论文信息

论文解读

AI 辅助整理
方法
按感知模态与决策方法梳理文献,涵盖RGB、热、光谱多模态。
发现
准确率68%至99%以上,但协议不一不可比,缺公开基准与田间验证。
机会点
可构建统一公开基准与田间验证平台,推动多模态分选模型可复现比较。

论文摘要

综述椰枣自动分选研究,按感知模态、决策方法与分选目标梳理经典视觉、模糊逻辑、化学计量学及深度学习(含RGB/热/光谱多模态)。报告准确率约68%至99%以上,但任务与数据协议不一不可直接比较;公开基准少、田间验证近乎缺失。

全文

MDPI Agriculture

Academic Editors: Heng Lu, Xiaoai
Dai and Lei Ma
Received: 15 October 2026
Revised: 9 November 2026
Accepted: 16 November 2026
Published: 17 November 2026
Copyright:© 2026 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license.
Review
Date Fruit Sorting: From Classical Vision to Deep Learning
Hanafy M. Omar *
and Saad M. S. Mukras
Department of Mechanical Engineering, College of Engineering, Qassim University,
Buraidah 51452, Saudi Arabia; mkras@qu.edu.sa
*Correspondence: hanafy@qu.edu.sa
Abstract
Date fruit (Phoenix dactyliferaL.) is a major economic and food-security crop across arid
and semi-arid regions, with global production around 10 million tons annually, led by
Saudi Arabia, Egypt, Algeria, Iran, and Iraq. Sorting and grading dates by size, color,
maturity, and surface quality remains largely manual, slow, costly, and inconsistent be-
tween inspectors, motivating automated systems that meet export-quality standards. This
review synthesizes the literature on automated date-fruit sorting by methodological family
(classical computer vision, rule-based and fuzzy-logic grading, chemometric classifica-
tion, and deep learning, including multimodal systems fusing RGB, thermal, weight, and
spectral data) and categorizes studies by sensing modality, decision method, and sorting
target. Reported accuracies span roughly 68% to above 99%, but they were obtained on
different tasks, datasets, and evaluation protocols and are not directly comparable; most
come from small, single-institution datasets, and physical sorting machines with published
mechanical detail remain rare relative to algorithm-only studies. Persistent gaps limit
real-world adoption: few public benchmark datasets, narrow varietal coverage, near-absent
field validation, inconsistent evaluation protocols, and sensing modalities (validated inline
internal-quality sensing, acoustic firmness, real-time tracking) still minimally explored on
dates despite maturity elsewhere. Closing these gaps is essential for moving automated
date sorting from laboratory demonstrations to industry-wide use.
Keywords:Phoenix dactylifera; date fruit sorting; computer vision; machine learning; deep
learning; fuzzy logic; chemometrics; quality assessment

  1. Introduction
    1.1. Background and Economic Context
    The date palm (Phoenix dactyliferaL.) is one of the oldest cultivated fruit trees, with ar-
    chaeological evidence of cultivation in Mesopotamia dating back several millennia [1]. It
    has long sustained human populations in arid environments, providing food, materials,
    and cultural value across the Middle East and North Africa [ 2]. Today, date cultivation
    extends across several dozen countries, predominantly in the Middle East and North Africa
    (MENA) region, with smaller-scale production emerging elsewhere.
    Global date production averaged about 9.7 million tons per year in 2022–2024 [ 3],
    led by Saudi Arabia (18.7%), Egypt (18.2%), Algeria (13.4%), Iran (10.9%), Iraq (6.5%),
    and Pakistan (5.9%) (Table 1). Keskin et al. [4], analyzing FAOSTAT data from 1961–2022,
    project world production in the range of 10.1–11.0 million tons annually through 2028.
    By continent, Asia accounted for about 55% of world production in 2022–2024 and
    Africa for about 44% (Table 1). The main Asian producers are Saudi Arabia, Iran, Iraq,
    Agriculture2026,16, 2532 https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 2 of 20
Pakistan, Oman, and the United Arab Emirates, and African production is concentrated
in Egypt, Algeria, Sudan, Tunisia, Libya, and Morocco. The Americas contributed less
than 1%, mainly from the United States (about 57,000 tons) and Mexico (about 18,000 tons),
where Medjool and Deglet Nour dominate; FAOSTAT records only a negligible amount for
Europe and no production for Oceania.
Table 1.Date production by country and continent (FAOSTAT, 2022–2024 means) [3].
Rank Country Share of World Production (%)
1 Saudi Arabia 18.7
2 Egypt 18.2
3 Algeria 13.4
4 Iran 10.9
5 Iraq 6.5
6 Pakistan 5.9
Continent
1 Asia 54.9
2 Africa 44.1
3 Americas 0.8
4 Europe 0.2
– Oceania Not recorded
Date fruits pass through four well-documented ripening stages, each with a distinct
moisture and sugar profile that directly affects sorting criteria [ 5] (Figure 1). Quality
parameters relevant to sorting include size, color, shape, skin integrity, sugar content,
moisture, texture, and defects such as mechanical damage, insect infestation, or fungal
contamination; these vary substantially across the thousands of documented cultivars and
across developmental stages [6].
(a) Kimri
(b) Khalal
(c) Rutab
(d) Tamr
Kimri
Unripe
Green, hard
Khalal
Full-size
Yellow/red, crunchy
Rutab
Softening
Soft, brown
Tamr
Fully ripe
Dark, low moisture
Ripening Ripening Ripening
Rising sugar → Rising sugar → Falling moisture → Low moisture
Figure 1.Date-fruit ripening stages and sugar/moisture trends [5,7].
1.2. Motivation for Automated Sorting
Manual sorting is widely described in the date-grading literature itself as labor-
intensive and inconsistent between inspectors [8,9]. This has motivated sustained interest
in automated alternatives that improve consistency, help meet export-grading standards,
and reduce reliance on skilled seasonal labor. This review covers automated sorting,
grading, and classification systems built specifically for date fruit; the wider fruit- and
vegetable-sorting literature is touched on only for context (Section 8). Readers wanting
a broader look at AI across the full date-palm harvesting and postharvest chain should
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 3 of 20
also consult the recent review by Noutfia and Ropelewska [10], which is complementary in
scope to this one.
1.3. Scope and Method of the Review
Studies were identified by searching Scopus, Google Scholar, and Google Patents; the
final search was run on 4 November 2026. Scopus titles, abstracts, and keywords were
searched with (“date fruit” OR “date fruits” OR “date palm” OR “Phoenix dactylifera”)
AND (sort* OR grad* OR classif* OR “quality inspection”) AND (“computer vision” OR
“machine vision” OR “image processing” OR “machine learning” OR “deep learning” OR
“neural network” OR fuzzy OR spectral OR hyperspectral OR “near-infrared” OR “X-
ray”). Equivalent terms were used in Google Scholar, where the first 100 results sorted
by relevance were screened, and in Google Patents abstracts, to capture physical sorting
machines not indexed by the academic databases (Section 7). Backward and forward
citation tracking supplemented the searches.
A study was eligible if it built, tested, or benchmarked an automated system for
sorting, grading, or classifying date fruit (Phoenix dactylifera) on any attribute, or measured
date-fruit quality non-destructively with a sensor suitable for sorting. Studies on date-palm
trees, palm-based materials, other foods, and jujube (“red date”) were excluded, as were
theses and preprints. Scopus returned 248 records, of which 126 were excluded on title and
122 were eligible; Google Scholar added 3 eligible studies, and Google Patents returned
9 records, of which 1 was eligible. Because the eligible literature is large, representative
studies of each method family, sensing modality, and sorting target are discussed individu-
ally, with priority given to highly cited, recent, or methodologically distinct work, physical
sorting machines, and non-destructive sensing studies; studies that reuse the public feature
dataset of Koklu et al. [11] are summarized as a group (Section 5). The remaining eligible
studies were counted but are not discussed individually. Quantitative results were checked
against each study’s original publication or abstract where accessible.
2. A Taxonomy of Date Fruit Sorting Approaches
The studies are categorized here along three largely independent axes: what physical
signal is sensed, what decision-making method turns that signal into a sorting decision,
and what attribute of the fruit is sorted on. These three axes map directly onto the stages
of a generic automated sorting system (Figure 2): a sensor acquires a signal from the
fruit, a decision method converts that signal into a class label, and an actuator (where one
exists) routes the fruit accordingly. Figure 3 summarizes the three axes and their categories,
described in full in the three subsections below; studies discussed later in this review are
cross-referenced against all three in Table 2.
Sensing Decision method Actuation Sorted fruit
Signal Class label Routing
Axis 1
RGB, thermal,
spectral, physical,
multimodal, other
Axis 2
Threshold, classical
ML, fuzzy,
deep learning
Conveyor, paddle
wheel, piston, air jet
Axis 3
Variety, maturity,
quality, hardness,
size, composition
Figure 2.Generic date-fruit sorting system and the three taxonomy axes.
