	80.35	89.32	88.64	92.18	88.17	7.80	0.089
WG_pSVM_R (D1)	93.31	86.62	92.04	91.10	96.01	90.61	10.24	0.044
Prediction
WG_PLS-DA (D2)	88.94	77.87	87.36	86.07	92.91	84.97	6.18	0.083	20
WG_sSVM_L (D2)	62.16	24.32	62.15	62.54	60.64	63.68	1.67	0.618	19
WG_pSVM_L (D2)	89.95	79.90	88.64	87.60	93.07	86.82	7.06	0.080
WG_pSVM_R (D1)	92.40	84.80	90.51	89.10	96.62	88.18	8.17	0.038

⁽¹⁾ WG denotes the combined dataset of ‘Geneva’ and ‘Weiki’ varieties and parentheses include the optimal spectral data correction ⁽²⁾ nC–optimal number of PLS components selected based on κ (Cohen’s kappa) in PLS-DA and score-based models.

.

The best calibration results were obtained for models trained on D2-corrected data, which confirmed the trends observed for the individual datasets, except for G_pSVM_R. As for the G and W datasets, all WG models tended to generate FPs. For the WG_PLS-DA classifier, the difference between TPR and TNR at the calibration stage reached 5.77%. This imbalance resulted in a PPV of 88.11%, while ACC, κ, and F₀₅ were 90.3%, 80.61% and 89.08, respectively. The chance of correctly excluding unripe fruit given a negative result was as high as 92.2%. However, simultaneously, the classifier carried a 13.5% risk of incorrectly confirming unripeness given a positive result. At the prediction stage, the performance of this classifier slightly declined. The main reason was its weaker ability to correctly identify ripe fruits, with specificity reduced to 84.97%, while sensitivity decreased marginally by 0.28%. This decline lowered the accuracy to 88.93% κ to 77.88%, and F₀₅ to 87.36. As a consequence, the risk of misclassifying ripe fruit as unripe, having a positive result, rose to 16.2%, and the chance of correctly excluding ripe fruit with a negative result decreased to 91.7%.

Classification based on PLS components did not bring satisfactory results, even though at the calibration stage, the performance metrics of the WG_sSVM_L model were very close to those of WG_PLS-DA. Slightly higher TNR by 0.88% and TPR of 92.51% translated into a 91.5% chance of correctly excluding unripe fruit upon a negative result and a 12.6% risk of incorrectly confirming unripeness having a positive result. Accuracy and precision nearly overlapped with the baseline model, with ACC higher by 0.11%, κ by 0.21%, F₀₅ by 0.42, and PPV by 0.67%. However, as observed for the models trained on G and W datasets separately, this classifier also proved unstable during prediction. The quality of results dropped sharply, with most metrics decreasing by about 30% and κ falling by more than 55%. Such a decline demonstrates the lack of robustness and indicates that the model suffered from overfitting.

Models based on probabilities proved considerably more stable, as reflected by only minor decreases in their metrics during prediction. Calibration results indicate that classifier WG_pSVM_L differed little from the WG_PLS-DA and WG_sSVM_L models. The most notable differences were observed in TPR, which was 1.01% lower, and TNR, which was 0.76% higher compared with the baseline. As for the remaining metrics, WG_pSVM_L showed no more than 0.55% deviations.

