Food and Bioprocess Technology (2025) 18:183–206

[Link]

REVIEW

Fresh Fruit Bunch Ripeness Classification Methods: A Review

Jin Yu Goh1 · Yusri Md Yunos1 · Mohamed Sultan Mohamed Ali1,2

Received: 1 November 2023 / Accepted: 12 June 2024 / Published online: 28 June 2024
© The Author(s) 2024

Abstract
The escalating demand for palm oil necessitates enhanced production strategies. As the trend shifts towards automated
harvesting to meet the demand, precise ripeness classification has become pivotal. Manual methods are inefficient and error-
prone because of workforce constraints. The present review scrutinizes the following non-destructive ripeness classification
methods: spectroscopy, inductive sensing, thermal imaging, light detection and ranging, laser-light backscattering imaging,
and computer vision. The review focuses on identifying reliable techniques capable of real-time and accurate classification
in dynamic and unstructured environments. All aforementioned techniques are discussed in intricate detail, accompanied
by thorough critiques. This review then presents a performance comparison and benchmarking process, providing com-
prehensive insights into the strengths and weaknesses of each technique. A compelling solution emerges in the fusion of
light detection and ranging and computer vision techniques. This synergy capitalizes on their strengths to offset individual
limitations, offering a potent approach. Furthermore, this fusion yields added value in terms of localization and mapping,
rendering it exceptionally suitable for real-time classification in complex environments. This review provides insights into
bridging the gap between automated harvesting needs and ripeness assessment precision, thereby fostering advancements
in the palm oil industry.

Keywords Computer vision in agriculture · Agriculture automation · Non-destructive assessment · Oil palm technology

Introduction Harvesting FFB from an oil palm tree requires experi-

enced workers because FFB can weigh 20–25 kg and must
The oil palm industry exerts a notable global impact, play- be cut from heights of up to 20 feet (Malaysian Palm Oil
ing a vital role in the production of various edible and non- Council, n.d.). Workers are at high risk of injury during this
edible consumer goods. Thus, global demand for fresh fruit process due to the force exerted to cut the FFB with a sickle
bunches (FFB) has increased significantly, accounting for and the impact threat when the FFB drops down (Syazwani
approximately 35% of global vegetable oil consumption et al., 2016). Likewise, the conventional harvesting process
(United States Department of Agriculture, 2023a). In 2022, is labor-intensive and employs approximately 8 million indi-
there was a record-breaking global production of 78.14 mil- viduals in Indonesia (Indonesian Palm Oil Association, n.d.).
lion metric tons of FFB (United States Department of Agri- The integration of innovative technology for autonomous
culture, 2023b), with Indonesia and Malaysia accounting for harvesting operations has emerged as a critical imperative
83.18% of that total. Because of FFB’s versatility in various for the field and mitigating potential human errors associated
industries, as well as increasing global demand due to popu- with decision-making tasks that directly affect productivity.
lation growth and shifting dietary patterns, efficient harvest- The first stage in implementing an autonomous FFB harvest-
ing of FFB is gaining attention to cope with production rates. ing system is to develop reliable and accurate methods for
recognizing and classifying FFB ripeness, thereby addressing
the existing difficulties in ripeness classification. The conven-
* Mohamed Sultan Mohamed Ali tional approach to determining the ripeness of FFB while still
[Link]@[Link]; sultan_ali@[Link]
on the tree involves visual examination by knowledgeable
1
Faculty of Electrical Engineering, Universiti Teknologi graders, which entails inspecting and classifying the FFB
Malaysia, Johor Bahru, 81310 UTM Johor, Malaysia based on their color and the number of loose fruitlets on the
2
Department of Electrical Engineering, College ground adjacent to the tree. The quality index identification
of Engineering, Qatar University, Doha, Qatar

Vol.:(0123456789)
184 Food and Bioprocess Technology (2025) 18:183–206

of FFB is verified based on shape, texture, and color, which Despite a considerable body of research in this domain, a
can be inefficient and inconsistent because of the limitations notable paucity exists in comprehensive review articles that
of human observation (Shabdin et al., 2016). Conventional explore the diverse aspects of the oil palm sector within the
FFB ripeness classifications in the palm oil industry are inad- framework of ripeness classification. Unlike previous review
equate because of a lack of standardization, and the potential articles, the present rigorous systematic review ensures a
inaccuracy of these classifications hinders optimal harvesting methodical meta-analysis that was evaluated by enabling
practices and overall efficiency of the oil extraction rate. qualitative and quantitative assessment benchmarks of the
Adopting ripeness classification technology is crucial for compiled literature. The present review focuses on notable
the oil palm industry’s long-term growth, and the methods advancements, prevailing trends, and pioneering concepts,
should be able to operate in a real-time outdoor dynamic and thereby stimulating further research in the oil palm domain
unstructured environment. There are two types of methods by using innovative ripeness classification techniques. A
for ripeness classification: destructive and non-destructive structured approach was used to select relevant sources for a
methods. Researchers have commonly used various tech- comprehensive and informative review. Priority was given to
niques under non-destructive methods, such as spectroscopy, journal articles published in highly regarded journals within
inductive sensing, thermal imaging, light detection and rang- the fields of FFB ripeness classification. However, there are
ing (LiDAR), laser-light backscattering imaging (LLBI), and limited articles because of the geographically concentrated
computer vision, for classification purposes. Because of the research effort on FFB ripeness classification in Southeast
importance of this research field, several efforts have been Asia. To address this gap and capture the latest advance-
made to review the use of ripeness classification techniques ments, some conference proceedings were incorporated in
in the context of oil palm. For instance, Lai et al. (2023) this review. These conference materials provide valuable
and Yee et al. (2020) reviewed FFB ripeness detection tech- insights into the potential of novel technologies for FFB
niques, providing valuable current insight into both destruc- ripeness assessment. Ultimately, this review offers high-
tive and non-destructive methods. However, the authors did quality recommendations for researchers, backed by detailed
not evaluate the collected results with quantitative assess- evaluation of the existing literature, thereby propelling the
ments, which are essential for determining the performance field of oil palm research into new frontiers.
metrics of the reported techniques. Additionally, Mohamed
Ahmed Mansour et al. (2022) specifically reviewed non-
destructive methods for FFB ripeness classification and Characteristics of FFB
evaluated various classification methods with comparative
performance indexes. However, the authors provided a gen- The assessment of FFB is primarily influenced by the exter-
eral overview of the non-destructive methods without delv- nal characteristics of the fruitlet and its chemical compo-
ing into each technique’s specific technical details and real- sition, with key elements in the fruitlet being moisture,
world scenarios, further limiting the reader’s understanding lipids, and fiber. Figure 1 illustrates the internal structure
of the methodologies. and content of the ripe and unripe FFB fruitlets based on

Fig. 1  A representation of an

oil palm tree (Elaeis Guineen-
sis) followed by the FFB and
fruitlets content for both ripe
and unripe conditions
Food and Bioprocess Technology (2025) 18:183–206 185

the report by Japan’s Food Technology Department (Harun sensing, thermal imaging, LiDAR, LLBI, and computer
et al., 2013). In differentiating the internal content of fruit- vision. Accuracy and coefficient of determination (R2) are
lets, > 58% of the ripe fruitlet is made up of lipids (oil), crucial performance metrics used to evaluate the effective-
while > 80% of the unripe fruitlet is made up of moisture ness of the aforementioned techniques used for the ripeness
(water); these elements provide important criteria for FFB classification. Accuracy measures the closeness of sensor
classification. The exocarp is the outer skin or rind of FFB measurements to actual ripeness, signifying the reliability
fruitlets, and the mesocarp is the fleshy middle layer beneath of the sensor’s output. In spectroscopy, thermal imaging,
the exocarp. The fruitlets ripen within an FFB, so the exo- inductive sensing, LiDAR, and LLBI, accuracy ensures
carp and mesocarp undergo noticeable changes in color; the correct identification of components, precise distance
these changes in color serve as visible indicators for assess- measurements, accurate temperature readings, and reliable
ing the ripeness (Fig. 1). Furthermore, the color of the exo- predictions of FFB properties. However, R2 evaluates the
carp’s surface is influenced by chlorophyll and triacylglyc- goodness of fit of regression models that establish rela-
erols in the mesocarp. As the fruitlets ripen, triacylglycerols tionships between sensor data and real-world quantities. A
accumulate as oil components, replacing chlorophyll in the higher R2 value indicates a stronger correlation, suggesting
mesocarp and changing the color from black or dark purple a better fit for the model and higher accuracy of the sensor
in the unripe stage to dark red or orange in the ripe stage data. Both accuracy and R2 play pivotal roles in ensuring the
(Harun et al., 2013). Several factors can influence oil quality, effectiveness and dependability of the ripeness classification
including oil content and free fatty acids (FFAs) produced techniques to meet specific application requirements. This
by glyceride hydrolysis in the presence of water (Makky review provides useful insights into robust and reliable non-
& Soni, 2014). The extraction of crude palm oil involves destructive FFB ripeness classification methods by exploring
obtaining the mesocarp (Basyuni et al., 2017), which pri- the latest advancements and findings.
marily consists of glycerides with minor quantities of non-
glyceride components, including FFAs, moisture, trace met- Spectroscopy
als, and impurities (Mba et al., 2015).
Table 1 shows the grading standard used by FFB asses- In the context of FFB ripeness classification, spectroscopy
sors globally. The standard comprises the following four cri- is used to determine the chemical and physical properties of
teria to determine the ripeness of oil palm FFBs: the number the fruitlets by measuring the wavelengths and intensities
of loose fruit sockets, percentages of the fruitlets that are of electromagnetic radiation absorbed, emitted, or diffused.
still attached to the FFB, mesocarp color, and exocarp color. Various molecules and elements across multiple ripeness
Besides the fruitlet color, which is the primary indicator of stages have distinctive spectral signatures that can be used
ripeness, the quantity of loose fruitlet sockets improves the to identify and analyze them. When FFB is exposed to elec-
classification accuracy. tromagnetic radiation, it absorbs or emits energy at specific
wavelengths that correspond to the energies required for
the FFB’s electronic, vibrational, or rotational transitions.
Ripeness Classification Techniques Researchers have determined the FFB’s chemical composi-
tion (lipid and moisture content), molecular structure (tri-
To achieve accurate ripeness classification of FFB, research- glycerides and fatty acids), and physical properties (bulk
ers have explored various techniques that leverage advanced density and oil viscosity) by measuring the wavelengths
technology and data analysis. The present review focuses and intensities of the absorbed or emitted radiation (Pavia
on non-destructive methods that classify the maturity of et al., 2013) (Fig. 2). Spectroscopy has numerous applica-
FFB without damaging or destroying it by extracting the tions in scientific research and industry; the present review
external features for further study. The review includes focuses on near-infrared spectroscopy (NIRS), hyperspectral
the following in-depth techniques: spectroscopy, inductive imaging, Raman spectroscopy, fluorescence spectroscopy,

Table 1  Fresh Fruit Bunch Ripeness level Characteristics

ripeness grading standards
(Utom et al., 2018) Number of Percentages of Mesocarp color Exocarp color
loose sockets attached fruitlets

Unripe <1 95–100% Yellow Yellowish purple black

Under-ripe < 10 > 90% Yellowish Orange Reddish purple
Ripe >  = 10 > 50% Orange Reddish yellow
Over-ripe > 10 10–50% Orange Reddish orange
186 Food and Bioprocess Technology (2025) 18:183–206