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 4 of 20
Axis 1
Sensing Modality
Axis 2
Decision Method
Axis 3
Sorting Target
Visible/RGB imaging
Physical/geometric
measurement
Thermal imaging
Spectral reflectance
Destructive
chemical analysis
(offline only, not real-time)
Multimodal fusion
Other non-imaging
(gas, dielectric, X-ray)
Threshold/rule-based
Classical statistical/ML
Fuzzy/neuro-fuzzy
Multivariate statistics
Deep learning:
scratch/transfer/attention/
detection
Variety/cultivar
Maturity/ripening stage
Quality/defect grading
Hardness/texture
Size
Internal composition
Figure 3.Taxonomy of date-fruit sorting approaches.
Table 2.Studies grouped by taxonomy category.
Sensing Modality Decision Method Sorting Target Studies
RGB imaging Threshold/rule-based Quality + size [12]
RGB imaging Threshold/rule-based Maturity stage [13]
RGB imaging Classical statistical/ML Variety [11,14–17]
RGB imaging Classical statistical/ML Quality grading [8,18]
RGB imaging Classical statistical/ML Hardness [19]
RGB imaging Deep learning Variety [20–31]
RGB imaging Deep learning Maturity stage [32,33]
RGB imaging Deep learning Defect detection [34]
RGB imaging Dataset only Quality + size [35]
RGB imaging Dataset only Maturity stage [36]
Physical measurement Fuzzy/neuro-fuzzy Quality grading [9,37]
Spectral sensing Threshold/rule-based Quality grading [38]
Spectral sensing Threshold/rule-based Quality + size [39]
Spectral sensing Threshold/rule-based Defect detection [40]
Spectral sensing Classical statistical/ML Defect detection [41,42]
Spectral sensing Chemometric Maturity stage [43,44]
Spectral sensing Chemometric Internal composition [45]
Multimodal Deep learning Quality grading [46–48]
Destructive chemical Chemometric Variety [49]
Gas, dielectric, X-ray Classical statistical/ML Maturity stage [50]
Gas, dielectric, X-ray Classical statistical/ML Defect detection [51]
Gas, dielectric, X-ray Classical statistical/ML Internal composition [52,53]
2.1. Axis 1: Sensing Modality
Most systems reviewed here rely on standard visible-light/RGB imaging, either alone
or as the base layer under every other imaging-based approach discussed below; it is the
dominant modality in this literature. A second, non-imaging family measures physical or
geometric properties directly: weight, length, width, thickness, or projected area, typically
via caliper, scale, or simple geometric analysis of a binarized image. This family appears
throughout the fuzzy-logic and chemometric work of Section 4. Thermal imaging appears
in only one line of work, always combined with RGB (Section 5). Spectral sensing has
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 5 of 20
a longer history on dates than the imaging literature suggests, from spectrophotometric
thresholds in 1993 [40] and near-infrared (NIR) maturity sorting around 2000 [43,44] to NIR
and hyperspectral imaging, mid-infrared, gas, dielectric, and X-ray sensing (Section 5). A
further, destructive family uses chemical analysis (high-performance liquid chromatog-
raphy, HPLC; texture-profile analysis; and colorimetry performed on sacrificed samples),
unsuited to real-time sorting but useful for offline variety characterization (Section 4).
Finally, a small number of studies deliberately fuse two or more of these modalities in a
single pipeline, discussed together as multimodal systems in Section 5. A study is classed
here as multimodal only when signals from two or more different sensors are used in the
sorting decision itself; features computed from a single camera image do not qualify, nor
does a spectrometer used offline only to choose filters or thresholds, as in [40].
2.2. Axis 2: Decision-Making Method
Decision-making methods in this literature span a spectrum. At one end are fixed,
expert-defined thresholds and decision trees (DTs) with no data-driven training step. Next
come classical statistical and machine-learning classifiers, such as linear and stepwise
discriminant analysis (LDA/SDA), k-nearest neighbors (k-NN), support vector machines
(SVMs), logistic regression (LR), Random Forest, Bayesian networks, and small (shallow)
neural networks like a multilayer perceptron (MLP) or back-propagation neural network
(BPNN), typically applied to hand-crafted features. Fuzzy and neuro-fuzzy inference
follows: expert-rule Mamdani systems and their data-adaptive extension (ANFIS) sit along-
side a separate, largely unsupervised strand of multivariate statistics (principal component
analysis, PCA, and hierarchical clustering) used exploratively in the chemometric literature.
Deep learning covers several approaches, distinguished by how the network acquires its
weights and what it outputs. A CNN may be trained from scratch, with no pretrained
weights, or pretrained on a large generic image corpus such as ImageNet and fine-tuned on
date-fruit images (transfer learning, the most common deep-learning approach here). It may
also extend a CNN backbone with an attention mechanism such as Squeeze-and-Excitation
(SE) to weight the most discriminative image regions more heavily. Object detectors such
as YOLO localize and classify fruit or defects within an image at once, instead of classifying
a single pre-cropped image.
2.3. Axis 3: Sorting Target
The literature reviewed here sorts dates on six largely distinct attributes. Variety
or cultivar identification asks which named cultivar (e.g., Khalas, Ajwa, Deglet Nour)
a given fruit belongs to, and is the single most common target across the studies be-
low. Maturity or ripening-stage classification instead asks where a fruit sits along the
Kimri→Khalal→Rutab→Tamr progression (Section 3), while external quality or defect
grading assigns an overall quality tier or identifies the presence and type of a specific
surface defect. Hardness or texture grading is treated separately from general “quality”
here because so few studies target it directly; Manickavasagan et al. [19] is the clearest exam-
ple. Size grading, sorting purely by physical dimension independent of quality or variety,
appears mainly as one component of a combined grading scheme (e.g., [12,35]) and rarely
as a task on its own. Finally, spectral and dielectric studies target internal composition
(moisture, dry matter, and sugar), which image-based studies do not measure directly.
Two patterns are visible once the literature is organized this way. First, nearly every
decision-making method in Axis 2 has been applied to nearly every sorting target in
Axis 3 by at least one study. Second, Axis 1 is much less balanced. The large majority
of studies use visible/RGB imaging alone; thermal sensing appears in only one line of
work, combined with RGB and weight [ 46,47]; only one study fuses RGB imaging with
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 6 of 20
VisNIR spectral sensing in a single decision [48]; and spectral, dielectric, gas, and X-ray
sensing, although applied to dates in twelve studies (Section 5), remain separate from
the image-based sorting pipelines. This is a gap relative to the broader produce-quality-
assessment literature.
Figure 4 tallies the 45 studies in Table 2 against each axis (studies reporting a combined
category, e.g., “quality + size,” are counted under that combined label); RGB imaging and
deep learning each dominate their axis, and variety identification is the most common
target. In Figure 4, Multimodal counts only studies that use two or more sensors in the
sorting decision, and Other covers gas, dielectric, and X-ray sensing.
RGB
PhysicalSpectral
MultimodalDestructive
Other
0
10
20
30 27
2
8
3 1 4
Axis 1: Sensing modality
Threshold
Classical ML
Fuzzy
MultivariateDeep learning
N/A
0
10
20
5
14
2
4
18
2
Axis 2: Decision method
VarietyQuality
Quality+Size
MaturityHardnessDefect
Composition
0
10
20 18
8
3
7
1
5
3
Axis 3: Sorting target
Figure 4.Number of studies per category on each taxonomy axis.
3. Classical Computer Vision and Rule-Based Grading
Early work on automated date sorting (roughly 1993–2015) relied on hand-crafted
color, shape, and texture features paired with statistical or rule-based classifiers, and, in
some cases, on direct image thresholding without a trained classifier at all.
Al-Ohali [8] built an early complete computer-vision date-grading system, extracting
external quality features (flabbiness, size, shape, surface wrinkling, visible defects) and
classifying using a back-propagation neural network. Trained on 1200 samples and tested
on 660 (three grades), the best model (color + diameter) reached 80% overall accuracy
(71%/80%/66% by grade). Flabbiness and bruise detection were the weakest points, and the
study recommended touch sensing and unsupervised learning; these recommendations
were only partially taken up, as later sections show.
Pourdarbani et al. [13] built and physically tested a complete online sorting machine
for Barhee-cultivar dates: a belt conveyor feeds fruit past a four-armed paddle wheel whose
exit port is set by two optical sensors and step motors (Figure 5). A weighted RGB color
index (Index = 0.167R+ 0.253G− 0.578B) classifies each fruit’s maturity stage (Khalal,
Rotab, Tamar, equivalent to Rutab and Tamr elsewhere in this review) as it arrives. Tested
online on 100 samples per stage across three replications, the system reached 88.33% overall
accuracy versus 91.08% for human expert graders (not a statistically significant difference),
with a per-fruit processing time of 0.34 s and a throughput of about 15.45 kg/h. Accuracy
was markedly worse (56.66% machine vs. 65.33% human) at the Rotab-verging-on-Tamar
transition, where color alone does not capture softening cues; the authors recommend
touch/firmness sensing as a complement.
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 7 of 20
Belt conveyor
(0.45 m×3 m)
Paddle wheel
(4 arms)
Optical sensor 1
(triggers capture)
RGB color index
classification
Optical sensor 2
(times actuation)
Khalal
exit port
Rotab
exit port
Tamar
exit port
Step motor 1
Step motor 2 (CW)
Step motor 2 (CCW)
Figure 5.Sorting machine of Pourdarbani et al. [13].