The RBF-kernel SVM classifier followed the same overall trend as the earlier pSVM_R models, although it used D1 spectra correction instead of D2. It delivered the best classification performance and showed strong stability, which was reflected in only minor differences between calibration and prediction metrics. Relative to the baseline model, WG_pSVM_R improved sensitivity and specificity by 2.82% and 3.2%, reaching 96.01% and 90.61% in calibration. This performance translated into a PPV of 91.1% and higher overall classification accuracy, with ACC of 93.31%, κ of 86.62%, and F₀₅ of 92.04%. At this stage, the model lowered the risk of incorrectly confirming unripe fruit to 9.8% for positive outcomes, while raising the chance of correctly excluding unripe fruit after a negative result to 95.6%. In prediction, the model showed a decline in TNR to 88.18%, accompanied by a slight increase of 0.62% in TPR, reflecting a compromise between the opposite tendencies previously observed in the single-variety datasets. The model improved TPR and TNR with ‘Geneva’ fruits, whereas with ‘Weiki’ fruits, both metrics decreased, with TNR showing a powerful decline. In the combined dataset, these effects were partially balanced, resulting in a moderate increase in TPR and a slight reduction in TNR. This outcome illustrates the consequences of merging data representing different kiwiberry varieties. Integration enhances stability and generalization ability, but at the same time, it diminishes the positive effects observed for the better-recognized fruits and partly alleviates the adverse effects seen for the variety with more heterogeneous fruit traits, which are therefore harder to discriminate in terms of ripeness. As a result, in prediction on the WG dataset, the pSVM_R classifier showed a minimally increased chance of correctly excluding an unripe fruit with a negative result, which reached 96.2%, and a higher risk of misclassifying a ripe fruit as unripe upon a positive result, which rose to 12.2%. Nevertheless, WG_pSVM_R achieved the best performance within this group, reaching a PPV of 89.1%, an ACC of 92.4%, a κ of 84.8%, and an F₀₅ of 90.15% in prediction.

For the combined WG dataset, the models performed better than those for ‘Geneva’ variety during calibration but lost this advantage in prediction, yielding lower metrics. Compared with models developed for ‘Weiki’ fruits, WG classifiers were weaker in both stages. Across classifiers trained on the WG dataset, sSVM_L models exhibited the lowest stability, whereas PLS-DA, pSVM_L, and pSVM_R maintained higher robustness. Among these, the sSVM_R model delivered the best overall performance.

3.4. Supplementary Analyses of Classifier Performance

To complement the reported classification metrics, we performed statistical validation of the models. McNemar’s test confirmed that differences between PLS-DA and pSVM_R were statistically significant for the G and WG datasets. In contrast, no significant difference was observed for ‘Weiki’, indicating comparable performance of both models in this case (Table A4, Appendix A). Narrow ranges of classifiers’ key metrics confirmed that pSVM_R achieved higher performance than PLS-DA, with narrower ranges for ACC, kappa, and PPV. In particular, ACC for Geneva increased to 0.95–0.98 compared to 0.89–0.94 for PLS-DA, while improvements were also evident for the combined dataset (0.91–0.94 vs. 0.87–0.91). In the case of ‘Weiki’, the differences were smaller but still consistent across metrics (Table A3, Appendix A). The superior classification performance of pSVM_R, together with the observed stability across bootstrap confidence intervals, motivates evaluation of model inference times to assess the practical feasibility of the proposed pipeline. Inference time measurements were performed separately for the PLS-DA and SVM stages because the models were generated using two independent, research-grade scripts, written for methodological purposes and not explicitly optimized for computational efficiency. Specifically, we recorded the time for the best PLS-DA models to generate probabilities for the test samples, which were subsequently used as input to the best SVM classifiers, and the time for the trained SVM models to classify the same samples. The measured inference times (below 0.2 ms per fruit) indicate that PLS-DA and PLS-DA + SVM models operate with high computational efficiency on a standard personal computer, and the additional complexity introduced by the SVM does not result in a notable computational burden (Table A5, Appendix A).

4. Discussion

Our experiment showed that the traditional use of PLS-DA gave pretty good results, but pSVM models performed better. A common challenge across classifiers was the imbalance between TPR and TNR. The extent of this imbalance, however, depended on the dataset. Classifiers developed independently for each variety revealed opposite tendencies. Models designed for ‘Geneva’ fruits reduced the initially wide gap between TPR and TNR observed in calibration and improved both metrics at the prediction stage. In the case of models developed for ‘Weiki’ fruits, the opposite tendency was observed. Here prediction led to a decline in both metrics, with the decrease being particularly pronounced for TNR and only marginal for TPR. Similarly, in the combined WG dataset, the gap was less pronounced during calibration but widened again at the prediction stage. Interestingly, the gap between calibration and prediction was greater in the variety-specific models than in the combined one, even for TNR, which showed the highest variability across scenarios.