Fig. 2  A working illustration

of a spectroscopy instrument
which has a probe to emit light
with various wavelengths to
determine the absorbance from
the reflected light. The spectros-
copy instrument illustrated is
general equipment that can be
replaced with several types of
spectroscopies

and microwave spectroscopy. These techniques were con- classification is made possible by comparing the absorbance
ducted by researchers, particularly in oil palm FFB ripeness of NIRS spectra (Fig. 2), which shows that the ripe FFB has
classification. a higher absorbance than the unripe FFB. Iqbal et al. (2019)
A notable constraint pertains to the additional variables developed a calibration model based on the NIRS approach
originating from the external surroundings, such as ambient using chemometrics to predict the water content and oil con-
illumination, underlying fluorescence, and electromagnetic tent of the oil palm FFB. The calibration model uses partial
interference, which present substantial obstacles and may least square (PLS) regression to predict the water content
result in imprecise or unreliable measurements. It is essential using multiplicative scatter correction preprocessing; how-
to thoroughly evaluate and decrease the effect of these inter- ever, the oil content calibration statistic parameter did not
ferences to guarantee the precision of fruit ripeness evalua- meet the requirements that restrict the application of PLS
tion. Furthermore, spectroscopic systems are impractical for on the oil content approach. The genetic algorithm neural
outdoor applications because of their lack of portability and network model (GANN) (Silalahi et al., 2016a, 2016b) was
dependence on a stable power source. used to process the NIRS spectral data for FFB ripeness
classification with the result of dimensional reduction from
Near Infrared Spectroscopy original NIRS spectral data using principal component anal-
ysis (PCA) validation scores of 84.5% for good classification
NIRS is a non-destructive analytical technique that involves accuracy. NIRS is conducted in the laboratory and requires
emitting near-infrared light onto an FFB and measuring how specialized equipment; heavy assets and lack of portabil-
much light is absorbed or reflected. The resulting spectra ity restrict the data-acquiring process in dynamic outdoor
can be used to identify and quantify chemical compounds environments, causing major drawbacks to the feasibility of
in the fruitlet. The near-infrared region of the electromag- implementation in dynamic outdoor environments.
netic spectrum is typically defined as wavelengths ranging
from 700 to 2500 nm. This region is useful for spectroscopy Hyperspectral Imaging
because it contains high molecular vibrations that are unique
to specific chemical bonds in molecules. By measuring the Hyperspectral imaging is a non-destructive, non-contact
absorption or reflection of light at various wavelengths in imaging technique involving the acquisition and process-
this range, NIRS can provide information about the chemi- ing of FFB images in hundreds or thousands of contiguous
cal composition of the fruitlet. For the use case of NIRS to narrows and closely spaced spectral bands across a broad
classify the ripeness of oil palm FFB, the oil and moisture spectrum of wavelengths. The acquired data is used to gen-
content in the mesocarp of FFB was evaluated using NIRS, erate a hyperspectral data cube, in which each pixel in the
and the result was validated with an average classifica- image represents a spectrum, and its spectral signature can
tion accuracy of > 81% (Silalahi et al., 2016a, 2016b). The be analyzed. The spectral signature of each pixel is used to
Food and Bioprocess Technology (2025) 18:183–206 187

identify and classify the fruitlet substances based on their and blackberries (Benelli et al., 2022; Gao et al., 2020; Oli-
chemical and physical properties. As FFB ripens, its com- sah et al., 2024; Pu et al., 2019). Hyperspectral imaging
position changes, affecting not only chemical properties offers a non-destructive and objective method for ripeness
like moisture content, oil concentration, and pigment com- assessment by analyzing the unique spectral signatures of
position but also physical properties like fruitlet firmness, these fruits at various ripeness stages.
texture, and even subtle changes in size. These combined
changes manifest in the fruitlet’s spectral response that are Raman Spectroscopy
captured using hyperspectral imaging. Hyperspectral imag-
ing is typically conducted with specialized imaging systems, Raman spectroscopy measures inelastic light dispersion
such as hyperspectral cameras and spectrometers, which can by irradiating an FFB with a laser source before collecting
capture high-resolution images over a broad spectrum of and analyzing the dispersed light to determine the molecu-
wavelengths. Chu et al. (2022) extracted and analyzed each lar structure and composition of the FFB. When laser light
pixel’s spectral information by using specialized software interacts with the fruitlet, some photons are elastically dis-
that was compatible with the equipment. The hyperspectral persed, which means they retain the same energy and wave-
imaging approach was then initiated to classify the ripeness length as the incident photons. Nonetheless, a small propor-
of oil palm FFB for an in-line monitoring system using red, tion of the dispersed photons are scattered inelastically, and
green, and blue (RGB) color images from a hyperspectral these photons endure a change in energy and wavelength as
camera that focuses on identifying the color reflectivity from a result of interactions with the FFB molecules. The energy
each pixel with intensity values ranging from 615 to 940 nm and wavelength shift of the dispersed photons provides
with an average 20 nm interval from 16 light-emitting diode information about the vibrational modes of the fruitlet’s
(LED). The statistical results from the study demonstrated molecules, and the resultant Raman spectra can be used
significant differences in ripeness at a wavelength of 660 nm. to identify and quantify the chemical composition. Raman
The captured images were processed with linear discrimi- spectra can reveal the types of chemical bonds present in an
nant analysis (LDA) with Mahalanobis distance, yielding an FFB, the relative concentrations of various chemical species,
average accuracy of 88.2% (Setiawan et al., 2019). Bensaeed and the structural properties, such as the presence of impuri-
et al. (2014) used a hyperspectral-based system to classify ties or defects (Dan et al., 2018). Raman spectroscopy is a
FFB ripeness by recording the reflectance at various wave- popular approach for oil palm FFB because of its carotene
lengths. The oil palm FFB was classified into three ripeness content. The Raman peak averaging at 1515cm−1 is proven to
categories (ripe, underripe, and overripe) using a hyperspec- be a crucial molecular fingerprint for carotene levels, which
tral device and an artificial neural network (ANN) model. can serve as a ripeness indicator in oil palm FFB. Further
The ANN model successfully classified the FFB with an signal analysis on the Raman peak reveals four primary
accuracy of > 95%. The spectral bands from 710 to 940 nm sub-bands (lycopene, β-carotene, lutein, and neoxanthin),
demonstrated the best test accuracy between the three types which originate from the stretching vibration of carotenoid
of oil palm FFB ripeness. FFB has a stable range from 770 molecules found in the peel of the oil palm fruitlet. The fine
to 890 nm with an accuracy of 83.3%, which considers the kernel nearest neighbors (KNN) classifier is found to provide
overall most optimum wavelengths to classify between the the highest overall accuracy of 100% (Raj et al., 2021) with
three ripeness categories of oil palm FFB. Studies have the use of limited samples (46 samples). Another study used
implemented a similar approach with several LEDs in vari- a confocal Raman spectrometer to extract spectral features
ous specific wavelengths (Cherie et al., 2015; Setiawan & from the FFB and further analyzed with ANN to obtain a
Prasetya, 2020; Utom et al., 2018) to acquire multispectral prediction accuracy of 95.48% (Hong et al., 2021), demon-
images. Hue, saturation, and value (HSV) color spaces in the strating the effectiveness of this method. Tzuan et al. (2022)
acquired images were the elements for evaluation, and the located the following eight hidden Raman peaks: protein,
most optimum result was obtained at a 940 nm wavelength. beta carotene, carotene, lipid, guanine/cytosine, chloro-
Hyperspectral imaging is prone to background noises in the phyll-a, and tryptophan. Analysis of variance revealed six
image; thus, obtaining clear and background-free images is peak intensities from proteins to amide III (β-sheet), beta-
only possible in a controlled laboratory with independent carotene, carotene, lipid, guanine/cytosine, and carotene;
FFB, which is a disadvantage in real-time applications. one peak location from lipid was found to be significant.
The present review primarily explores the use of hyper- An automated oil palm FFB ripeness classification system
spectral imaging for FFB ripeness classification. However, deployed with ANN using the seven significant features
Hyperspectral imaging holds immense potential for assess- demonstrated an overall performance of 97.9% accuracy
ing the ripeness of various other fruits. Researchers have (Tzuan et al., 2022), and KNN achieved 90.9% accuracy
actively applied this technology to monitor the ripening (Azmi et al., 2022). Raman spectroscopy can classify the
stages of diverse fruits like strawberries, bananas, kiwifruits, ripeness of FFB, which is robust against lights and noise
188 Food and Bioprocess Technology (2025) 18:183–206

caused by moisture. The downside of Raman spectroscopy Microwave Spectroscopy

is that Raman scattering is an inherently weak process that
requires specialized instrumentation operated in a lab envi- Microwave spectroscopy uses microwave radiation to exam-
ronment for precise laser sources, high-quality optics, and ine the rotational energy levels of fruitlet molecules. It is
sensitive detectors, which limits the outdoor application and based on the observation that molecules can absorb and
pre-harvesting classification. emit electromagnetic radiation at specific frequencies cor-
responding to their rotational energy levels by conveying
Fluorescence Spectroscopy microwave radiation through a fruitlet containing molecules.
As the microwave radiation interacts with the fruitlet’s mol-
Fluorescence spectroscopy is a spectroscopic method that ecules, the molecules can absorb the radiation and attain
uses fluorescence to determine the properties of FFB fruitlet higher rotational energy levels. The assimilated energy alters
molecules. It entails the excitation of an FFB with a specific the dipole moment of the molecule, which can be measured
wavelength of light and the measurement of the emitted fluo- and analyzed to reveal information about the molecule’s
rescence light at an extended wavelength. Because of the rotational energy levels (Townes & Schawlow, 2012). You
relaxation of excited electronic states in the fruitlet, the fluo- et al. (2020) used various microwave sensors used in oil
rescence emission provides information about the molecular palm monitoring for FFB ripeness classification to obtain the
structure and properties of the fruitlet. The FFB’s properties shifting resonance frequency/wavelength or the change in
can be determined by measuring and analyzing the fluores- the reflection/transmission coefficient. The dielectric proper-
cence light emitted (Lakowicz, 2011). A fluorescence sensor ties and moisture content are the important elements when
is used to obtain the blue-to-red fluorescence ratio param- determining the corresponding classification for FFB. Fur-
eter from blue-green (447 nm) to far-red (685 nm) emission thermore, Yeow et al. (2010) developed a low-cost coaxial
signals by using UV light emitting diode as the excitation moisture sensor for determining the moisture content of oil
light source; the classification and regression tree method palm FFB of various degrees of fruit ripeness. The sensor
successfully distinguished between the three categories with operates between 1 and 5 GHz and is suitable for single-
an overall classification accuracy of 89.7%. An optical probe fruit measurement. The actual moisture content was deter-
consisting of a 635 nm diode laser and two photodiodes was mined using the standard oven drying method and compared
used to study the relation of fruitlet firmness, which resulted with the predicted value of fruit moisture content obtained
in photodiode voltages. The output voltages were amplified using the studied sensor. The sensor was used to monitor
from a commercial multi-board circuit, and firmness levels fruit ripeness based on the measurement of the phase or
were obtained from the penetrometer. The linear correlation magnitude of the reflection coefficient. The profundity of
was calculated to get the conclusion that ripe FFB produces microwave penetration can vary depending on the size of the
higher voltage and vice versa (Sari et al., 2019). Hazir et al., fruit samples, which can consequently affect the accuracy
(2012a, 2012b) investigated the potential of flavonoids and of the measurements. Based on the limitations of the probe
anthocyanins because predictors were used to classify the configuration at frequencies below 4 GHz, the minimum
degree of oil palm FFB ripeness. The research collected 180 sample thickness for significant signal penetration was estab-
oil palm FFB samples from a private plantation in Malaysia lished to be 3.5 mm (Yeow et al., 2010). For larger or thicker
and scanned them using a hand-held Multiplex ® 3 multi- fruit samples, microwave radiation may not penetrate deeply
parameter fluorescence sensor. The findings show that both enough to provide reliable information about the internal
predictors can be good indicators for classifying oil palm properties of the fruit. This limitation makes it difficult to
FFB into under-ripe, ripe, and over-ripe maturity categories. use microwave spectroscopy to determine the maturity of
The highest overall classification accuracy was 87.7% using FFB of varying sizes and thicknesses. Moreover, this method
a stochastic gradient boosting trees model with both predic- lacks performance evaluation in the reference papers, limit-
tors. The positive outcomes of this research show the poten- ing its comparability with other techniques in this review.
tial of flavonoids and anthocyanins as predictors for oil palm
FFB ripeness. Fluorescence spectroscopy for outdoor real- Inductive Sensing
time FFB ripeness detection faces obstacles such as inter-
ference from ambient light and background fluorescence, Inductive sensing is an electromagnetic induction-based
variability in fruit samples, difficulties in optical access and method for determining FFB ripeness classification. Induc-
alignment, sensor stability, and the computational demands tive sensors can be used to measure certain ripeness-depend-
of real-time data processing. Fluorescence spectroscopy ent physical characteristics of FFB, such as electrical con-
must be carefully designed and optimized to detect outdoor ductivity and physical properties. The inductive sensor used
FFB maturity to overcome these limitations and improve for FFB ripeness classification, known as a flat-type induc-
accuracy and reliability. tor, consists of an electromagnetic field-generating air-cored
Food and Bioprocess Technology (2025) 18:183–206 189