Ibrahim et al. [12] built a low-cost prototype for Khalas-variety dates combining CIE
Lab* color analysis and projected-area size estimation (verified against computer-aided
design (CAD) measurements) via a rule-based decision tree, yielding ten color×size grades
across 100 samples (200 images) at about 0.3 s per fruit; no held-out test-set accuracy was
reported, since the system was threshold-based rather than trained.
Manickavasagan et al. [19] used RGB imaging with discriminant analysis (LDA/SDA,
Section 2) to classify date hardness, a distinct attribute from maturity or variety that is
rarely targeted directly, across three Omani varieties (Fard, Khalas, Naghal; 3300 samples,
39 color features). Accuracy ranged from about 68% (Fard, three-class) to 96% (Naghal,
two-class); roughly 10 of the 39 features proved sufficient to match the full-feature result.
Noutfia and Ropelewska [14] built purely classical models from image-texture features
(combined across all color channels) to distinguish five Moroccan date varieties, reaching
about 98% average accuracy.
Muhammad [15] decomposed each color channel into a texture descriptor (Local
Binary Patterns, LBPs, or a Weber Local Descriptor, WLD), appended shape/size fea-
tures, and classified with an SVM; the WLD outperformed LBPs, and fusing texture with
shape/size gave the best result, above 98% accuracy.
An earlier, largely pre-machine-learning study by Wulfsohn et al. [40] used spectropho-
tometric reflectance (400–1100 nm) and image thresholding to flag defective dry dates by
percentage of “defective” pixels, detecting decay and insect/bird damage well (e.g., 36.18%
defective-pixel area for Zahidi decay vs. 0.02–0.08% for good fruit) but failing to distinguish
a red-tinge/wrinkled defect from sound fruit; an early illustration of the limits of color-only
detection that later multimodal work (Section 5) addresses.
Haidar et al. [16] combined roughly fifteen hand-crafted visual features with several
classifiers across seven date categories, reporting top accuracies between 89% and 99%
depending on the category.
Mohana and Prabhakar [18] proposed a shape-and-texture fusion pipeline (bilateral
filtering, threshold segmentation, contour shape features, curvelet + LBP texture) for six-
grade classification, comparingk-NN, SVMs, and LDA;k-NN gave the best grading rate.
Koklu et al. [11] classified 898 images across seven date varieties (Barhee, Deglet Nour,
Sukkary, Rotab Mozafati, Ruthana, Safawi, Sagai) using 34 morphological, shape, and color
features with logistic regression and an artificial neural network. The two classifiers reached
91.0% and 92.2% accuracy respectively, and stacking them raised this to 92.8%.
Özaltin [17] compared four classical classifiers (decision trees, k-NN, SVMs, and an ar-
tificial neural network), each swept over multiple hyperparameters with an 80/20 train/test
split and five-fold cross-validation, on image features from seven date varieties. The neural
network (hidden layer size 25) performed best, reaching 93.85% test accuracy; the paper
does not report the underlying image count.
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 8 of 20
4. Fuzzy-Logic and Chemometric Approaches
A separate strand of date-grading research uses expert-knowledge rule systems or
chemical/physical measurement instead of image classification.
Alavi [37] built a Mamdani fuzzy inference system using three physical inputs (juice
quantity, size, freshness) with 27 expert-defined rules (3 × 3 × 3), grading 100 Mozafati
dates (Iran) into five classes, and agreed with human expert grading 86% of the time overall
(73.7–100% by class). A related study by the same author reports 91% agreement [54].
Nozari and Mazloumzadeh [9] extended this to an Adaptive Network-based Fuzzy
Inference System (ANFIS), using weight, length, width, and thickness measurements on
1000 Mozafati dates (700 train/300 test) to predict five expert-assigned grades. The ANFIS
reached 93.5% agreement with expert grading, higher than the 91% reported for the prior,
non-adaptive Mamdani system [54], and processed 1000 samples in about two minutes
versus roughly two hours for the manual method it replaced.
Kamal-Eldin and Ghnimi [49] took a chemometric approach: HPLC sugar/organic-
acid analysis, texture profile analysis, and Hunter Lab* colorimetry across 20 UAE vari-
eties (Tamr stage, 29 quantitative traits). PCA reduced these to three components explaining
52.3% of variance, and hierarchical clustering separated the 20 varieties into three groups
distinguished mainly by color, size, and malic-acid concentration. Texture, despite its
emphasis in the vision literature of Section 3, was not a significant discriminator. As a
destructive, lab-based method it is unsuited to real-time sorting, but offers a quantitative,
non-imaging basis for variety classification.
5. Deep Learning, Transfer Learning, and Multimodal Sensing
Alresheedi et al. [ 20], at the authors’ own institution (Qassim University), ran a
head-to-head comparison between classical and deep-learning features on a self-built
nine-variety dataset. Classical hand-crafted features (Color Layout, FCTH, Gabor, PHOG),
tested individually and fused across four classifiers, peaked at a single-feature accuracy
of 86.3% (PHOG + Bayesian network) and a fused F1 = 0.938 (PHOG + FCTH + MLP).
A deep-learning pipeline (AlexNet features + SVM, averaged over ten runs) reached 94.2%
accuracy, only 1–2 points above the best classical setup. The authors attribute this small
margin to the modest dataset size, cautioning that the deep-learning advantage may not
hold at larger class counts.
Nasiri et al. [55] published one of the earliest image-based deep-learning date-sorting
studies, applying a CNN to classify the processing stage of the Shahani cultivar; a specific
accuracy figure could not be confirmed for this review and is not restated here. In the same
year, Altaheri et al. [27] used transfer learning with pretrained CNNs to classify date variety
and maturity in a natural orchard environment and published the underlying 8079-image
dataset [36].
Faisal et al. [ 32] built an “Intelligent Harvesting Decision System” (IHDS) using
transfer learning from VGG-19, Inception-v3, and NASNet to classify maturity stage
on the public King Saud University dataset of Altaheri et al. [ 36] (about 8079 images,
350 bunches, 29 palms, seven stages from Immature to Tamar; 5-fold cross-validation).
VGG-19 performed best throughout: up to 99.4% accuracy for coarse two-stage classification
and 97% (97.6% F1) for the full seven-stage task, outperforming two comparison studies
the authors ran on similar data.
Rybacki et al. [23] built a custom CNN, DateNET, fusing color and geometric-shape
features for variety classification on 500 images (five Algerian varieties, Tamr stage). Color
alone reached 85.24% accuracy, geometry alone 87.62%, and the fused model 93.41% (F-
score 95.03%), matching a transfer-learned ResNet50 baseline (93.24%) and outperforming
MobileNet (91.98%) and ShuffleNet (89.62%).
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 9 of 20
Azimi and Rezaei [34] built a two-stage pipeline for Piarom dates (a premium Iranian
variety): YOLOv8-based object detection plus separate defect-classification models across
11 defect categories, on a custom 9900-image dataset (900/category before augmentation).
YOLOv8-small classification performed best at 91.95% accuracy. The separate, single-class
detection stage reached near-perfect scores (mAP50-95 up to 0.995), a substantially easier
task. Four visually similar defect classes dominated classification error; the authors suggest
that non-visual modalities, rather than more image data, may be needed to resolve them.
The dataset and code are publicly released (Section 6).
Hassan et al. [24] classified 11 Saudi date varieties from 1658 images (1868 after aug-
mentation), comparing a DenseNet121 backbone with a Squeeze-and-Excitation attention
block against YOLOv8n. The attention-augmented DenseNet121 reached 98.25% accuracy
(F1 = 97.49%), while YOLOv8n reached 96.04% accuracy but higher detection-style preci-
sion/recall (99.76%/99.70%); classification accuracy and detection metrics are therefore not
directly comparable, even within one study.
Alfarhood et al. [25] compared five architectures on one task, classifying eight date
varieties at the Rutab stage from 1659 smartphone photographs, and found performance
rose sharply with model sophistication: a plain CNN reached only 44.8% F1; ResNet-50
and EfficientNetB0 about 68–69%; YOLOv8 82.4%; and YOLOv12 93% precision, recall,
and F1 simultaneously. Because all five models were trained and tested on the same data,
this comparison is internally consistent, although it does not by itself establish a general
advantage of detection architectures.
Al-Din [26] ran the largest head-to-head comparison identified for this review: ten
pretrained CNNs, fine-tuned on a merged dataset combining two public sources (1658 Saudi
images, nine varieties; 1399 Pakistani/Saudi images, seven varieties; fifteen named cultivars
total after preprocessing and augmentation). DarkNet-50 achieved the highest accuracy
(99.33%), while MobileNetV2 and ShuffleNet offered the best accuracy-to-efficiency trade-
off for embedded deployment.
Pérez-Pérez et al. [33] compared eight CNN architectures, trained by transfer learning,
for classifying the maturity of Medjool dates, with VGG-19 performing best, and Albar-
rak et al. [28] fine-tuned a pretrained CNN for date variety classification.
Sabanci et al. [31] proposed a hybrid ResNet–ViT–SVM model, and Abuowda et al. [30]
classified date type and maturity stage for robotic harvesting (95.1% and 94.7% with ResNet-
50). Zaki et al. [ 29] compared five pretrained CNNs and used Grad-CAM to show the
image regions behind variety predictions, the first application of explainable AI to date
classification identified for this review.