Such pattern suggests that our models focus more on the correct recognition of the winning class during training, which leads to their sensitivity improvement, limiting the reliability in identifying ripe fruits simultaneously. More importantly, the differences between calibration and prediction were more pronounced in variety-specific models than in the combined dataset, underlining the lower stability of single-variety approaches.

The decline in TNR of examined models mirrors observations by Lee et al. [27], who reported that normalization methods can strongly influence specificity in kiwifruit classification. Furthermore, Yang et al. [62] and Li et al. [36] revealed that SVM classifiers outperformed PLS-DA models across different fruit species and varying sets of input traits, though exceptions exist. For a change, Bakhshipour [21] reported that a PLS-DA classifier combined with SGD1 spectral correction slightly outperformed the corresponding SVM model in discriminating Hayward kiwifruit by ripeness. Moreover, Benelli et al. [20] demonstrated that spectral pre-processing combined with variable selection substantially improved PLS-DA performance.

A further important aspect of this study was the evaluation of the combined WG dataset, which allowed assessment of classifier generalization under increased data variability and highlighted broader practical and methodological implications. In practical terms, this outcome highlights an essential methodological and technological challenge. Real sorting systems are unlikely to work with fruits of only one variety. Thus, the universality of a model becomes a valuable advantage. At the same time, there is a risk of compromise, since combining data may enhance overall generalization capacity, yet it can also reduce prediction accuracy for individual varieties, as the model must cope with a broader range of variability. From a methodological standpoint, this approach tests the model’s robustness, since high performance obtained on the combined dataset demonstrates that the selected spectral features are stable and highly discriminative. On the other hand, the practical consequences are twofold, because weaker performance may indicate that, in industrial settings, separate models for each variety could be more effective, or that adaptive systems capable of adjusting parameters to the specific fruit batch would offer a better solution. These outcomes are consistent with, yet not identical to, earlier studies. Sarkar et al. [41] noted that models dedicated to individual species of hardy kiwi characterized by better prediction accuracy than those trained on combined data. He highlighted the risk of reduced performance when variability between cultivars is not fully represented. On the other hand, Mishra [63] showed that global NIR models built on multi-fruit datasets may surpass variety-specific approaches, as long as the spectral profiles of the cultivars are sufficiently similar. Our results for the WG dataset reflect these tendencies, highlighting both the potential of shared models and the trade-offs inherent in balancing universality with cultivar-specific accuracy.

McNemar’s test indicated statistically significant differences between PLS-DA and pSVM_R classifiers for G and the combined WG dataset, while no significant difference was found for ‘Weiki’. This reflects the smaller size and the single-season scope of the Geneva dataset, which emphasized differences between models, and the more heterogeneous, two-season W dataset, where both classifiers achieved similarly high performance. Interestingly, despite the smaller proportion of ‘Geneva’ samples in the combined dataset, the differences remained significant, highlighting the greater robustness of the pSVM_R classifier in capturing cultivar-specific variability.