coil sandwiched between two non-conducting materials such using discriminant analysis to determine the ripeness. Unlike
as Perspex (Fig. 3a). When the FFB fruitlet approaches the the previously mentioned papers, Sinambela et al. used a
sensor, the electromagnetic field interacts with the elec- low frequency (270–500 Hz) that resulted in lower resonant
trical and magnetic properties of the fruitlet. Changes in frequencies obtained for ripe FFB. Based on the field test
the dielectric properties of the fruitlet because of ripeness performance, the inductive sensor system can determine the
influence the capacitance and inductance value, allowing oil palm ripeness with an accuracy of 100%. Furthermore,
for a shift in resonance frequency. The distinctive resonance the proposed system can forecast the harvest time by trac-
frequency contributes to the classification of FFB ripeness. ing the ripeness after anthesis. Because the ripe FFB had a
Misron et al. (2017) proposed a flat-type inductive concept lower resonant frequency, the ripeness was predicted using
based on a resonant frequency technique, which focuses on polynomial regression.
the resonance frequency of the air coil and the change in the The calibration process is essential to achieve high-
air coil’s stray capacitance. The experiment was conducted accuracy FFB ripeness classification but is time-consuming
by selecting three FFB and observing the change of resonant and labor-intensive. The process requires the collection of
frequency from the 8th to the 18th week of anthesis. The reference data from various fruitlet samples representing
resonant frequencies of the ripe FFB were recorded from 9.8 multiple phases of ripeness. The inherent variability of fruit
to 9.9 MHz, while unripe started at 6.3 MHz. The research characteristics presents a significant challenge in achieving
further analyzed the moisture content in the fruitlet, which consistent and accurate ripeness classification. Furthermore,
can be determined from the capacitance of the sensor. As the sensitivity and specificity of the sensing system must be
the lipid content in the fruitlet increases when the number of thoroughly evaluated to guarantee that it can reliably detect
weeks progresses, the capacitance value obtained decreases. minor variations in ripeness levels. Fruit maturity, moisture
Harun et al. (2014) analyzed dual resonant frequencies content, and variations in fruit composition can affect the
by having two sets of flat-type inductive sensors measur- accuracy and dependability of the classification results.
ing from the top and bottom of the fruitlet. However, both External interferences (e.g., electromagnetic interference)
papers lack relevant performance results that are comparable and environmental factors (e.g., temperature and humid-
with those in the present review. Sinambela et al. (2020) ity) can influence sensor readings, necessitating the use of
designed a portable flat-type inductive sensor (Fig. 3b) to shielding techniques or signal-processing algorithms to miti-
collect resonant frequency data, which was then analyzed gate these effects. Overcoming these limitations necessitates

Fig. 3  (a) Flat-type inductive sensor consists of air-cored coil with two non-conductive materials. (b) A fresh fruit bunch ripeness assessor is
holding the Inductive Sensing device to identify the ripeness of the fresh fruit bunch
190 Food and Bioprocess Technology (2025) 18:183–206

meticulous system design, resilient calibration methods, and between the average temperature of the FFB and the ambient
additional research to optimize the technology for practica- temperature (∆Temp) to normalize the measurement (Fig. 4).
ble and accurate FFB ripeness classification in a real-time A classification analysis was performed to assess the poten-
dynamic environment. tial of ∆Temp of FFB as an index for ripeness classification.
The ∆Temp values were used in linear discriminant analysis,
Thermal Imaging Mahalanobis discriminant analysis, ANN, and KNN algo-
rithms. Notably, the ANN method achieved the highest over-
Thermal imaging is a technique for detecting and measuring all accuracy of 99.1%, demonstrating the promising potential
thermal radiation from an FFB. Thermal camera devices can of thermal imaging for oil palm maturity classification.
capture and quantify infrared radiation, which represents the FFB ripeness is a multifaceted attribute influenced by a
surface temperature of FFB fruitlets. The captured thermal complex interplay of factors beyond just temperature. These
images provide a visual representation of the temperature factors include changes in sugar composition, particularly
radiation distribution across the observed FFB, enabling the increases in non-structural, structural, and water-soluble
identification of surface temperature gradients and thermal sugars during the ripening (Thang et al., 2017). Additional
distribution variations. Oil has a lower specific heat capac- factors include potential variations in acidity and anthocya-
ity and thermal conductivity than water, so ripe FFB with nin content. The observed decrease in anthocyanin content
higher oil content dissipates heat less efficiently than unripe with increasing ripeness (Hazir et al., 2012a, 2012b) is
FFB with higher moisture content. Thus, ripe FFB tends to associated with a possible shift in FFB pH as it matures.
radiate higher temperatures than unripe FFB. Fauziah et al. Anthocyanins, as pH-sensitive pigments, may decrease in
(2021) presented a thermal-vision system to observe the concentration or change color as the environment becomes
thermal properties of oil palm FFB before harvest. The sys- less acidic (Luiza et al., 2022). Consequently, relying merely
tem uses the multilayer perceptron ANN algorithm to pro- on thermal imaging to evaluate the viability of fruit may
cess thermal images by identifying temperature values that not provide a comprehensive understanding of the actual
are represented by the intensity of RGB color. The R2 value ripeness stage. Furthermore, thermal imaging requires cali-
for FFB ripeness was found to be 0.9058 when correlated bration and validation to establish meaningful correlations
with oil content and 0.8039 when correlated with tempera- between temperature measurements and particular ripeness
ture. Makky & Cherie (2021) developed a linear regression indicators. A potential limitation of using thermal imaging
model to predict the ratio of oil and moisture content of for oil palm FFB maturity classification lies in the inherent
observed FFB with an R2 of 0.6732. Moreover, a multiple- variability of fruit size. The thermal camera captures the
regression model was successfully developed to estimate the entire FFB (Zolfagharnassab et al., 2022), so the measured
concentration of carotene in FFB mesocarp with an R2 of average radiation temperature is influenced by bunch size.
0.8122. Zolfagharnassab et al. (2022) used the difference Compared with smaller bunches, larger bunches tend to have

Fig. 4  Overview of thermal imaging technique, which comprises a three-stage workflow

Food and Bioprocess Technology (2025) 18:183–206 191

a higher average surface temperature, even if the internal of LiDAR has expanded significantly in the classification of
fruit maturity is similar. Further research must quantify this FFB ripeness. By harnessing the distinct variations in reflec-
effect and explore mitigation strategies. Moreover, environ- tion caused by varying degrees of FFB ripeness, research-
mental factors such as ambient temperature, humidity, and ers have unlocked the potential to derive crucial ripeness
external heat sources can influence thermal readings, poten- information by calculating the reflectance patterns in the
tially leading to inaccuracies if not properly accounted for. form of point clouds exhibited by FFB (Fig. 5b). The point
Thermal imaging may have limitations in differentiating sub- cloud data collected using the LiDAR sensor can be used
tle ripening stage variations. The thermal contrast between to develop algorithms for accurate FFB ripeness classifica-
various maturation levels of FFB may be insufficient to yield tion. Mat Seri et al. (2021) developed a LiDAR scanning
precise and consistent classification results. This result can system to predict the ripeness of oil palm FFB. The system
be especially difficult when dealing with FFB at similar was built using a LiDAR sensor and servo motors to obtain
maturity stages or FFB with intricate internal contents that the point cloud data and intensity value reflected from the
influence thermal conductivity. oil palm FFB. The authors designed the scanning mecha-
nism to overcome the vertical range limitation of the used
LiDAR and LLBI sensor, which is a single-stripe laser transmitter. Ripe oil
palm FFB tends to reflect a lower mean intensity value than
LiDAR has long been recognized as a powerful tool, par- unripe FFB. Furthermore, Husin et al. (2021) took a similar
ticularly for tasks such as localization and mapping. The approach, including coordinating information on the ripe
ability to measure the time it takes for the sensor light to FFB in a distribution map. This study validated the use of
return from the targeted object enables accurate mapping of LiDAR sensors to distinguish FFB ripeness. The LiDAR
surroundings and determination of the object’s precise posi- system was used to scan the FFB, and a reflectance percent-
tion. The LiDAR sensor can be mounted on a mobile robot age ranging from 0 to 100% was calculated based on the
to perform the task (Fig. 5a). Furthermore, the application concept of linearity (Zulkifli et al., 2019). This reflectance

Fig. 5  a) Camera or LiDAR mounted on a mobile robot to obtain real-time Fresh Fruit Bunch detection and ripeness classification. b) LiDAR
sensor used to scan and retrieve the point cloud data from the field
192 Food and Bioprocess Technology (2025) 18:183–206

value corresponds to the amount of infrared light at 905 nm ease of implementation. With advances in imaging technology
wavelength reflected by the FFB, which is proportional to and computational algorithms, the use of image processing in
its lipid content. FFB ripeness detection has gained significantly in recent years.
Apart from LiDAR, similar imaging methods such as Computer vision is a multidisciplinary field of study that seeks
LLBI have also shown promise in differentiating FFB to enable machines to perceive and comprehend visual data in
ripeness levels. LLBI systems use diode lasers with a manner analogous to human vision. The workflow for FFB
two distinct wavelengths, a CMOS camera, and an opti- ripeness classification through computer vision unfolds seam-
cal component. The system analyzes the relationship lessly in a well-orchestrated workflow (Fig. 6). The process
between the reflectance intensities represented using begins with the acquisition of high-resolution images of palm
RGB values in the captured FFB images and FFB height oil FFB at varying ripeness stages, followed by precise image
variations to assess ripeness (Shiddiq et al., 2017). annotation. Each fruit bunch is carefully delineated and labeled
Another research investigated a hybrid system of LLBI based on its ripeness, constituting an annotated dataset system-
and computer vision, combining the determination of atically organized in a structured database. This dataset serves
oil content and color values to achieve > 85% accuracy as the cornerstone for the ensuing machine-learning phase.
(Mohd Ali et al., 2020). Laser-light backscattering is Prior to model training, a crucial preprocessing stage refines
used to categorize FFB’s level of ripeness and to assess the dataset to ensure its optimal format. Feature extraction then
the maturity of plants, including watermelon, potato, identifies key characteristics such as color distribution and tex-
apricot, banana, and macaw oil palm (Sunilkumar & ture. The core of the process lies in training a machine learning
Sparjan, 2013; Adebayo et al., 2016; Costa et al., 2017; model, commonly CNN, utilizing the annotated dataset. Vali-
Zulkifli et al., 2019; Mozaffari et al., 2022). The chemi- dation and fine-tuning enhance the model’s adaptability to new,
cal and textural characteristics of the FFB can be identi- unseen data. Following the training, the model is ready for
fied through laser-light backscattering. The laser diode’s inference, delivering real-time ripeness predictions for input
wavelength can be adjusted to obtain the physical charac- FFB images. A popular approach in computer vision involves
teristics of the fruit. The FFB’s absorption and scattering integrating technology with a mobile robot. This approach
characteristics are also connected to its backscattering allows for automated, real-time classification of oil palm
profile. Similar research on cocoa pods (Lockman et al., FFB directly on the plantation, as explored by Suharjito et al.
2019) with LLBI revealed the disadvantage of detecting (2021). The robot is equipped with all-terrain wheels (Fig. 5a)
intact FFB on the tree because it requires a laser diode to navigate uneven terrain effectively, thereby developing
pointed at the specific fruit to obtain the reflectance algorithms and systems that derive meaningful characteristics
data, which inhibits effectiveness in a dynamic outdoor from digital images or videos. These characteristics serve as
environment. the foundation for more complex tasks for real-time dynamic
There are some potential drawbacks to using LiDAR and unstructured environment ripeness classification. Image
and LLBI techniques for FFB ripeness classification filtering, edge detection, feature detection, and segmentation
in outdoor dynamic environments. These techniques are among the most important techniques used in the feature
may have limitations in penetrating dense vegetation or extraction process in computer vision. These techniques are
opaque surfaces, hindering their ability to capture data used to recognize and isolate pertinent visual patterns, objects,
from deep within objects. Additionally, environmental or regions of interest within the visual data. Using machine
conditions such as fog, rain, or dust can affect the accu- learning and deep learning techniques, computer vision algo-
racy and performance of these techniques by scattering rithms (Mansour et al., 2022) can recognize and understand
or absorbing laser light. The implementation costs of complex visual patterns for FFB ripeness, enabling machines
LiDAR and LLBI systems can be high, and processing to comprehend and interpret visual content. Shabdin et al.
the large volume of generated data can be computation- (2016) investigated a color classification method using the hue,
ally intensive. Furthermore, these techniques primarily saturation, and intensity (HSI) model. This model can catego-
provide geometric information and may be limited in rize oil palm FFB into underripe and ripe classes. The model
their ability to offer detailed spectral or color informa- uses an ANN model that achieves 70% accuracy. Fadilah et al.
tion about the fruit bunches. Harsh weather conditions (2012) further improved accuracy to 71.66% by incorporating
and mobility or range limitations may also affect their a data reduction technique called PCA into the ANN model.
performance in dynamic outdoor environments. Alfatni et al., (2014, 2020) and Ibrahim et al. (2018a) explored
another approach based on real-time analysis of RGB color
Computer Vision spaces. This method involves transforming the RGB data into
HSI. The HSI color space allows researchers to analyze the
Computer vision has emerged as the preferred method for properties of digital color values and derive information about
object classification because of its commercial viability and the intensity and perceived color of light. The hue value can
Food and Bioprocess Technology (2025) 18:183–206 193