A further group of studies, mostly conference papers, reuses the public seven-variety
feature dataset of Koklu et al. [11] (898 images, 34 features) to compare classical, ensemble,
and neural classifiers (e.g., [17]); as they share one dataset and differ mainly by classifier,
they are treated here as a group.
Boumaraf et al. [46] proposed a multimodal fusion approach: four RGB viewing angles,
one thermal (FLIR) image, and a weight measurement per fruit encoded into a single
composite three-channel image, classified with several CNNs. On roughly 1103 Algerian
dates (five quality grades, 70/20/10 split), VGG16 performed best (93.69% test accuracy, F1
= 93.69%, κ = 0.929); InceptionV3 overfitted substantially (100% training vs. 76.6% test),
and ResNet50 and a basic custom CNN both reached about 84.7% test accuracy. The same
group later reported feature-level fusion of these modalities [47].
Eskaf [48] proposed SmartDate, a different kind of multimodal system: high-resolution
imaging fused with Visible-Near-Infrared (VisNIR) spectral sensors to estimate moisture,
sugar content, and texture, combined with a deep-learning classifier whose hyperpa-
rameters a genetic algorithm tunes and whose behavior a reinforcement-learning (RL)
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 10 of 20
component adapts to production conditions in real time. The system reached 94.5% ac-
curacy, an F1-score of 93.1%, and an AUC-ROC of 0.96. Unlike the earlier near-infrared
studies on dates discussed below, which used the spectrum alone, it fuses spectral and
image data in one decision; however, as a six-page conference paper it does not report the
underlying dataset size or disaggregate the metrics by task.
Vikraman et al. [21] demonstrated an accessible, low-code route using Google’s Teach-
able Machine and MIT App Inventor’s Personal Image Classifier to classify three Omani
varieties, paired with a working conveyor-and-piston sorting prototype. No held-out test-
set accuracy was reported, only example per-image probabilities, so the result is best read
as a proof-of-concept for low-cost technology transfer rather than a benchmarked model.
Alaskar et al. [22] described a custom CNN trained from scratch for three date varieties
(Ekhlas, Nbute Sultan, Shayshi; 3165 images total, 1055 per variety), with a systematic
comparison of gradient-descent optimizers. Only the abstract is accessible, and it gives no
accuracy figure.
Near-infrared sensing of dates predates most image-based work. Schmilovitch et al.
determined date maturity by NIR spectrometry [43] and built a machine that sorted ‘Barhi’
dates by maturity [44]. Lee et al. [39] graded Medjool dates by size and skin delamination
with reflective NIR imaging in a commercial packing facility, improving accuracy and
reducing operating costs, and Zhang et al. [ 38] assigned Medjool dates to four surface-
quality levels in real time with short-wave infrared imaging.
Hyperspectral and NIR imaging have classified fungal-infected dates [41], detected
Aspergillus flavusinfection at 900–1700 nm [ 42], and predicted the moisture, dry matter,
and firmness of intact Khalas and Sukkari dates at 950–1750 nm [45]. Other sensors include
gas-sensor arrays for ripening stage [50], dielectric measurements for moisture and storage
quality [52,53], and X-ray imaging of internal infestation [51]. Most of these are laboratory
or single-installation demonstrations with little independent validation across varieties,
seasons, and production lines (Section 9).
6. Public Datasets and Benchmarks
At least four public date-image benchmark datasets already exist. Aiadi and Kherfi [56]
introduced a public benchmark of 5000 images across 10 date varieties alongside a Gaussian-
Mixture-Model classification method (full text paywalled; dataset existence and scale con-
firmed via the published abstract). Separately, the dataset of Altaheri et al. [ 36], used in
the IHDS study of Faisal et al. [ 32] (roughly 8079 images across seven maturity stages,
from 350 bunches on 29 palms), was published inData in Briefand released via IEEE
DataPort. More recently, Maitlo et al. [35] published, on the open Mendeley Data reposi-
tory (CC BY license), a dataset of 3004 preprocessed images (19,229 after augmentation)
spanning four date varieties (Aseel, Fasli Toto, Gajar, Kupro), each labeled by both size
(large/medium/small) and quality grade (1–3). Azimi and Rezaei’s [34] 9900-image Piarom
defect dataset is likewise publicly released, per the code and data availability statement
in that paper. None of these approaches the scale of general-purpose image-classification
benchmarks (which run to hundreds of thousands or millions of images), but the field
already has a basis to build on. Future benchmark-creation efforts should build on and
consolidate these existing resources, ideally by pooling them into a single multi-source,
multi-variety benchmark. Figure 6 puts these four public datasets alongside a sample of
the self-built (non-public) datasets discussed in Sections 3–5, on a log scale, using the pre-
augmentation counts reported by each source; the largest self-built datasets are comparable
in size to the smaller public ones.
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 11 of 20
102 103 104
Alavi 2012
Ibrahim 2014
Rybacki 2024
Koklu 2021
Boumaraf 2023
Hassan 2025
Maitlo 2024 – public
Al-Din 2025
Aiadi 2017 – public
Azimi 2024 – public
Altaheri 2019 – public
100
200
500
898
1103
1658
3004
3057
5000
9900
8079
Images/samples (log scale)
Figure 6.Dataset sizes of selected studies (log scale) [11,12,23,24,26,34–37,46,56].
7. Physical Sorting Machines and Commercial Deployment
This section covers physical machines (conveyors, actuators, sensors, and control logic)
built to sort date fruit.
The system of Pourdarbani et al. [ 13] (detailed in Section 3) is the most fully docu-
mented: a belt conveyor feeding a paddle wheel, two optical sensors, and two step motors
that physically divert each fruit to one of three exit ports by maturity stage, validated
online at 100 fruit per stage. Vikraman et al. [21] (Section 5) built a smaller-scale but still
physically working prototype, using a microcontroller-driven conveyor and pneumatic-
piston-actuated bins to sort fruit by variety in real time following a Teachable-Machine
classifier. The original system of Al-Ohali [8] (Section 3) also included a physical imaging
chamber and sorting strip, though its published description gives little detail on the me-
chanical actuation. Two earlier systems used NIR sensing (Section 5): the maturity-sorting
machine of Schmilovitch et al. [44] and the packing-facility installation of Lee et al. [ 39],
which improved grading accuracy and reduced operating costs relative to manual grading.
The clearest recent, patented example of a physical sorting machine is a 2024 patent
by Mohammed and Alqahtani [ 57], assigned to King Faisal University, Al-Ahsa, Saudi
Arabia. It describes a prototype with a feeding unit (hopper plus conveyor belt), a sensor-
based sorting unit, and an air-jet-actuated sorting mechanism directing fruit into multiple
output bins by size and at least one additional characteristic. Mechanically, its air-jet
actuation is a different approach from the step-motor/paddle-wheel design of [13] or the
pneumatic-piston design of [21], illustrating that no single actuation mechanism dominates
this small literature.
Apart from the installation reported by Lee et al. [ 39], published case studies with
cost or return-on-investment figures remain few in the peer-reviewed and patent literature
searched for this review.
Design Considerations for Feeding and Calibration
None of the systems reviewed here describe how sticky or dusty fruit is handled, so the
following principles are proposals rather than validated practice. Sticky Tamr-stage fruit
may be fed with low-adhesion food-grade surfaces, gentle vibratory feeding, and successive
belts at increasing speeds, with a vision check that recirculates clumps; dry brushing and
air cleaning before imaging would keep dust from being read as defects, and a paddle [13]
or piston [21] may release sticky fruit more reliably than an air jet [ 57]. For calibration,
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 12 of 20
an enclosed imaging chamber under filtered positive pressure, air-knife window cleaning,
and stable LED lighting would limit dust effects, and white and dark reference targets
would allow automatic correction, drift alarms, and recalibration scheduled from the
measured drift.
8. General Fruit-Sorting Context
For methodological grounding beyond date-specific work, Pandey et al. [58] reviewed
color-feature-extraction and machine-learning techniques (LDA,k-NN, SVMs, ANNs, fuzzy
rule-based systems, C4.5) across several fruits including dates, with reported accuracies
of roughly 75–96%. Jaya Vineela et al. [ 59] similarly surveyed traditional and machine-
learning/deep-learning fruit grading techniques across several fruits, including dates.
Hayat et al. [60] built a multi-fruit deep-learning grading system, “FruitVision,” evalu-
ated by 5-fold cross-validation across nine fruit types, including two date varieties alongside
apple, banana, guava, lime, orange, pomegranate, and mango; grading accuracy for Ajwa
dates reached 98.86% and for Mabroom dates 99.38%, comparable to the system’s accuracy
on the non-date fruits (97.96–99.50%). That dates graded as well as the other fruits supports
the cross-fruit transfer discussed in Section 9.
Separately, Ahmadi et al. [61] demonstrated a low-cost, non-vision, belt-and-roller
mechanical sizing machine on apples and oranges, reporting best-case sizing accuracies
of 95.28% (Red Delicious apple), 92.48% (Golden Delicious apple), and 94.28% (orange).