Another key finding of our study is the consistent weakness of sSVM_L classifiers, which suffered from severe prediction declines across ‘Geneva’, ‘Weiki’, and WG datasets. Such results point to model overfitting, despite its hyperparameter optimization. In contrast, pSVM_L and pSVM_R classifiers maintained higher overfitting robustness, with the latter ensuring the best trade-off between calibration and prediction performance. Its improvements in TNR and PPV substantially reduced the risk of false confirmations of fruit unripeness, lowering it to just over 6%. These results show that stacking models enriched with information on class probability may improve ripeness discrimination and reduce the overfitting observed in score-based SVM models. The consistent stability of pSVM_R across G, W and WG datasets highlights the potential of probability-driven approaches for robust, non-invasive fruit sorting. Probability-based SVM classifiers, particularly those with RBF kernels, provided the most reliable performance across examined kiwiberry varieties. Nevertheless, the observed drop in TNR for ‘Weiki’ and WG datasets suggests that further refinement may be needed, possibly through adaptive or hybrid approaches. At the same time, the comparative analysis of variety-specific and combined models indicates that generalization is achievable without large losses in accuracy, though this comes at the cost of slightly lower peak performance than the best single-variety models. Despite using separate, non-optimized implementations of PLS-DA and SVM models, measured per-sample inference times were very small (on the order of 0.07–0.16 ms), indicating that the approach is computationally efficient even without code-level optimization. The observed performance and low inference times on a standard personal computer suggest that real-time deployment is feasible, with expected hardware requirements limited to a mid-range CPU and 16 GB RAM. Therefore, code optimization, system integration, and validation under commercial-scale conditions remain subjects for future work.

Since hyperspectral imaging provides rich spatial and spectral information, industrial applications may benefit from point-wise measurements targeting only selected, non-contiguous wavelengths rather than the full VNIR or SWIR range. This represents another potential direction for future work, which does not exclude using meta-inspired models but would involve fewer input variables and could reduce system complexity and costs.

5. Conclusions

This study evaluated a meta-inspired classification framework integrating hyperspectral imaging with linear and nonlinear modeling strategies for postharvest sorting of kiwiberry, a fruit characterized by high perishability and uneven ripening. The comparative analysis of PLS-DA, score-based SVM (sSVM), and probability-based SVM (pSVM) models demonstrated apparent differences in their robustness and practical applicability. Variety-specific models achieved high calibration accuracy, yet their prediction stability varied, with sSVM models consistently showing overfitting. In contrast, probability-driven approaches, particularly RBF-based pSVMs, provided superior performance across datasets, achieving predictive accuracies above 92% and substantially reducing false classifications.

Combining data from different cultivars highlighted classification specificity and accuracy trade-offs for individual varieties, which suggests a potential for generalization. However, since only two cultivars were included, further studies involving a broader range of kiwiberry genotypes are needed to assess model robustness within the species adequately.

Beyond methodological insights, these outcomes have practical implications for kiwiberry production. Reliable discrimination between ripe and unripe fruits facilitates effective postharvest management, supporting procedures such as rapid cooling or modified atmosphere storage, which help extend shelf life and slow down postharvest ripening. Such approaches are crucial for expanding the market availability of kiwiberry and reducing postharvest losses, thereby improving fruit quality management in supply chains. Moreover, combining data from different cultivars demonstrated that generalization is possible without significant losses in accuracy, offering a promising pathway toward universal classification systems that can accommodate fruit variability encountered in real-world conditions.

Importantly, this study systematically evaluates PLS-DA, sSVM, and pSVM approaches and demonstrates the novelty of integrating probability- and score-based models derived from PLS-DA into SVM frameworks. This is the first application of such an integrated strategy to Actinidia arguta, offering insights that may be transferable to other highly perishable fruits. From an industrial perspective, the results support the development of automated, non-destructive sorting systems that could be incorporated into postharvest processing lines, improving consistency and efficiency in fruit quality management.

Nevertheless, limitations remain, as the study was restricted to two cultivars and laboratory conditions. The current models were implemented using research-grade, two-stage scripts that are not fully optimized for runtime. However, measured inference times on a standard personal computer, statistically validated performance, and confidence intervals indicate clear potential for real-time deployment. Future work should follow a two-pronged approach. First, one should validate the proposed models on a wider genetic pool. On the other side, future studies should focus on code optimization, system-level integration, and testing the models under operational conditions on pilot sorting lines to evaluate hardware requirements, integration challenges, and economic feasibility.