Fig. 6  The overall conventional workflow for computer vision method for fresh fruit bunch ripeness classification

potentially be used to estimate the oil content within the meso- the technique achieved 70% accuracy, specifically for ripe-
carp of the FFB. The accuracy of the data sets still has room ness classification. The technique relies on a concept called
for improvement. The hue color component of a growth model “bag-of-visual-words,” which was inspired by algorithms
was used to predict the maturity of FFB using the captured used in natural language processing (NLP). In NLP, words
digital image for estate planning, resulting in an R2 value of are the basic units of meaning. However, the accuracy of vari-
0.95 (Kassim et al., 2014). Robi et al. (2022) explored a deep ous grade classifications could be more consistent, requiring
learning approach for automated FFB classification using color extensive fine-tuning of hyperparameters (Fahmi et al., 2018;
vision technology. The researchers used two pre-trained deep Khamis et al., 2022). Similarly, an approach that used the slid-
learning models: ResNet50, which specializes in image clas- ing window method and retrained CNN to classify the ripeness
sification; and YOLOv3, designed for object detection. These of FFB from various categories in HSV color spaces yielded an
models were developed using the Darknet framework, which accuracy of 85%. By valuing the HSV color spaces, the envi-
helps address the challenge of limited data size in training deep ronment lighting condition is relatively essential to ensure the
learning models. Robi et al. (2022) achieved an average accu- collected data or images can extract the feature (Wong et al.,
racy of 90% using this approach. They further used YOLOv4 2020). The AlexNet model for CNN was conducted for FFB
and YOLOv4-tiny models combined with image processing ripeness grading identification that achieved an accuracy of
techniques based on deep learning. This approach yielded a 96.67% (Ibrahim et al., 2018b).
higher accuracy of 98.70%. Notably, poor lighting conditions Beside oil palm, other crops grown in other cultivation
can hinder the performance of these deep-learning models. areas that demonstrate notable contributions to the field
Aslamiah et al. (2019; 2022) validated another technique include banana, tomato, olive, mango, apple, and coco-
for FFB classification based on color extraction and image nut (Nandi et al., 2014; Chandra et al., 2014; Nagle et al.,
analysis. This method achieved an average accuracy of 57% 2016; Cárdenas-Pérez et al., 2017; Mazen & Nashat, 2019;
when classifying the FFB into four categories. Additionally, Malin et al., 2019; Mohamedon et al., 2021; Ortenzi et al.,
194 Food and Bioprocess Technology (2025) 18:183–206

2021; Gururaj et al., 2022; Sabzi et al., 2022; Aldakhil & detection and classification. The oil palm estate is typically
Almutairi, 2024; Meng et al., 2024; Vo et al., 2024; Wang located in an outdoor non-structured and dynamic environ-
et al., 2024; Xiao et al., 2024); these crops use a similar ment, so the weather and lighting conditions may differ,
method of computer vision with CNN to train the classifier resulting in an image that is not ideal for feature extraction.
from a color feature extraction. Computer vision techniques
are actively being explored and implemented in various FFB Ripeness Classification Techniques Comparison
agricultural activities. For instance, researchers have been
using machine learning on geophysical monitoring data to Performance Metrics Benchmark Comparison
optimize drilling processes (Osipov et al., 2023), developing
methods for estimating the spatial position of apples using The gathered benchmark comparison data provides valu-
computer vision systems (Andriyanov et al., 2022), and cre- able insights into the performance metrics of various non-
ating algorithms for detecting damaged or diseased potato destructive FFB ripeness classification techniques. The
tubers during harvest (Korchagin et al., 2021). Additionally, discussion primarily revolves around accuracy and R2 as
computer vision plays a role in identifying and classifying prominent indicators of the efficacy and reliability of each
mechanical damage inflicted on root crops during con- technique of FFB ripeness classification (Table 2). NIRS
tinuous harvesting operations (Osipov et al., 2022). These exhibits a remarkable R2 of 0.93 from Iqbal et al., indicat-
diverse applications highlight the versatility and potential of ing a strong correlation between predicted and actual val-
computer vision in transforming agricultural practices across ues. However, the experiment was restricted to the samples
a wide range of crops and cultivation stages. Most image taken from the same plantation estate because of the calibra-
processing techniques can achieve significant accuracy tion setting. Silalahi et al. reported an accuracy of 84.50%
performance; these are also popular approaches for object for the FFB ripeness classification with GANN on NIRS,

Table 2  Performance metrics comparison between different techniques for fresh fruit bunch ripeness classification
Techniques Accuracy (%) R2 Detected component References

NIRS N/A 0.93 Water content (Iqbal et al., 2019)

84.50 N/A Water content and oil to dry matter ratio (Silalahi et al., 2016a, 2016b)
Hyperspectral Imaging 98.67 N/A Chlorophyll absorption hole (Bensaeed et al., 2014)
88.20 N/A Oil content (Setiawan et al., 2019)
Raman spectroscopy 100.00 N/A Carotene (Raj et al., 2021)
95.48 N/A Carotene (Hong et al., 2021)
97.90 N/A Protein, β-carotene, carotene, lipids, guanine/ (Tzuan et al., 2022)
cytosine, chlorophyll-a, and tryptophan
90.90 N/A β-carotene, cuticular wax, and amino acid (Azmi et al., 2022)
Fluorescence spectroscopy 89.70 N/A Flavonoids and anthocyanin (Hazir et al., 2012a, 2012b)
Inductive sensing 100.00 N/A Oil content (Sinambela et al., 2020)
Thermal imaging N/A 0.95 Oil content (Fauziah et al., 2021)
N/A 0.81 Oil content, water content, and carotene (Makky & Cherie (2021))
99.10 N/A Oil content (Zolfagharnassab et al., 2022)
LiDAR & LLBI 85.00 N/A Oil content and color value (Mohd Ali et al., 2020)
Computer vision 70.00 N/A HSI (Shabdin et al., 2016)
N/A 0.93 Hue value (Fadilah et al., 2012)
N/A 0.98 RGB, HSI, HSV, YCbCr, and YIQ (Septiarini et al., 2021)
98.70 N/A RGB (Robi et al., 2022)
57.00 N/A Hue value (Aslamiah et al., 2019)
74.00 N/A RGB (Aslamiah et al., 2022)
76.00 N/A Edge, shape, and RGB (Khamis et al., 2022)
85.00 N/A HSV (Wong et al., 2020)
94.00 N/A RGB (Alfatni et al., 2020)
92.00 N/A RGB (Ibrahim et al., 2018b)

NIRS near infrared spectroscopy; N/A not available light detection and ranging; LLBI laser-light backscattering imaging; HSI hue, saturation, and
intensity; RGB red, green, blue; HSV hue, saturation, and value; YCbCr blue-difference chroma, red-difference chroma; YIQ luminance, in-phase,
quadrature
Food and Bioprocess Technology (2025) 18:183–206 195

highlighting its potential for precise classification. Bensaeed approach, particularly within dynamic outdoor and unstruc-
et al. demonstrated that hyperspectral imaging can achieve tured environments. Table 3 presents the advantages and
an impressive accuracy of 98.67% under a stationed equip- disadvantages of ripeness classification techniques. Spec-
ment setup for data acquisition, which further hinders the troscopy, renowned for its capacity to delve into the chemi-
possibility of deployment in an outdoor environment. Raman cal composition of FFB, presents an invaluable resource
spectroscopy emerges as a technique of exceptional accu- for assessing ripeness. Nevertheless, it is imperative to rec-
racy, with Raj et al. using a fine KNN classifier to achieve ognize that when deployed in real-world outdoor environ-
an accuracy of 100.00%. These results suggest Raman’s ments, spectroscopy encounters challenges stemming from
potential for precise and reliable measurements. Notably, environmental interferences, which may introduce potential
subsequent Raman readings demonstrated consistently high inaccuracies. The adaptability of spectroscopy and the task
accuracy rates of 95.48% from Hong et al., 97.90% from of calibrating spectroscopic instruments to suit the nuanced
Tzuan et al., and 90.90% from Azmi et al. This remark- conditions of outdoor and unstructured environments repre-
able accuracy stems from Raman spectroscopy’s ability to sent critical factors influencing the feasibility of deploying
detect specific chemical compositions within the fruitlets. this technique in real time. Although inductive sensing dem-
This intrinsic capability serves as a valuable compensatory onstrates swiftness in internal analysis, it requires calibration
factor for the relative lack of portability associated with the tailored to each distinct fruit type. This calibration process
setup. Hazir et al. achieved an accuracy level of 89.70% introduces complexity and may necessitate time-consuming
using fluorescence spectroscopy. This achievement can be adjustments in practical, real-world applications. Notably,
attributed to the heightened chemical reactivity observed the technology’s depth of penetration might not entirely
during the ripe stage of the fruitlets, which is a critical capture the breadth of internal variations, thereby restrict-
insight that underscores the technique’s effectiveness in FFB ing its efficacy in assessing FFB ripeness within dynamic
ripeness classification. Although inductive sensing boasts outdoor environments. The setup of thermal imaging is
a promising 100.00% accuracy rate for FFB ripeness clas- comparatively straightforward and demands less equipment.
sification by Sinambela et al., its application necessitates Nevertheless, its application encounters challenges because
direct contact with the fruitlet for accurate measurement. of the capricious nature of outdoor, unstructured environ-
This requirement, however, introduces an operational chal- ments. Susceptibility to external factors such as fluctuating
lenge. This specific requirement mandates the expertise of surface temperatures and the need for precise calibration
experienced workers to ensure precise placement of the sen- emerge as practical hurdles, demanding consideration dur-
sor on the fruitlet rather than the spikelet, introducing an ing deployment. LiDAR emerges as a standout choice for
efficiency drawback and a potential learning curve in the FFB ripeness classification in dynamic outdoor settings. Its
process. Thermal imaging proves to be a robust technique remarkable advantages, including the provision of precise
for FFB ripeness classification, evidenced by its high aver- 3D structural data, independence from fruit size variations,
age R2 values of 0.95 from Fauziah et al. and 0.81 by Makky and adaptability to changing lighting and foliage conditions,
et al., in tandem with impressive accuracy rates of 99.10% align seamlessly with the demands of the palm oil indus-
by Zolfagharnassab et al. These results reveal the technique’s try’s environment. The ability to generate real-time insights,
substantial potential. However, it is important to note that the capture external structural attributes, and offer consistent
surrounding environment exerts a direct effect on the FFB’s measurements positions LiDAR as a valuable tool for on-
actual temperature. This environmental influence is particu- field deployment, especially when size and shape assessment
larly pronounced in outdoor settings, ultimately limiting the is critical. The drawbacks include the requirement of com-
feasibility of using this technique for outdoor classification putational power for processing 3D data and rainy condi-
applications. Computer vision, despite a varied range of tions that might potentially affect the travel of emitted light.
results, displays noteworthy accuracy levels, such as 98.70% Computer vision, armed with its prowess in extracting rich
by Robi et al., 85.00% from Wong et al., and 92.00% from visual features, emerges as a potent option for real-world
Ibrahim et al. The R2 values (0.88–0.98) are also promising, deployment in FFB ripeness classification. Its adaptability
suggesting strong correlations between predicted and actual to diverse fruit types, rapid processing capabilities, and non-
outcomes. Notably, accuracy varies among instances, with invasive nature hold considerable promise. In the context of
values ranging from 57.00% to 92.00%. This variation could the palm oil industry, where variations in FFB appearance
be attributed to the complexity of computer vision tasks and dynamic outdoor environments are the norm, computer
and diverse datasets. Although the provided references lack vision is more suited for the role. However, the nuances of
specific performance metrics, microwave spectroscopy and lighting variations, complex backgrounds, and the resource-
LiDAR stand as notable techniques in the realm of sensing. intensive nature of model training should be navigated.
The choice of FFB ripeness classification techniques In this dynamic domain, where the palm oil industry
plays a pivotal role in ensuring an accurate and real-time intersects with outdoor unstructured environments, LiDAR
 196