Although not a date-fruit study, it illustrates a non-vision-based sizing alternative that has
not been demonstrated on dates specifically, a plausible and comparatively inexpensive
avenue for size-only pre-sorting ahead of a vision-based quality stage.
9. Possibilities for Adopting Sorting Techniques from Other Fruits
Section 2 showed that visible/RGB imaging dominates the date-sorting literature,
with other sensing modalities studied mostly in the laboratory, and Section 8 noted a
broader fruit-sorting literature that this review only touches on. Several sensing and
automation techniques already mature and routinely deployed for other fruits remain
barely explored on dates, with few date-specific attempts identified. Each of the four
directions below maps onto a gap identified in the sections above.
The first is non-destructive measurement of internal quality. Sugar content and mois-
ture are core date quality parameters (Section 1.1), and most studies reviewed here classify
based on external appearance only. The chemometric HPLC analysis of Kamal-Eldin and
Ghnimi [49] (Section 4) measures internal composition directly but is destructive and un-
suited to real-time sorting. Non-destructive internal measurement has been demonstrated
on dates with hyperspectral imaging [45], with dielectric sensing [52,53], and with VisNIR
sensing fused with imaging [48] (Section 5), but mostly on laboratory samples and without
independent validation across varieties and seasons. In citrus, visible–near-infrared hyper-
spectral imaging combined with wavelength-selection and regression modeling already
predicts soluble-solid content directly from intact fruit at production scale (Kim et al. [62]:
R2 = 0.75, RMSEP = 0.559 ◦Brix, comparable to commercial sorting-machine standards);
an equivalent, independently validated inline pipeline for dates would close a real mea-
surement gap.
The second is real-time, learning-based detection integrated with physical actuation.
The date-sorting machines reviewed in Section 7 all pair a single measurement or single-
frame classifier (a color index, an NIR measurement, or a CNN evaluated on one static
image) with comparatively simple actuation: a paddle wheel, a pneumatic piston, an air
jet. In citrus sorting, Chen et al. [63] combined a lightweight CNN detector (MobileNetV2
backbone) with a SORT tracker and a Kalman filter to follow each fruit’s trajectory on a
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 13 of 20
moving line at 43 FPS (93.6% detection accuracy), designed to guide a robotic picking arm
to defective fruit in real time. The detection-plus-tracking architecture sets a precedent for
upgrading date-sorting machines beyond single-frame classification toward continuous,
per-fruit tracking on a moving line.
The third responds to a recommendation already made within the date-specific lit-
erature: both Al-Ohali [8] and Pourdarbani et al. [13] (Section 3) recommended touch or
firmness sensing as a complement to color-only grading, since color could not reliably
capture the softening transition between the Rutab and Tamr stages. Acoustic-vibration
firmness sensing, which excites a fruit and measures its resonance response, is already
validated and deployed in commercial apple packinghouses at several fruit per second per
lane. It is a mature, non-contact technology that answers this recommendation directly, yet
no date-sorting study reviewed here has applied it.
The fourth, less immediate, is X-ray imaging for internal defects that surface imaging
cannot see. Wulfsohn et al. [40] (Section 3) found that a specific red-tinge/wrinkled defect
could not be reliably distinguished from sound fruit using color and reflectance alone.
X-ray imaging, already used in other fruits to detect internal defects, insect infestation,
and pit or seed abnormalities invisible from the surface, is a plausible route to exactly the
class of defect that color-based date sorting has already been shown to miss; on dates, it
has so far been tested for internal infestation by saw-toothed beetles [51].
Some transfers are harder for dates than for apples or citrus; the following difficulties
are hypotheses, because little date-specific measurement exists. For acoustic firmness
sensing, the hard seed, any air gap around it, and the damped, viscoelastic flesh complicate
the resonance response, and the apple stiffness coefficient f 2m2/3 assumes a near-spherical,
uniform fruit, whereas an elongated, light date is expected to show bending modes at
higher frequencies; a seed-and-flesh finite element model, with non-contact laser Doppler
vibrometry, could separate these effects. For NIR, high sugar and low moisture give
overlapping absorption bands, wrinkled skin scatters light, and penetration is shallow, so
scatter correction ahead of partial least squares regression [45] is a natural starting point;
dielectric measurements depend on moisture and sugar together [52,53]. Defects would be
detected as departures from the response of sound fruit, and X-ray imaging remains the
more direct route for internal infestation [51].
10. Summary of Quantitative Results
Table 3 collects the primary quantitative result from each date-specific study discussed
above that reported a single accuracy or agreement figure; the values are not directly
comparable (see “Conditions Behind the Reported Results”). Studies with no row here
(e.g., [12,16,18,21,22,55]) reported either no formal accuracy metric, a range rather than a
single figure, or a figure not accessible beyond the abstract.
Table 3.Quantitative results, sensing hardware, and data partitioning by study (chronological).
Study Year Method Acc. Dataset Sensor and Setting Data Partition
Al-Ohali [8] 2011 BPNN, external qualityfeatures 80% 1860 samples,3 grades Color camera, imagingchamber, conveyor; lab Fruit (1200/660)
Alavi [37] 2012 Mamdani fuzzyinference 86% 100 Mozafati dates Physical measurement;lab Not applicable
Nozari and
Mazloumzadeh [9] 2013 ANFIS 93.5% 1000 Mozafatidates Weight and dimensions;lab Fruit (700/300)
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 14 of 20
Table 3.Cont.
Study Year Method Acc. Dataset Sensor and Setting Data Partition
Manickavasagan
et al. [19] 2014 RGB + LDA/SDA(hardness) 68–96% 3300 samples,3 varieties CCD color camera; lab n.r.
Muhammad [15] 2015 WLD/LBP + shape/size,SVM >98% Multiple varieties n.r. n.r.
Pourdarbani
et al. [13] 2015 RGB color index 88.3% 300 Barhee dates(online) Color camera, LEDs,conveyor; inline Fruit (onlinetest)
Faisal et al. [32] 2020 VGG-19 transfer learning(two-stage/seven-stage) 99.4/97% 8079 images,public Two DSLR cameras(4272×2848); orchard n.r. (5-fold overimages)
Koklu et al. [11] 2021 LR + ANN, stacked 92.8% 898 images,7 varieties n.r. n.r.
Alresheedi et al. [20] 2022 PHOG + FCTH + MLP/AlexNet + SVM 93.8%F1/94.2% 9 varieties,self-built n.r. n.r. (ten runs)
Noutfia and
Ropelewska [14] 2022 Classical ML, imagetexture 98% 5 Moroccanvarieties n.r. n.r.
Boumaraf et al. [46] 2023
VGG16,
RGB + thermal + weight
fusion
93.7% 1103 Algeriandates RGB (4 views), FLIRthermal, scale; lab Fruit (onesample per fruit)
Abuowda et al. [30] 2023 ResNet-50(type/maturity) 95.1/94.7% 5 types, 4 maturitystages n.r. n.r.
Rybacki et al. [23] 2024 DateNET/ResNet50(transfer) 93.4/93.2% 500 images,5 varieties n.r. n.r.
Azimi and
Rezaei [34] 2024 YOLOv8s-classification 92.0% 9900 images,11 defect classes Tray images, 50 dateseach; lab n.r. (tray datesmay span sets)
Özaltin [17] 2024 ANN (classical ML) 93.85% 7 varieties,nnotreported Feature table of [11] n.r. (80/20,5-fold)
Eskaf [48] 2024 SmartDate, deep +VisNIR fusion 94.5%nnot reported Camera + VisNIRsensors n.r.
Alfarhood et al. [25] 2025 YOLOv12/YOLOv8 93.0/82.4%F1 1659 images,8 varieties (Rutab) Smartphone camera n.r. (random70/10/20)
Hassan et al. [24] 2025 DenseNet121 +SE/YOLOv8n 98.3/96.0% 1658 images,11 varieties n.r. n.r. (split afteraugmentation)
Sabanci et al. [31] 2025 ResNet–ViT–SVM(laboratory/field) 100/83.1% 1658 images Laboratory and fieldimages Separatelaboratory andfield tests
Al-Din [26] 2025 DarkNet-50 (10-waycomparison) 99.33% 3057 images,15 varieties Merged public datasets n.r.
“Acc.” is overall reported accuracy or, where noted, agreement with human/expert grading. “n.r.”:
not reported in the sources checked.
Figure 7 plots the values in Table 3 by sorting task, with separate markers for accuracy,
F1-score, and expert agreement; detection metrics such as precision, recall, and mAP are
not plotted. The figure is descriptive and does not support a ranking or a conclusion about
improvement over time (see “Conditions Behind the Reported Results”).
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 15 of 20
2011 2014 2017 2020 2023 202660
70
80
90
100
Publication year
Reported value (%)
Variety
2011 2014 2017 2020 2023 202660
70
80
90
100
Publication year
Reported value (%)
Maturity
2011 2014 2017 2020 2023 202660
70
80
90
100
Publication year
Reported value (%)
Quality and defect grading
2011 2014 2017 2020 2023 202660
70
80
90
100
Publication year
Reported value (%)
Hardness
Accuracy F1-score Expert agreement
Figure 7.Reported performance by sorting task and metric.
Conditions Behind the Reported Results
The values in Table 3 cannot be ranked, because they were obtained under different
conditions: tasks and class numbers differ (from two to more than twenty classes), datasets
range from about 100 samples to about 10,000 images and are mostly private, and acquisi-
tion ranges from laboratory chambers to smartphones and orchards. One model can score
very differently on two versions of a task, as VGG-19 did for two and seven maturity stages
(99.4% and 97%) [32]. Hardware also limits any algorithm: surface RGB imaging cannot
measure internal sugar or moisture directly, and resolution, illumination, and background
affect how well defects are resolved. Where tested, accuracy fell from 100% on laboratory
images to 83.11% on field images [31]. Comparisons are therefore made here only between
models evaluated on the same data within one study.
Data partitioning matters as much: when images of the same fruit, bunch, or tray,
or augmented copies of one image, fall in both the training and the test sets, accuracy can be
inflated. Table 3 records the partition level where reported. Faisal et al. [32] cross-validated
8079 images from 350 bunches without stating that bunches were kept together, the samples
of Azimi and Rezaei [34] came from tray images of 50 dates each, and Hassan et al. [ 24]
split their data after augmentation. A missing report does not show that leakage occurred,
only that the risk cannot be assessed; by contrast, several earlier studies partitioned by fruit,
explicitly [18] or with one sample per fruit [8,9,46].
11. Challenges and Research Gaps
Across the studies reviewed here, several consistent limitations emerge:
• Small sample sizes.Most date-specific studies reviewed here use dataset sizes in the
hundreds to low thousands of images or samples (e.g., 100 samples in [37], 200 images
in [12], 1103 fruit in [46]), typically from a single institution or growing season.
• Narrow varietal and geographic coverage.The studies reviewed here concentrate on
Gulf (Saudi, Omani, Emirati), Iranian, and North African cultivars, with a few studies
on Medjool dates in the United States and Mexico [33,38,39]; broader coverage would
strengthen generalizability claims, and cross-variety transfer is rarely tested explicitly.
• Little field validation.Nearly all reported results come from laboratory or prototype
settings (important partial exceptions being the online-throughput test in [13] and the
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 16 of 20
packing-facility installation in [39]); field-condition validation under dust, vibration,
and variable lighting is rare in the literature reviewed here.
• Inconsistent evaluation protocols.Studies differ in accuracy definitions (overall vs.
per-class), class schemes (2 through 27 classes), and train/test splitting, and most
deep-learning studies do not report whether data were partitioned by image, fruit,
or bunch (Table 3), which makes direct cross-study comparison difficult.
• Benchmark datasets exist but remain small and scattered.As discussed in Section 6,
at least four public date-image datasets now exist [34–36,56], but none offer the scale
or multi-institution/multi-season coverage needed for robust, field-representative
benchmarking, and no study identified combines them.
• Method families are rarely compared to one another, though this is improving.
Alresheedi et al. [20] and, more recently, Rybacki et al. [23], Azimi and Rezaei [ 34],
Alfarhood et al. [ 25], and, at the largest scale so far, Al-Din [ 26] are notable excep-
tions. Fuzzy/chemometric methods (Section 4) have not been benchmarked, in the
studies identified, against either classical computer vision or deep learning on the
same dataset.
12. Future Directions
Several directions are already underway in the recent literature (Sections 5 and 7):
attention-augmented architectures [24], object-detection-based grading [25,34], large-scale
multi-architecture comparison [26], optimization-hybridized deep learning [48], and low-
cost physical prototypes [21] each have at least one published example. What remains open
is broader, as follows:
• Explainable AIremains underexplored on date classification, with at least one prelim-
inary attempt [29]; open questions are whether the image regions highlighted by such
methods match the visual cues used by human graders and whether explanations
help to detect dataset bias.
• Fuzzy/chemometric methodshave not yet been benchmarked against deep learning
on shared data in the studies identified.
• Techniques mature in other fruits are still minimally explored in dates:non-
destructive internal-quality sensing validated inline at production speed, beyond the
laboratory NIR, hyperspectral, and dielectric studies (Section 5); X-ray internal-defect
detection, with only preliminary date-specific work [51]; and, with no date-specific
attempt identified, acoustic-vibration firmness sensing and real-time detection-plus-
tracking on a moving line.
• Cross-dataset generalization remains largely untested:no study identified trains on
one existing public dataset (Section 6) and evaluates on another.
• Physical sorting machineswith published mechanical detail (Section 7) remain far
outnumbered by algorithm-only studies evaluated purely on static image sets.
12.1. Generalization to New Varieties and Conditions
Because most deep-learning results come from small, single-institution datasets, accu-
racy should be expected to fall on new varieties: a model trained on Medjool, a large, soft
date, may rely on color, size, and wrinkling that differ in Deglet Nour, a smaller, semi-dry
date. This fall has rarely been measured; accuracy dropped from 100% to 83.11% between
laboratory and field images [ 31], and a gas-sensor ripening model was validated on a
second year’s data [ 50]. A staged approach is proposed: fine-tuning on a small labeled
set from the new variety, unsupervised domain adaptation when only unlabeled images
exist, and synthetic images from generative models used for training only, with leave-one-
variety-out and leave-one-season-out tests to measure robustness.
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 17 of 20
12.2. Multimodal Fusion and Line Speed
Early fusion combines raw data and needs accurate sensor alignment; late fusion
combines separate decisions, is the simplest, and tolerates a failed sensor; intermediate
fusion combines learned features and can capture relations between sensors, but usually
at the highest cost, especially with transformers [64,65]. On dates, a composite image [46]
and feature-level fusion [47] have been reported, but no comparison of fusion levels on the
same data was identified, so late or lightweight intermediate fusion is the recommended
starting point.
The time available for each reading is the sensing-zone length divided by the belt
speed; a 0.1 m zone at 0.5 m/s leaves 0.2 s per fruit, comparable to the 0.34 s per fruit
of Pourdarbani et al. [13]. Faster belts capture fewer thermal frames per fruit, with more
motion blur from slow uncooled cameras, and leave less time to excite and record an
acoustic response; triggering all sensors from a shared belt encoder matches each reading
to the correct fruit.
12.3. A Benchmarking Protocol Linked to Industrial Indicators
A fair comparison of classical, chemometric, and deep-learning methods needs shared
multi-site, multi-season, multi-variety data with physical and chemical reference measure-
ments; fruit- or bunch-level partitioning; separate unseen-variety and unseen-season test
sets; grading by several graders; and common reporting of per-class metrics, confidence
intervals, latency, and hardware. Assigning a monetary cost to each cell of the confusion
matrix links these metrics to industrial indicators. The share of exportable fruit rejected
gives the false-rejection rate and the premium lost, and accepted defective fruit provides the
risk of a rejected shipment. Combined with grade mix, throughput, labor saved, and equip-
ment cost, these costs give the return on investment, so two systems with equal F1-scores
may differ greatly in commercial value.
13. Conclusions
A substantial literature on automated date-fruit sorting exists, spanning classical
computer vision, fuzzy-logic and chemometric grading, deep learning, and spectral and
multimodal sensing. Reported accuracies span roughly 68% to above 99%, but they were
obtained on different tasks, datasets, and protocols and do not support a ranking of
methods. The taxonomy (Section 2) shows a field with many algorithms but narrow sensing,
as most studies use RGB imaging, and physical sorting machines remain a small part of
the literature (Section 7). Validated inline internal-quality sensing, real-time detection and
tracking, acoustic firmness sensing, and X-ray defect detection remain open opportunities
(Section 9), as do large shared benchmarks, field validation, published deployment data,
and head-to-head comparison of method families. This synthesis covers a substantial part
of the literature without claiming to be exhaustive.
Author Contributions:Conceptualization, H.M.O.; methodology, H.M.O.; validation, H.M.O. and
S.M.S.M.; investigation, H.M.O.; resources, H.M.O. and S.M.S.M.; writing—original draft preparation,
H.M.O.; writing—review and editing, H.M.O. and S.M.S.M.; supervision, S.M.S.M. All authors have
read and agreed to the published version of the manuscript.
Funding:This research received no external funding.
Institutional Review Board Statement:Not applicable.
Informed Consent Statement:Not applicable.
Data Availability Statement:No new data were created in this study. All information discussed is
drawn from the cited publications.
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 18 of 20
Acknowledgments:The Researchers would like to thank the Deanship of Graduate Studies and
Scientific Research at Qassim University (https://www.qu.edu.sa) for financial support (QU-APC-
2026).
Conflicts of Interest:The authors declare no conflicts of interest.
References

  1. Zohary, D.; Hopf, M.Domestication of Plants in the Old World, 3rd ed.; Oxford University Press: Oxford, UK, 2000.
  2. Chao, C.T.; Krueger, R.R. The date palm (Phoenix dactyliferaL.): Overview of biology, uses, and cultivation.HortScience2007,42,
    1077–1082. https://doi.org/10.21273/hortsci.42.5.1077.
  3. FAO.FAOSTAT: Crops and Livestock Products; Food and Agriculture Organization of the United Nations: Rome, Italy, 2026.
    Available online: https://www.fao.org/faostat/en/#data/QCL (accessed on 4 November 2026).
  4. Keskin, B.; Uzundumlu, A.S.; Karabacak, T. Production estimates of leading countries in date production for the period 2023–2028.