Table 3  Summary of working principles, advantages, and disadvantages of all mentioned techniques
Techniques Working principles Advantages Disadvantages

Spectroscopy Analyses FFB’s interaction with light to identify ripe- • Provides detailed chemical composition of FFB • Tedious sample preparation
ness based on spectral characteristics • Various equipment to capture different aspects of • Not portable, bonded with laboratory set up
FFB characteristics • Surface conditions affect real-time consistency
• Well-calibrated models offer accurate predictions
Inductive sensing Measures changes in electrical properties to • Fast measurements suitable for real-time applica- • Tedious calibration process from the inherent vari-
determine FFB ripeness through electromagnetic tions ability
interactions • Portable to operate outdoors • Affected by external interference
Thermal imaging Detects variations in FFB temperature to classify • Simple setup for real-time monitoring • Surface temperature readings are easily influenced
ripeness based on heat patterns by external conditions
• Exocarp thickness affects the accuracy
LiDAR & LLBI Uses laser pulses to measure distance and light scat- • Detailed insights into the size and shape of FFB • Advanced processing, increasing computational
tering for ripeness classification • Vulnerable to fruit size variation requirements
• Capable of encompassing varying lighting condi- • Severe rain conditions may affect the accuracy
tions and foliage
Computer vision Utilizes image analysis and pattern recognition to • Overall the most significant performance of accu- • Lighting variations that can impact consistency and
classify FFB ripeness based on visual features racy accuracy
• Real-time classification capability • Challenge in complex background or overlapping
• High adaptability of CNN for FFB classification FFB
• Require intensive training datasets
Destructive methods Directly analyze the fruit itself, often involving physi- • Provide highly accurate data on internal fruit • Destroy the fruit being analyzed, reducing potential
cal damage properties yield
• Often established and well-understood techniques • Can be time-consuming and labor-intensive
Non-destructive methods Assess ripeness without harming the fruit • Preserve the fruit for further processing • Accuracy can be influenced by external factors like
• Can be faster and more automated for large-scale lighting or weather
operations • May require sophisticated equipment and calibration
• Offer real-time assessment in field settings for reliable results

FFB fresh fruit bunch, LiDAR light detection and ranging, LLBI laser-light backscattering imaging, CNN convolutional neural network
Food and Bioprocess Technology (2025) 18:183–206
Food and Bioprocess Technology (2025) 18:183–206 197

and computer vision shine as beacons of potential. Although testament to the increasing confidence in the efficacy of deep
LiDAR’s structural insights and adaptability resonate with learning for image-based classification tasks. The consen-
size and shape assessment, computer vision’s visual features sus seems to be that, as datasets grow in complexity, CNNs
and real-time processing prowess align compellingly with provide the most viable and scalable solutions for accurate
the requirements of the field. FFB ripeness classification.
Previous researchers have believed that the shift towards
Machine Learning Models and Algorithms Comparison deep learning and advanced models in the context of FFB
ripeness classification can be attributed to several compel-
The landscape of FFB ripeness classification has undergone ling reasons. These trends are reflective of the growing
significant evolution, as reflected in the diverse array of recognition among researchers of the unique capabilities
techniques and models employed in recent years (Table 4). and advantages offered by deep learning methodologies.
NIRS has seen a progression from GANN in 2016 to more Agricultural datasets, including those for FFB ripeness, are
advanced ANN in 2022 by Novianty et al., indicating a inherently complex, featuring intricate patterns and varia-
growing reliance on sophisticated machine learning mod- tions. Deep learning models, particularly CNNs, excel at
els. Hyperspectral imaging, introduced by Bensaeed et al. in capturing and learning from these complex, non-linear
2014 by using ANN, has exhibited a trend towards simplic- relationships in data. Deep learning models automatically
ity with the adoption of linear regression by Cherie et al. in learn hierarchical representations of features from raw data,
2015 and, subsequently, LDA by Setiawan et al. in 2019. allowing them to extract meaningful patterns and features
This shift suggests a strategic move towards more interpret- without explicit manual feature engineering. This ability is
able and less complex models. On the other hand, Raman particularly beneficial in scenarios where the relevant fea-
spectroscopy has exhibited a strong preference for ANN, tures for classification may not be immediately apparent.
as demonstrated by Dan et al. in 2018, Raj et al. in 2021, Deep learning models, especially CNNs, are highly scalable.
Hong et al. in 2021, Tzuan et al. in 2022, and Azmi et al. As datasets grow in size and complexity, these models can
in 2022. This preference likely stems from ANNs’ ability adapt and continue to provide accurate and efficient results.
to effectively handle the complex, non-linear relationships This scalability is crucial in agricultural applications where
present within Raman spectral data for FFB ripeness clas- datasets may vary significantly because of factors such as
sification. Fluorescence spectroscopy, introduced in 2012 by climate, soil conditions, and crop varieties. Unlike conven-
Hazir et al., Shariff et al., and Amiruddin et al., has shown tional methods, deep learning models support end-to-end
a reduced number of recent studies, possibly indicating a learning, indicating that the model learns directly from the
saturation point in its application or a shift towards newer raw input data to the output, thereby eliminating the need
and more advanced techniques in the domain. Microwave for manual preprocessing and feature engineering. This
spectroscopy, when absent from Table 4 because of the lack process streamlines the development process and often
of machine learning applications in prior research, presents leads to more efficient and accurate models. In the case of
an untapped avenue for potential exploration in FFB ripeness computer vision and spectral analysis, such as with hyper-
classification. Inductive sensing, a relatively recent entrant in spectral imaging, thermal imaging, and LiDAR, deep learn-
2020, showcases the utilization of LDA by Sinambela et al. ing models, especially CNNs, are well suited for handling
This technique is yet to witness extensive exploration, pre- image and spatial data. Their ability to understand spatial
senting opportunities for further development and research. hierarchies and patterns makes them ideal for tasks such as
Thermal imaging, introduced by Fauziah et al. in 2021, has image-based ripeness classification. Deep learning models
seen a blend of traditional linear regression and advanced demonstrate a remarkable ability to generalize across diverse
ANN in subsequent studies by Makky et al. and Cherie et al. conditions. This ability is crucial in agriculture, where vari-
in 2021 and Zolfagharnassab et al. in 2022. This combina- ations in environmental factors, crop types, and growing
tion of approaches suggests a nuanced approach to leverage conditions can be substantial. Models trained using deep
the strengths of various models for comprehensive analysis. learning techniques can adapt to these variations and provide
LiDAR and LLBI, as explored by Mohd Ali et al. in 2020, robust performance across various scenarios. The increasing
and combined PLS and PCA showcase a trend towards inte- availability of computational resources, including powerful
grating multiple techniques for a more holistic understanding GPUs and TPUs, has facilitated the training of deep learn-
and accurate classification. Computer vision experienced a ing models on large-scale datasets. This accessibility has
dynamic shift over the years since the early use of ANN in empowered researchers to experiment with and deploy more
2012; the field has rapidly progressed through various mod- complex models, leading to advancements in the field. Deep
els such as CNN-AlexNet, SVM, and, eventually, the YOLO learning models, especially advanced architectures such as
series. The consistent adoption of advanced CNN models in YOLOv3, YOLOv4, and YOLOv5, have consistently dem-
recent years, such as YOLOv3, YOLOv4, and YOLOv5, is a onstrated state-of-the-art performances in various computer
198 Food and Bioprocess Technology (2025) 18:183–206

Table 4  The list of machine learning models and algorithms used in fresh fruit bunch ripeness classification
Techniques Year Machine learning Description References
models and
algorithms

NIRS 2016 GANN – Able to capture complex patterns and relationships in the (Silalahi et al., 2016a, 2016b)
data
– Training can be challenging and computationally expen-
sive
– Interpretability may be a challenge as often considered as
“black box” models
2016 LDA – Computationally efficient and straightforward (Makky, 2016)
– May not perform well when the classes are not well-
separated
– Generally, less computationally demanding compared to
more complex models
– Lack of flexibility to capture complex, non-linear rela-
tionships in the data
2019 PLS – Suitable for situations with a high number of predictor (Iqbal et al., 2019)
variables
– Interpretability can be challenging, especially with many
latent variables
– Moderate computational resources
– May not perform well if the assumptions regarding
linearity are violated
2022 ANN – Able to model complex relationships and patterns in the (Novianty et al., 2022)
data
– Robust to noisy data and can generalize well
– Can be computationally expensive
Hyperspectral 2014 ANN – Effective in capturing complex relationships within data, (Bensaeed et al., 2014)
imaging suitable for nonlinear patterns
– Prone to overfitting, requires a large amount of data for
training
– High computational demand during training
2015 Linear regression – Simple and interpretable (Cherie et al., 2015)
– Assumes linearity, may not capture complex patterns,
sensitive to outliers
– Low computational demand
2019 LDA – Efficient in dimensionality reduction, provides optimal (Setiawan et al., 2019)
linear discriminants
– Assumes normality and equal covariance matrices, sensi-
tive to outliers
– Moderate computational demand
Raman spectros- 2018 ANN – Offers flexibility in capturing complex relationships in (Dan et al., 2018)
copy data
– Suffer from overfitting, especially with limited data
– Training can be time-consuming
2021 Fine KNN – Robust for small datasets and simple to implement (Raj et al., 2021)
– Sensitive to irrelevant features, requires careful feature
selection
– Low computational requirement during training, higher
during prediction
2021 ANN – Similar to (Dan et al., 2018) but in a different Raman (Hong et al., 2021)
Spectrum
2022 ANN – Similar to (Dan et al., 2018) but in a different Raman (Tzuan et al., 2022)
Spectrum
2022 KNN & trilayered – Enhanced capacity to learn complex patterns (Azmi et al., 2022)
neural network – Increased risk of overfitting and requires careful tuning
– Higher computational power due to additional layers
Food and Bioprocess Technology (2025) 18:183–206 199

Table 4  (continued)
Techniques Year Machine learning Description References
models and
algorithms

Fluorescence 2012 Stochastic gradient – Effective at capturing complex relationships in data due (Hazir et al., 2012a, 2012b)
spectroscopy boosting trees to ensemble learning
– Training can be computationally demanding, especially
with large datasets
– Requires significant computational resources during
training
– Models lack interpretability compared to simpler models
2012 Classification and – Offer clear interpretability, aiding in understanding deci- (Hazir et al., 2012a, 2012b)
regression tree sion paths
– Prone to overfitting, particularly with deep trees and
small datasets
– Generally, computationally less intensive
– Lacks the global optimization capabilities
Inductive sensing 2020 LDA – Efficient for binary and multiclass classification (Sinambela et al., 2020)
– May not perform well if assumptions are violated
– Suitable for real-time applications with reasonable com-
putational resources
– Lack of sensitivity to outliers and violations of distribu-
tion assumptions
Thermal imaging 2021 ANN – Capable of learning complex patterns in data (Fauziah et al., 2021)
– Prone to overfitting, especially with limited data
– High computational demand during training
2021 Linear regression – Simple and computationally efficient (Makky & Cherie (2021))
– Provides transparency and interpretability in model
predictions
– Lack of the ability to capture nonlinear patterns inherent
in FFB ripeness characteristics
2022 ANN – Effective for nonlinear relationships (Zolfagharnassab et al., 2022)
– Requires substantial computational resources
LiDAR & LLBI 2020 PLS and PCA – Effective for handling collinear data, suitable for high- (Mohd Ali et al., 2020)
dimensional datasets
– Linear assumptions may not capture non-linear relation-
ships
– Low to moderate computational requirement
– Lack of interpretability in the extracted latent variables
200 Food and Bioprocess Technology (2025) 18:183–206

Table 4  (continued)
Techniques Year Machine learning Description References
models and
algorithms

Computer vision 2012 ANN – Versatile and effective in capturing complex patterns (Fadilah et al., 2012)
– Requires a large amount of data for training
– Moderate computational requirement
– Lack of explicit spatial awareness
2014 ANN – Similar to the previous Computer Vision ANN descrip- (Alfatni et al., 2014)
tion
2016 ANN – Similar to the previous Computer Vision ANN descrip- (Shabdin et al., 2016)
tion
2017 SVM – Effective for binary classification tasks (Sabri et al., 2017)
– Struggle with complex, non-linear data
– Low to moderate computational power requirement
– Lack of the inherent ability to handle spatial relationships
2018 Back Propagation – Efficient for supervised learning tasks (Fahmi et al., 2018)
– Vulnerable to vanishing/exploding gradient problems
– Moderate computational requirement
– Lack of robustness in handling unstructured data
2018 CNN-AlexNet – Effective feature extraction with multiple convolutional (Ibrahim et al., 2018b)
layers
– May be computationally demanding
– Moderate to high computational requirement
– Lack of optimization for lightweight applications
2019 SVM – Similar to the previous computer vision SVM description (Aslamiah et al., 2019)
2020 ANN – Similar to the previous computer vision ANN description (Alfatni et al., 2020)
2020 CNN-AlexNet – Similar to the previous computer vision CNN-AlexNet (Wong et al., 2020)
description
2021 ANN – Similar to the previous computer vision ANN description (Septiarini et al., 2021)
2021 CNN-MobileNetV1 – Efficient in terms of computational resources (Suharjito et al., 2021)
– May sacrifice some accuracy compared to larger models
– Low to moderate computational requirement
– Lack of high accuracy in certain intricate tasks
2022 CNN-YOLOv3 – High accuracy and real-time object detection capabilities (Khamis et al., 2022)
– Can be resource intensive
– High computational requirement
– Lack of precision in small object detection
2022 CNN-YOLOv3 – Similar to the previous Computer Vision CNN-YOLOv3 (Aslamiah et al., 2022)
description
2022 CNN-YOLOv4 – Improved accuracy over YOLOv3 (Robi et al., 2022)
– Higher computational demand
– Lack of efficiency on less powerful hardware
2022 CNN-YOLOv5 – Balanced accuracy and efficiency (Mansour et al., 2022)
– May not match the highest accuracy models
– Moderate to high computational requirement
– Lack of specialization for extreme-scale datasets

NIRS near infrared spectroscopy, GANN genetic algorithm neural network, LDA linear discriminant analysis, PLS partial least square, ANN arti-
ficial neural network, KNN kernal nearest neighbors, FFB fresh fruit bunch, PCA principal component analysis, SVM support vector machine,
CNN convolutional neural network.
Food and Bioprocess Technology (2025) 18:183–206 201

vision tasks, including object detection and classification. Challenges and Future Prospects
The promising results obtained from these models further
drive the adoption of deep learning in agriculture. When Navigating the intricate world of FFB ripeness classifi-
discussing the various machine learning models used in FFB cation presents substantial challenges that require crea-
ripeness classification, it is crucial to delve into the specifics tive solutions. Integrating these techniques into automatic
of each technique to provide a comprehensive understand- harvesting systems holds immense potential but demands
ing. Table 4 presents a detailed analysis of each model used strategies to match the speed of harvesting machinery. The
by various classification methods, unveiling nuanced advan- palm oil industry often encounters obstacles in outdoor
tages and disadvantages associated with their application. environments, dealing with varying lighting conditions,
The study not only discusses the computational requirements dense foliage, and uneven terrain. These challenges have
of these models but also highlights the lacking aspects that been addressed by combining LiDAR and computer vision
may influence their suitability for specific FFB ripeness clas- (Kang et al., 2022). The present review discusses previ-
sification tasks. A detailed description of these models can ous research focusing on developing hybrid fusion models
provide a clearer insight into the strengths and limitations that integrate LiDAR and computer vision, extending the
of each model, facilitating informed decision-making in the fusion system into 3D mapping and localization, which
selection of the most appropriate technique for real-world introduces complexities in capturing and interpreting
applications in palm oil industry settings. spatial information accurately. The fusion of computer
Although the focus of ripeness classification techniques vision and LiDAR can lead to optimized algorithms for
has been thoroughly discussed in previous sections, the efficient 3D mapping, revolutionizing on-field navigation
integration of these methods with harvesting devices holds and resource utilization. This exploration opens up excit-
immense potential for the palm oil industry. Unlike tomato ing possibilities for overcoming obstacles in FFB ripeness
harvesting, FFB are heavy and require careful handling dur- classification and enhancing the efficiency of harvesting
ing harvest. Current harvesting technology is not yet sophis- operations in dynamic outdoor environments. Acknowl-
ticated enough to handle these complexities. Extensive edging the pivotal role of data quality and size in training
research and development efforts are necessary to bridge deep learning models, researchers are increasingly turn-
the gap and create truly automated harvesting systems for ing to synthetic data to bolster their datasets. This trend
the palm oil industry. The most crucial part is being able to holds great promise for improving the accuracy of FFB
classify the ripe FFB and effectively locate it with specific ripeness classification, especially in complex and dynamic
coordinates. Combining LiDAR and camera technologies outdoor environments. By seamlessly blending synthetic
can offer a powerful approach to overcome the challenges data with real-world datasets, researchers can enhance the
because this combined approach can potentially improve the training process, expose models to a wider array of sce-
efficiency and precision of harvesting operations. Fu et al. narios, and address challenges related to dataset size and
(2023) demonstrated the potential of LiDAR and camera diversity (Goh et al., 2024). This trend not only offers a
fusion for fruit detection tasks in agricultural settings. Their cost-effective alternative for model optimization but also
findings suggest that this integration approach can lead to lays the groundwork for a more resilient and adaptable
robust and accurate fruit recognition, paving the way for its FFB ripeness classification system in diverse agricultural
future application in automated harvesting systems. settings. Future research must refine these methodolo-
Conventional methods for FFB ripeness assessment often gies, focusing on improving the realism of synthetic data
rely on destructive techniques that involve physically dam- and establishing standardized benchmarks for evaluating
aging the fruit. Despite providing some level of accuracy, model efficacy. The integration of synthetic data emerges
these methods come with significant drawbacks. Moreover, as a transformative strategy, ushering in a new era of pre-
non-destructive techniques offer a more efficient approach. cision in FFB ripeness classification. In the lush, natu-
When using computer vision or machine learning in con- ral settings where FFBs thrive, the abundance of foliage,
junction with sensors such as cameras or LiDAR, the FFB branches, and neighboring fruits creates visually clut-
remains intact on the tree throughout the assessment pro- tered backgrounds. This environmental richness poses a
cess. Furthermore, non-destructive methods are generally challenge in distinguishing FFB from their surroundings,
less labor-intensive and require simpler equipment compared leading to substantial occlusion. The boundaries of indi-
to destructive techniques, reducing operational costs. vidual fruit bunches become blurred or obscured, mak-
ing accurate identification and classification a tricky task.
Besides the complexity, the images are captured from a
considerable distance during field surveys or from aerial
platforms such as drones. This distance results in the fruit
202 Food and Bioprocess Technology (2025) 18:183–206

bunches appearing relatively small in the images, reducing ripeness classification seamlessly aligns with the evolving
the available visual detail for analysis. Consequently, accu- demands of agricultural practice and the broader ecosys-
rately assessing ripeness becomes a formidable challenge tem. The trajectory of research in this direction not only
due to the limited visual information that was empha- fosters innovation within the palm oil industry but also the
sized in the previous discussion. One promising solution overall agricultural landscape for other plantations as well,
involves leveraging the RT-DETR model for classifica- which serves as a pivotal component in the sustainable and
tion in FFB image analysis that has been implemented efficient evolution of agricultural methodologies.
recently for anomaly detection in wireless sensor networks
(Siva Kumar et al., 2023), underwater small object detec-
Author Contribution Jin Yu Goh performed conceptualization, inves-
tion (Chen et al., 2023), and marine ship detection (Xing tigation, analysis, writing, editing, and validation. Yusri Md Yunos
et al., 2023). This solution could potentially enhance the contributed to analysis, validation, and review. Mohamed Sultan
accuracy of ripeness assessment by effectively navigating Mohamed Ali performed supervision, conceptualization, writing
through the challenges posed by visually cluttered back- review, editing, methodology, investigation, analysis, validation, and
funding acquisition.
grounds, occlusion, and limited visual details.
Funding Open Access funding provided by the Qatar National Library.
This work was supported by Ministry of Higher Education under Fun-
damental Research Grant Scheme FRGS/1/2022/TK07/UTM/02/42
and Universiti Teknologi Malaysia under UTM Fundamental Research
Conclusion Grant Q.J130000.3823.22H55. Open Access funding provided by the
Qatar National Library.
In conclusion, the realm of FFB ripeness classification
within the palm oil industry is a multifaceted challenge, Data Availability No data was used for the research described in the
article.
demanding a strategic selection of techniques to ensure
accurate and real-time assessments. As described in this Declarations
comprehensive comparative analysis across spectroscopy,
inductive sensing, thermal imaging, LiDAR, LLBI, and Competing Interests The authors declare no competing interests.
computer vision have provided a nuanced understanding
of their distinct merits and limitations within the dynamic Open Access This article is licensed under a Creative Commons Attri-
and unstructured outdoor environments inherent to this bution 4.0 International License, which permits use, sharing, adapta-
tion, distribution and reproduction in any medium or format, as long
ripeness classification. Computer vision emerges as the as you give appropriate credit to the original author(s) and the source,
linchpin in FFB ripeness classification, outshining other provide a link to the Creative Commons licence, and indicate if changes
techniques with its unparalleled ability to discern intri- were made. The images or other third party material in this article are
cate visual details and swift adaptability. This adaptability included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in
positions it as the preferred choice for real-world appli- the article’s Creative Commons licence and your intended use is not
cations in the palm oil industry, promising efficient and permitted by statutory regulation or exceeds the permitted use, you will
accurate assessments. Additionally, LiDAR proves to be a need to obtain permission directly from the copyright holder. To view a
valuable tool, offering crucial 3D structural data and adept copy of this licence, visit [Link]
handling of diverse lighting and environmental conditions.
This finding positions LiDAR as a strong supporting tech-
nique for FFB ripeness classification, particularly when
external structural attributes carry significant importance. References
The challenges identified, including the optimization of
Adebayo, S. E., Hashim, N., Abdan, K., & Hanafi, M. (2016). Appli-
LiDAR and computer vision fusion, the need for high- cation and potential of backscattering imaging techniques in
quality datasets in computer vision, and the exploration agricultural and food processing - A review. In Journal of Food
of a higher classification accuracy model, call for inno- Engineering (Vol. 169, pp. 155–164). Elsevier Ltd. [Link]
vative solutions. Advanced sensor integration technolo- org/​10.​1016/j.​jfood​eng.​2015.​08.​006
Aldakhil, L. A., & Almutairi, A. A. (2024). Multi-fruit classification
gies, synthetic data implementation, and the deployment and grading using a same-domain transfer learning approach.
of RT-DETR as a new object detection model represent IEEE Access, 12, 44960–44971. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 109/​
promising prospects for overcoming these challenges. ACCESS.​2024.​33792​76
The present review provides a beacon light on the current Alfatni, M. S. M., Mohamed Shariff, A. R., Ben Saaed, O. M., Albh-
bah, A. M., & Mustapha, A. (2020). Colour feature extraction
state of FFB ripeness classification techniques and sets the techniques for real time system of oil palm fresh fruit bunch
stage for future advancements. By addressing challenges maturity grading. IOP Conference Series: Earth and Environ-
and embracing prospective solutions, researchers can pave mental Science, 540(1). [Link]
the way for transformative innovations, ensuring that FFB 540/1/​012092
Food and Bioprocess Technology (2025) 18:183–206 203