    Appl. Fruit Sci.2025,67, 83. https://doi.org/10.1007/s10341-025-01307-9.
  5. Awad, M.A. Increasing the rate of ripening of date palm fruit (Phoenix dactyliferaL.) cv. Helali by preharvest and postharvest
    treatments.Postharvest Biol. Technol.2007,43, 121–127. https://doi.org/10.1016/j.postharvbio.2006.08.006.
  6. Haider, M.S.; Khan, I.A.; Naqvi, S.A.; Jaskani, M.J.; Khan, R.W.; Nafees, M.; Pasha, I. Fruit developmental stages effects on
    biochemical attributes in date palm.Pak. J. Agric. Sci.2013,50, 577–583.
  7. Wikimedia Commons. Date Fruit Photographs (Kimri: M. Toorani, CC BY-SA 4.0; Others Public Domain). Available online:
    https://commons.wikimedia.org (accessed on 8 November 2026).
  8. Al-Ohali, Y. Computer vision based date fruit grading system: Design and implementation.J. King Saud Univ.—Comput. Inf. Sci.
    2011,23, 29–36. https://doi.org/10.1016/j.jksuci.2010.03.003.
  9. Nozari, V .; Mazloumzadeh, S.M. Date fruits grading based on some physical properties.J. Agric. Technol.2013,9, 1703–1713.
  10. Noutfia, Y.; Ropelewska, E. What can artificial intelligence approaches bring to an improved and efficient harvesting
    and postharvest handling of date fruit (Phoenix dactyliferaL.)? A review.Postharvest Biol. Technol.2024,213, 112926.
    https://doi.org/10.1016/j.postharvbio.2024.112926.
  11. Koklu, M.; Kursun, R.; Taspinar, Y.S.; Cinar, I. Classification of date fruits into genetic varieties using image analysis.Math. Probl.
    Eng.2021,2021, 4793293. https://doi.org/10.1155/2021/4793293.
  12. Ibrahim, A.A.A.; Amer Eissa, A.H.; Alghannam, A.R.O. Image processing system for automated classification date fruit.Int. J.
    Adv. Res.2014,2, 702–715.
  13. Pourdarbani, R.; Ghassemzadeh, H.R.; Seyedarabi, H.; Nahandi, F.Z.; Vahed, M.M. Study on an automatic sorting system for date
    fruits.J. Saudi Soc. Agric. Sci.2015,14, 83–90. https://doi.org/10.1016/j.jssas.2013.08.006.
  14. Noutfia, Y.; Ropelewska, E. Innovative models built based on image textures using traditional machine learning algo-
    rithms for distinguishing different varieties of Moroccan date palm fruit (Phoenix dactyliferaL.).Agriculture2022,13, 26.
    https://doi.org/10.3390/agriculture13010026.
  15. Muhammad, G. Date fruits classification using texture descriptors and shape-size features.Eng. Appl. Artif. Intell.2015,37,
    361–367. https://doi.org/10.1016/j.engappai.2014.10.001.
  16. Haidar, A.; Dong, H.; Mavridis, N. Image-based date fruit classification. InProceedings of the 2012 IV International Congress on
    Ultra Modern Telecommunications and Control Systems (ICUMT), St. Petersburg, Russia, 3–5 October 2012; IEEE: New York, NY, USA,
    2013; pp. 357–363. https://doi.org/10.1109/ICUMT.2012.6459693.
  17. Özaltin, Ö. Date fruit classification by using image features based on machine learning algorithms.Res. Agric. Sci.2024,55, 26–35.
    https://doi.org/10.5152/AUAF.2024.23171.
  18. Mohana, S.H.; Prabhakar, C.J. A novel technique for grading of dates using shape and texture features.Mach. Learn. Appl. Int. J.
    2014,1, 15–29. https://doi.org/10.5121/mlaij.2014.1202.
  19. Manickavasagan, A.; Al-Mezeini, N.K.; Al-Shekaili, H.N. RGB color imaging technique for grading of dates.Sci. Hortic.2014,175,
    87–94. https://doi.org/10.1016/j.scienta.2014.06.003.
  20. Alresheedi, K.M.; Aladhadh, S.; Khan, R.U.; Qamar, A.M. Dates fruit recognition: From classical fusion to deep learning.Comput.
    Syst. Sci. Eng.2022,40, 151–166. https://doi.org/10.32604/csse.2022.017931.
  21. Vikraman, B.P .; Mahadevan, V .; Al Hashmi, D.; Al Harrasi, A. Date fruit classification and sorting system using Artificial
    Intelligence: Application of transfer learning.Res Mil.2022,12, 1151–1160.
  22. Alaskar, H.; Alhewaidi, S.; Obaid, B.; Alzahrani, G.; Abdulahi, A.; Sbai, Z.; Vaiyapuri, T. Dates fruit classification using convolution
    neural networks. InProceedings of the Sixth International Congress on Information and Communication Technology (ICICT 2021); Lecture
    Notes in Networks and Systems; Springer: Singapore, 2021; Volume 216, pp. 757–775. https://doi.org/10.1007/978-981-16-1781-
    2_66.
    https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 19 of 20
23. Rybacki, P .; Niemann, J.; Derouiche, S.; Chetehouna, S.; Boulaares, I.; Seghir, N.M.; Diatta, J.; Osuch, A. Convolutional Neural
Network (CNN) model for the classification of varieties of date palm fruits (Phoenix dactyliferaL.).Sensors2024,24, 558.
https://doi.org/10.3390/s24020558.
24. Hassan, E.; Abu Ghazalah, S.; El-Rashidy, N.; Abd El-Hafeez, T.; Shams, M.Y. Sustainable deep vision systems for
date fruit quality assessment using attention-enhanced deep learning models.Front. Plant Sci.2025,16, 1521508.
https://doi.org/10.3389/fpls.2025.1521508.
25. Alfarhood, M.; Alsahw, N.; Almajed, M.; Alzahrani, M.; Alawfi, A.; Alanazi, M.; Alalwan, A. A machine learning approach for
classifying date fruit varieties at the Rutab stage.Front. Plant Sci.2025,16, 1678757. https://doi.org/10.3389/fpls.2025.1678757.
26. Al-Din, M.S.N. Date fruit varieties identification using convolutional neural networks: A comparative study.J. Soft Comput. Data
Min.2025,6, 103–117. https://doi.org/10.30880/jscdm.2025.06.03.008.
27. Altaheri, H.; Alsulaiman, M.; Muhammad, G. Date Fruit Classification for Robotic Harvesting in a Natural Environment Using
Deep Learning.IEEE Access2019,7, 117115–117133. https://doi.org/10.1109/ACCESS.2019.2936536.
28. Albarrak, K.; Gulzar, Y.; Hamid, Y.; Mehmood, A.; Soomro, A.B. A Deep Learning-Based Model for Date Fruit Classification.
Sustainability2022,14, 6339. https://doi.org/10.3390/su14106339.
29. Zaki, N.; Singh, H.; Krishnan, A.; Alnaqbi, A.; Alneyadi, S.; Alnaqbi, S.; Alhindaassi, S.; Alam, M.; Eldin, A.K. Transfer Learning
and Explainable Artificial Intelligence Enhance the Classification of Date Fruit Varieties. InProceedings of the 2023 15th International
Conference on Innovations in Information Technology (IIT), Al Ain, United Arab Emirates, 14–15 November 2023; IEEE: Piscataway,
NJ, USA, 2023; pp. 222–227. https://doi.org/10.1109/IIT59782.2023.10366495.
30. Abuowda, Z.; Ramadan, S.; Salam, N.; Gad, A.; Yousaf, J.; Hassan, T.; Ghazal, M.; Almajali, E. Date Fruit Classifi-
cation System using Deep Transfer Learning. InProceedings of the 2023 10th International Conference on Future Internet
of Things and Cloud (FiCloud), Marrakesh, Morocco, 14–16 August 2023; IEEE: Piscataway, NJ, USA, 2023; pp. 393–398.
https://doi.org/10.1109/FiCloud58648.2023.00063.
31. Sabanci, K.; Aslan, M.F.; Aslan, B. ResNet-ViT-SVM: A new hybrid architecture proposal and experimental comparisons on date
fruit.J. Food Compos. Anal.2025,148, 108519. https://doi.org/10.1016/j.jfca.2025.108519.
32. Faisal, M.; Alsulaiman, M.; Arafah, M.; Mekhtiche, M.A. IHDS: Intelligent harvesting decision system for date
fruit based on maturity stage using deep learning and computer vision.IEEE Access2020,8, 167985–167997.
https://doi.org/10.1109/ACCESS.2020.3023894.
33. Pérez-Pérez, B.D.; García Vázquez, J.P .; Salomón-Torres, R. Evaluation of convolutional neural networks’ hy-
perparameters with transfer learning to determine sorting of ripe Medjool dates.Agriculture2021,11, 115.
https://doi.org/10.3390/agriculture11020115.
34. Azimi, N.; Rezaei, D.M. Automated defect detection and grading of Piarom dates using deep learning.arXiv2024,
arXiv:2410.18208. https://doi.org/10.48550/arXiv.2410.18208.