Alfatni, M. S. M., Shariff, A. R. M., Abdullah, M. Z., Marhaban, M. evaluation using hyperspectral imaging. Agriculture (Switzer-
H., Shafie, S. B., Bamiruddin, M. D., & Saaed, O. M. B. (2014). land), 12(4). [Link]
Oil palm fresh fruit bunch ripeness classification based on rule- Costa, A. G., Pinto, F. de A. de C., JÃonior, R. B. A., Motoike, S.
based expert system of ROI image processing technique results. Y., & Gracia, L. M. N. (2017). Determination of macaw fruit
IOP Conference Series: Earth and Environmental Science, 20(1). harvest period by biospeckle laser technique. African Journal of
[Link] Agricultural Research, 12(9), 674–683[Link]
Andriyanov, N., Khasanshin, I., Utkin, D., Gataullin, T., Ignar, S., Shu- AJAR2​016.​12039
maev, V., & Soloviev, V. (2022). Intelligent system for estimation Dan, S. A. M., Hashim, F. H., Raj, T., Huddin, A. B., & Hussain, A.
of the spatial position of apples based on YOLOv3 and real sense (2018). Classification of oil palm fresh fruit bunches (FFB) using
depth camera D415. Symmetry, 14(1). [Link] Raman spectroscopy. International Journal of Engineering and
sym14​010148 Technology(UAE), 7(4), 184–188. [Link]
Aslamiah Ghazalli, S., Selamat, H., Omar, Z., & Yusof, R. (2019). v7i4.​11.​20798
Image analysis techniques for ripeness detection of palm oil fresh Fadilah, N., Mohamad-Saleh, J., Halim, Z. A., Ibrahim, H., & Ali, S. S.
fruit bunches. Journal of Electrical Engineering, 18(3), 57–62. S. (2012). Intelligent color vision system for ripeness classifica-
www.​elekt​rika.​utm.​my tion of oil palm fresh fruit bunch. Sensors (switzerland), 12(10),
Aslamiah, S., & Ghazalli, B. (2022). Oil palm fresh fruit bunches ripe- 14179–14195. [Link]
ness classification with color and texture features extraction. Fahmi, F., Palti, H., Emerson, S., & Suherman, S. (2018). Oil palm
Azmi, M. H. I. M., Hashim, F. H., Huddin, A. B., & Sajab, M. S. fresh fruit bunch ripeness classification using back propagation
(2022). Correlation study between the organic compounds and and learning vector quantization. IOP Conference Series: Mate-
ripening stages of oil palm fruitlets based on the raman spectra. rials Science and Engineering, 434(1). [Link]
Sensors, 22(18). [Link] 1757-​899X/​434/1/​012066
Basyuni, M., Amri, N., Putri, L. A. P., Syahputra, I., & Arifiyanto, D. Fauziah, W. K., Makky, M., Santosa, & Cherie, D. (2021). Thermal
(2017). Characteristics of fresh fruit bunch yield and the physico- vision of oil palm fruits under difference ripeness quality. IOP
chemical qualities of palm oil during storage in north Sumatra, Conference Series: Earth and Environmental Science, 644(1).
Indonesia. Indonesian Journal of Chemistry, 17(2), 182–190. [Link]
[Link] Fu, K., Wei, P., Villacres, J., Kong, Z., Vougioukas, S. G., & Bai-
Benelli, A., Cevoli, C., Fabbri, A., & Ragni, L. (2022). Ripeness evalu- ley, B. N. (2023). Fusion-driven tree reconstruction and fruit
ation of kiwifruit by hyperspectral imaging. Biosystems Engi- localization: Advancing precision in agriculture. [Link]
neering, 223, 42–52. [Link] org/​abs/​2310.​15138
2021.​08.​009 Gao, Z., Shao, Y., Xuan, G., Wang, Y., Liu, Y., & Han, X. (2020).
Bensaeed, O. M., Shariff, A. M., Mahmud, A. B., Shafri, H., & Alfatni, Real-time hyperspectral imaging for the in-field estimation of
M. (2014). Oil palm fruit grading using a hyperspectral device strawberry ripeness with deep learning. Artificial Intelligence
and machine learning algorithm. IOP Conference Series: Earth in Agriculture, 4, 31–38. [Link]
and Environmental Science, 20(1). [Link] oi.o​ rg/1​ 0.1​ 088/1​ 755-​ 04.​003
1315/​20/1/​012017 Goh, J. Y., Md Yunos, Y., & Mohamed Ali, M. S. (2024). Synthetic
Bin Mat Seri, A. D. I., Bin Mohd Kassim, M. S., Binti Abdul Rahman, data generation for fresh fruit bunch ripeness classification. In
S. R., & Sajak, A. A. B. (2021, August 23). Development of Communications in Computer and Information Science (pp.
virescens fresh fruit bunch ripeness prediction using LiDAR for 276–288). [Link]
smart agriculture. TENSYMP 2021 - 2021 IEEE Region 10 Sym- Gururaj, N., Vinod, V., & Vijayakumar, K. (2022). Deep grading of
posium. [Link] oi.o​ rg/1​ 0.1​ 109/T
​ ENSYM ​ P5285​ 4.2​ 021.9​ 55098​ 1 mangoes using convolutional neural network and computer
C. H. Townes, & A.L. Schawlow. (2012). Microwave spectroscopy. vision. Multimedia Tools and Applications[Link]
[Link] 1007/​s11042-​021-​11616-2
gAAQB​ A J& ​ o i= ​ f nd& ​ p g= ​ P P1& ​ d q=% ​ 2 2Mic ​ r owave+ ​ S pect​ Harun, N. H., Misron, N., Sidek, R. M., Aris, I., Ahmad, D., Wakiwaka,
rosco​py%​22+​by+​Townes+​and+​Schaw​low,+​Dover+​Publi​catio​ H., & Tashiro, K. (2013). Investigations on a novel inductive
ns,+​2012.​&​ots=​Rs_​lBsB2_-​&​sig=​DxXeo​2w5Vk​G7gCb​jJI3n​ concept frequency technique for the grading of oil palm fresh
XZYBw​KY#v=​onepa​ge&​q&f=​false fruit bunches. Sensors (switzerland), 13(2), 2254–2266. [Link]
Cárdenas-Pérez, S., Chanona-Pérez, J., Méndez-Méndez, J. V., Calde- doi.​org/​10.​3390/​s1302​02254
rón-Domínguez, G., López-Santiago, R., Perea-Flores, M. J., & Harun, N. H., Misron, N., Sidek, R. M., Aris, I., Wakiwaka, H., &
Arzate-Vázquez, I. (2017). Evaluation of the ripening stages of Tashiro, K. (2014). Dual resonant frequencies effects on an
apple (Golden Delicious) by means of computer vision system. induction-based oil palm fruit sensor. Sensors (switzerland),
Biosystems Engineering, 159, 46–58. [Link] 14(11), 21923–21940. [Link]
biosy​stems​eng.​2017.​04.​009 Hazir, M. H. M., Shariff, A. R. M., & Amiruddin, M. D. (2012a). Deter-
Chandra Sekhar, N., Bipan, T., & Chiranjib, K. (2014, December mination of oil palm fresh fruit bunch ripeness-Based on flavo-
28). Computer vision based mango fruit grading system. Inter- noids and anthocyanin content. Industrial Crops and Products,
national Conference on Innovative Engineering Technologies. 36(1), 466–475. [Link]
[Link] Hazir, M. H. M., Shariff, A. R. M., Amiruddin, M. D., Ramli, A. R., &
Chen, G., Mao, Z., Wang, K., & Shen, J. (2023). HTDet: A hybrid Iqbal Saripan, M. (2012b). Oil palm bunch ripeness classifica-
transformer-based approach for underwater small object detec- tion using fluorescence technique. Journal of Food Engineering,
tion. Remote Sensing, 15(4). [Link] 113(4), 534–540. [Link] oi.o​ rg/1​ 0.1​ 016/j.j​ foode​ ng.2​ 012.0​ 7.0​ 08
Cherie, D., Herodian, S., Ahmad, U., Mandang, T., & Makky, M. Hong, T. S., Hanim Hashim, F., Raj, T., & Huddin, A. B. (2021). Clas-
(2015). Optical characteristics of oil palm fresh fruits bunch sification of oil palm fruit ripeness using artificial neural net-
(FFB) under three spectrum regions influence for harvest deci- work. 2021 IEEE International Conference on Automatic Control
sion. 5(3). and Intelligent Systems, I2CACIS 2021 - Proceedings, 358–363.
Chu, X., Miao, P., Zhang, K., Wei, H., Fu, H., Liu, H., Jiang, H., & [Link]
Ma, Z. (2022). Green banana maturity classification and quality Husin, H. S., Amar, N., Bakar Sajak, A. A., & Sallehin Mohd Kassim,
M. (2021). Distribution map of oil palm fresh fruit bunch using
204 Food and Bioprocess Technology (2025) 18:183–206