35. Maitlo, A.K.; Shaikh, R.A.; Arain, R.H. A novel dataset of date fruit for inspection and classification.Data Brief2024,52, 110026.
https://doi.org/10.1016/j.dib.2023.110026.
36. Altaheri, H.; Alsulaiman, M.; Muhammad, G.; Amin, S.U.; Bencherif, M.; Mekhtiche, M. Date fruit dataset for intelligent
harvesting.Data Brief2019,26, 104514. https://doi.org/10.1016/j.dib.2019.104514.
37. Alavi, N. Date grading using rule-based fuzzy inference system.J. Agric. Technol.2012,8, 1243–1254.
38. Zhang, D.; Lee, D.-J.; Desai, A. Using short-wave infrared imaging for fruit quality evaluation. InProceedings of the SPIE 9025,
Intelligent Robots and Computer Vision XXXI: Algorithms and Techniques, San Francisco, CA, USA, 3 February 2014; SPIE: Bellingham,
WA, USA, 2014; p. 90250B. https://doi.org/10.1117/12.2045406.
39. Lee, D.-J.; Schoenberger, R.; Archibald, J.; McCollum, S. Development of a machine vision system for automatic date grading
using digital reflective near-infrared imaging.J. Food Eng.2008,86, 388–398. https://doi.org/10.1016/j.jfoodeng.2007.10.021.
40. Wulfsohn, D.; Sarig, Y.; Algazi, R.V . Defect sorting of dry dates by image analysis.Can. Agric. Eng.1993,35, 133–139.
41. Teena, M.A.; Manickavasagan, A.; Ravikanth, L.; Jayas, D.S. Near infrared (NIR) hyperspectral imaging to classify fungal infected
date fruits.J. Stored Prod. Res.2014,59, 306–313. https://doi.org/10.1016/j.jspr.2014.09.005.
42. Teena, M.; Manickavasagan, A.; Al-Sadi, A.M.; Al-Yahyai, R.; Deadman, M.L.; Al-Ismaili, A. Near infrared imag-
ing to detectAspergillus flavusinfection in three varieties of dates.Eng. Agric. Environ. Food2018,11, 169–177.
https://doi.org/10.1016/j.eaef.2018.04.002.
43. Schmilovitch, Z.; Hoffman, A.; Egozi, H.; Ben-Zvi, R.; Bernstein, Z.; Alchanatis, V . Maturity determination of fresh dates by
near infrared spectrometry.J. Sci. Food Agric.1999,79, 86–90. https://doi.org/10.1002/(SICI)1097-0010(199901)79:1<86::AID-
JSFA175>3.0.CO;2-R.
44. Schmilovitch, Z.; Hoffman, A.; Egozi, H.; Ben-Zvi, R. Machine for automatic sorting ’Barhi’ dates according to maturity by near
infrared spectrometry.Acta Hortic.2001,553, 481–485. https://doi.org/10.17660/actahortic.2001.553.113.
https://doi.org/10.3390/agriculture16222532

Agriculture2026,16, 2532 20 of 20
45. Ibrahim, A.; Alghannam, A.; Eissa, A.; Firtha, F.; Kaszab, T.; Kovacs, Z.; Helyes, L. Preliminary Study for Inspecting Moisture
Content, Dry Matter Content, and Firmness Parameters of Two Date Cultivars Using an NIR Hyperspectral Imaging System.
Front. Bioeng. Biotechnol.2021,9, 720630. https://doi.org/10.3389/fbioe.2021.720630.
46. Boumaraf, I.; Djeffal, A.; Setta, S.; Taleb-Ahmed, A. Improving date fruit sorting with a novel multimodal approach and CNNs.
Int. J. Adv. Soft Comput. Its Appl.2023,15, 190–206. https://doi.org/10.15849/IJASCA.231130.13.
47. Boumaraf, I.; Djeffal, A.; Taleb-Ahmed, A. Feature-level multimodal deep learning for date fruit quality assessment.J. Food Meas.
Charact.2026,20, 8353–8371. https://doi.org/10.1007/s11694-026-04193-1.
48. Eskaf, K. SmartDate: AI-driven precision sorting and quality control in date fruits. InProceedings of the 2024 21st International
Conference on High Performance Computing and Networking (HONET), Doha, Qatar, 1–3 December 2024; IEEE: New York, NY, USA,
2024. https://doi.org/10.1109/HONET63146.2024.10822938.
49. Kamal-Eldin, A.; Ghnimi, S. Classification of date fruit (Phoenix dactylifera, L.) based on chemometric analysis with multivariate
approach.J. Food Meas. Charact.2018,12, 1020–1027. https://doi.org/10.1007/s11694-018-9717-4.
50. Savand-Roumi, E.; Keramat-Jahromi, M. Optimizing gas sensor array to classify three edible ripening stages of date
fruit using machine learning and feature selection (case study: Cv. Shahani).Span. J. Agric. Res.2025,23, 21120.
https://doi.org/10.5424/sjar/2025232-21120.
51. Al-Mezeini, N.; Manickavasagan, A.; Al-Yahyai, R.; Al-Wahaibi, A.K.; Al-Raeesi, A.A.; Khriji, L. X-ray Imaging of Stored Dates to
Detect Infestation by Saw-Toothed Beetles.Int. J. Fruit Sci.2016,16, 42–56. https://doi.org/10.1080/15538362.2015.1044692.
52. Mohammed, M.; Munir, M.; Aljabr, A. Prediction of Date Fruit Quality Attributes during Cold Storage Based on Their Electrical
Properties Using Artificial Neural Networks Models.Foods2022,11, 1666. https://doi.org/10.3390/foods11111666.
53. Karimi, H. Intelligent dielectric method for evaluating some qualitative characteristics of date fruit.Postharvest Biol. Technol.2025,
219, 113195. https://doi.org/10.1016/j.postharvbio.2024.113195.
54. Alavi, N. Quality determination of Mozafati dates using Mamdani fuzzy inference system.J. Saudi Soc. Agric. Sci.2013,12,
137–142. https://doi.org/10.1016/j.jssas.2012.10.001.
55. Nasiri, A.; Taheri-Garavand, A.; Zhang, Y.D. Image-based deep learning automated sorting of date fruit.Postharvest Biol. Technol.
2019,153, 133–141. https://doi.org/10.1016/j.postharvbio.2019.04.003.
56. Aiadi, O.; Kherfi, M.L. A new method for automatic date fruit classification.Int. J. Comput. Vis. Robot.2017,7, 692.
https://doi.org/10.1504/IJCVR.2017.087751.
57. Mohammed, M.E.A.; Alqahtani, N.K. Intelligent Sorting for Date Palm Fruit. U.S. Patent 11,919,043, 5 March 2024.
58. Pandey, R.; Naik, S.; Marfatia, R. Image processing and machine learning for automated fruit grading system: A technical review.
Int. J. Comput. Appl.2013,81, 29–39. https://doi.org/10.5120/14209-2455.
59. Jaya Vineela, P .; Sai Manvitha, M.; Rishitha, P .; Prabu, U. A comprehensive study on fruit classification and grading techniques.
InProceedings of the 2022 3rd International Conference on Electronics and Sustainable Communication Systems (ICESC), Coimbatore, India,
17–19 August 2022; IEEE: New York, NY, USA, 2022; pp. 970–978. https://doi.org/10.1109/ICESC54411.2022.9885492.
60. Hayat, A.; Morgado-Dias, F.; Choudhury, T.; Singh, T.P .; Kotecha, K. FruitVision: A deep learning based automatic fruit grading
system.Open Agric.2024,9, 20220276. https://doi.org/10.1515/opag-2022-0276.
61. Ahmadi, V .F.; Ziyaee, P .; Bazyar, P .; Cavallo, E. Development and testing of a low-cost belt-and-roller machine for spheroid fruit
sorting.AgriEngineering2020,2, 596–606. https://doi.org/10.3390/agriengineering2040040.
62. Kim, M.J.; Yu, W.H.; Song, D.J.; Chun, S.W.; Kim, M.S.; Lee, A.; Kim, G.; Shin, B.S.; Mo, C. Prediction of soluble-solid content in
citrus fruit using visible–near-infrared hyperspectral imaging based on effective-wavelength selection algorithm.Sensors2024,24,
1512. https://doi.org/10.3390/s24051512.
63. Chen, Y.; An, X.; Gao, S.; Li, S.; Kang, H. A deep learning-based vision system combining detection and tracking for fast on-line
citrus sorting.Front. Plant Sci.2021,12, 622062. https://doi.org/10.3389/fpls.2021.622062.
64. Baltrusaitis, T.; Ahuja, C.; Morency, L.-P . Multimodal Machine Learning: A Survey and Taxonomy.IEEE Trans. Pattern Anal. Mach.
Intell.2019,41, 423–443. https://doi.org/10.1109/tpami.2018.2798607.
65. Ramachandram, D.; Taylor, G.W. Deep Multimodal Learning: A Survey on Recent Advances and Trends.IEEE Signal Process.
Mag.2017,34, 96–108. https://doi.org/10.1109/msp.2017.2738401.
Disclaimer/Publisher’s Note:The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
https://doi.org/10.3390/agriculture16222532

论文原文与出版信息

MDPI Agriculture
查看论文原文

继续探索