LiDAR*. 2021 12th International Conference on Information non-destructive approach. Journal of Food Engineering, 120(1),
and Communication Systems, ICICS 2021, 4–9. [Link] 248–259. [Link]
10.​1109/​ICICS​52457.​2021.​94645​75 Malaysian Palm Oil Council. (n.d.). The oil palm tree. Retrieved May
Ibrahim, Z., Sabri, N., & Isa, D. (2018a). Palm oil fresh fruit bunch 17, 2023, from [Link]
ripeness grading recognition using convolutional neural network. Malin Sutan, S., Hendrawan, Y., Amini, A., & Maya Maharani, D.
Journal of Telecommunication, Electronics and Computer Engi- (2019). Science & technology intelligent non-invasive sensing
neering, 10. method in identifying coconut (Coco nucifera var. Ebunea) ripe-
Ibrahim, Z., Sabri, N., & Isa, D. (2018b). Palm oil fresh fruit bunch ness using computer vision and artificial neural network. Per-
ripeness grading recognition using convolutional neural net- tanika J. Sci. & Technol, 27(3), 1317–1339.
work. Journal of Telecommunication, Electronic and Computer Mansour, M. Y. M. A., Dambul, K. D., & Choo, K. Y. (2022). Object
Engineering. detection algorithms for ripeness classification of oil palm fresh
Indonesian Palm Oil Association. (n.d.). G20 - Indonesia palm oil fruit bunch. International Journal of Technology, 13(6), 1326–
facts. Retrieved May 17, 2023, from [Link] 1335. [Link]
almoi​lfacts.​com/​g20/ Mazen, F. M. A., & Nashat, A. A. (2019). Ripeness classification of
Iqbal, Z., Herodian, S., & Widodo, S. (2019). Development of partial bananas using an artificial neural network. Arabian Journal for
least square (PLS) prediction model to measure the ripeness of Science and Engineering, 44(8), 6901–6910. [Link]
oil palm fresh fruit bunch (FFB) by using NIR spectroscopy. IOP 1007/​s13369-​018-​03695-5
Conference Series: Earth and Environmental Science, 347(1). Mba, O. I., Dumont, M. J., & Ngadi, M. (2015). Palm oil: Processing,
[Link] characterization and utilization in the food industry - A review.
Joseph R. Lakowicz. (2011). Principles of fluorescence spectroscopy. In Food Bioscience (Vol. 10, pp. 26–41). Elsevier Ltd. [Link]
Kang, H., Wang, X., & Chen, C. (2022). Accurate fruit localisation doi.​org/​10.​1016/j.​fbio.​2015.​01.​003
using high resolution LiDAR-camera fusion and instance seg- Meng, Z., Du, X., Xia, J., Ma, Z., & Zhang, T. (2024). Real-time statis-
mentation. Computers and Electronics in Agriculture, 203. tical algorithm for cherry tomatoes with different ripeness based
[Link] on depth information mapping. Computers and Electronics in
Kassim, M. S. M., Ismail, W. I. W., Ramli, A. R., & Bejo, S. K. Agriculture, 220. [Link]
(2014). Image clustering technique in oil palm fresh fruit Misron, N., Aliteh, N. A., Harun, N. H., Tashiro, K., Sato, T., &
bunch (FFB) growth modeling. Agriculture and Agricultural Wakiwaka, H. (2017). Relative estimation of water content for
Science Procedia, 2, 337–344. [Link] flat-type inductive-based oil palm fruit maturity sensor. Sensors
2014.​11.​047 (Switzerland), 17(1). [Link]
Khamis, N., Selamat, H., Ghazalli, S., Md Saleh, N. I., & Yusoff, N. Mohamed Ahmed Mansour, M. Y. (2022). A review of non-destructive
(2022). Comparison of palm oil fresh fruit bunches (FFB) ripe- ripeness classification techniques for oil palm fresh fruit bunches.
ness classification technique using deep learning method. ASCC Journal of Oil Palm Research. [Link]
2022 - 2022 13th Asian Control Conference, Proceedings, 64–68. 2022.​0063
[Link] Mohamedon, M. F., Abd Rahman, F., Mohamad, S. Y., & Omran
Korchagin, S. A., Gataullin, S. T., Osipov, A. V., Smirnov, M. V., Khalifa, O. (2021). Banana ripeness classification using com-
Suvorov, S. V., Serdechnyi, D. V., & Bublikov, K. V. (2021). puter vision-based mobile application. Proceedings of the 8th
Development of an optimal algorithm for detecting damaged International Conference on Computer and Communication
and diseased potato tubers moving along a conveyor belt using Engineering, ICCCE 2021, 335–338. [Link]
computer vision systems. Agronomy, 11(10). [Link] ICCCE​50029.​2021.​94672​25
3390/​agron​omy11​101980 Mohd Ali, M., Hashim, N., & Abdul Hamid, A. S. (2020). Combina-
Lai, J. W., Ramli, H. R., Ismail, L. I., & Wan Hasan, W. Z. (2023). Oil tion of laser-light backscattering imaging and computer vision
palm fresh fruit bunch ripeness detection methods: A systematic for rapid determination of oil palm fresh fruit bunches maturity.
review. In Agriculture (Switzerland) (Vol. 13, Issue 1). MDPI. Computers and Electronics in Agriculture, 169. [Link]
[Link] 10.​1016/j.​compag.​2020.​105235
Lockman, N. A., Hashim, N., & Onwude, D. I. (2019). Laser-based Mohd Zulkifli, Z., Hanim Hashim, F., Raj, T., & Baseri Huddin, A.
imaging for cocoa pods maturity detection. Food and Bio- (2018). A rapid and non-destructive technique in determining
process Technology, 12(11), 1928–1937. [Link] oi.o​ rg/1​ 0.1​ 007/​ the ripeness of oil palm fresh fruit bunch (FFB) (Teknik Cepat
s11947-​019-​02350-7 dan Tidak Merosakkan dalam Menentukan Kematangan Tandan
Luiza Koop, B., Nascimento da Silva, M., Diniz da Silva, F., dos San- Buah Sawit Segar (FBB)). Jurnal Kejuruteraan, 30(1), 93–101.
tos, T., Lima, K., Santos Soares, L., José de Andrade, C., Ayala [Link]
Valencia, G., & Rodrigues Monteiro, A. (2022). Flavonoids, Mozaffari, M., Sadeghi, S., & Asefi, N. (2022). Prediction of the qual-
anthocyanins, betalains, curcumin, and carotenoids: Sources, ity properties and maturity of apricot by laser light backscatter-
classification and enhanced stabilization by encapsulation and ing imaging. Postharvest Biology and Technology, 186, 111842.
adsorption. Food Research International, 153, 110929. [Link] [Link]
doi.​org/​10.​1016/J.​FOODR​ES.​2021.​110929 Nagle, M., Intani, K., Romano, G., Mahayothee, B., Sardsud, V., &
Makky, M. (2016). A portable low-cost non-destructive ripeness Müller, J. (2016). Determination of surface color of “all yellow”
inspection for oil palm FFB. Agriculture and Agricultural Sci- mango cultivars using computer vision. International Journal
ence Procedia, 9, 230–240. [Link] Agriculture and Biological Engineering, 9(1). [Link]
2016.​02.​139 3965/j.​ijabe.​20160​901.​1861
Makky, M., & Cherie, D. (2021). Pre-harvest oil palm FFB nondestruc- Nandi, C. S., Tudu, B., & Koley, C. (2014). A machine vision-based
tive evaluation technique using thermal-imaging device. IOP maturity prediction system for sorting of harvested mangoes.
Conference Series: Earth and Environmental Science, 757(1), IEEE Transactions on Instrumentation and Measurement, 63(7),
012003. [Link] 1722–1730. [Link]
Makky, M., & Soni, P. (2014). In situ quality assessment of intact oil Novianty, I., Gilang Baskoro, R., Iqbal Nurulhaq, M., & Achirul
palm fresh fruit bunches using rapid portable non-contact and Nanda, M. (2022). Empirical mode decomposition of near-
infrared spectroscopy signals for predicting oil content in palm
Food and Bioprocess Technology (2025) 18:183–206 205

fruits. Information Processing in Agriculture[Link] Enabling Technologies, ICIoT 2020, 85–88. [Link]
1016/j.​inpa.​2022.​02.​004 1109/​ICIoT​48696.​2020.​90894​31
Olisah, C. C., Trewhella, B., Li, B., Smith, M. L., Winstone, B., Whit- Shabdin, M. K., Shariff, A. R. M., Johari, M. N. A., Saat, N. K., &
field, E. C., Fernández, F. F., & Duncalfe, H. (2024). Convo- Abbas, Z. (2016). A study on the oil palm fresh fruit bunch (FFB)
lutional neural network ensemble learning for hyperspectral ripeness detection by using hue, saturation and intensity (HSI)
imaging-based blackberry fruit ripeness detection in uncon- approach. IOP Conference Series: Earth and Environmental
trolled farm environment. [Link] Science, 37(1). [Link]
Ortenzi, L., Figorilli, S., Costa, C., Pallottino, F., Violino, S., Pagano, Shiddiq, M., Fitmawati, Anjasmara, R., Sari, N., & Hefniati. (2017).
M., Imperi, G., Manganiello, R., Lanza, B., & Antonucci, F. Ripeness detection simulation of oil palm fruit bunches using
(2021). A machine vision rapid method to determine the ripe- laser-based imaging system. AIP Conference Proceedings, 1801.
ness degree of olive lots. Sensors, 21(9). [Link] oi.o​ rg/​10.3​ 390/​ [Link]
s2109​2940 Silalahi, D. D., Reaño, C. E., Lansigan, F. P., Panopio, R. G., & Ban-
Osipov, A., Pleshakova, E., Bykov, A., Kuzichkin, O., Surzhik, D., tayan, N. C. (2016). Using genetic algorithm neural network
Suvorov, S., & Gataullin, S. (2023). Machine learning methods on near infrared spectral data for ripeness grading of oil palm
based on geophysical monitoring data in low time delay mode (Elaeis guineensis Jacq.) fresh fruit. In Information Processing
for drilling optimization. IEEE Access, 11, 60349–60364. [Link] in Agriculture (Vol. 3, Issue 4, pp. 252–261). China Agricultural
doi.​org/​10.​1109/​ACCESS.​2023.​32840​30 University. [Link]
Osipov, A., Shumaev, V., Ekielski, A., Gataullin, T., Suvorov, S., Silalahi, D. D., Reaño, C. E., Lansigan, F. P., Panopio, R. G., Bantayan,
Mishurov, S., & Gataullin, S. (2022). Identification and clas- N. C., Davrieux, F., Caliman, J. P., Yuana, Y. Y., & Sudarno.
sification of mechanical damage during continuous harvesting (2016). Near infrared spectroscopy: A rapid and non-destructive
of root crops using computer vision methods. IEEE Access, 10, technique to assess the ripeness of oil palm (Elaeis guineensis
28885–28894. [Link] Jacq.) fresh fruit. Journal of Near Infrared Spectroscopy, 24(2),
Pavia, D. L., Lampman, G. M., Kriz, G. S., & Vyvyan, J. R. (2013). 179–190. [Link]
Introduction to spectroscopy (L. P. Donald, M. L. Gary, S. K. Sinambela, R., Mandang, T., Subrata, I. D. M., & Hermawan, W.
George, & R. V. James, Eds.; 5th ed.). Cengage Learning. (2020). Application of an inductive sensor system for identifying
Pu, Y. Y., Sun, D. W., Buccheri, M., Grassi, M., Cattaneo, T. M. P., ripeness and forecasting harvest time of oil palm. Scientia Hor-
& Gowen, A. (2019). Ripeness classification of bananito fruit ticulturae, 265. [Link]
(Musa acuminata, AA): A comparison study of visible spec- Siva Kumar, A., Raja, S., Pritha, N., Raviraj, H., Babitha Lincy, R., &
troscopy and hyperspectral imaging. Food Analytical Methods, Jency Rubia, J. (2023). An adaptive transformer model for anom-
12(8), 1693–1704. [Link] aly detection in wireless sensor networks in real-time. Measure-
Raj, T., Hashim, F. H., Huddin, A. B., Hussain, A., Ibrahim, M. F., & ment: Sensors, 25. [Link] oi.o​ rg/1​ 0.1​ 016/j.m
​ easen.2​ 022.1​ 00625
Abdul, P. M. (2021). Classification of oil palm fresh fruit matu- Suharjito, Elwirehardja, G. N., & Prayoga, J. S. (2021). Oil palm fresh
rity based on carotene content from Raman spectra. Scientific fruit bunch ripeness classification on mobile devices using deep
Reports, 11(1). [Link] learning approaches. Computers and Electronics in Agriculture,
Robi, S. N. A. B. M., Izhar, M. A. B. M., Sahrim, M. Bin, & Ahmad, 188. [Link]
N. B. (2022). Image detection and classification of oil palm fruit Sunilkumar, K., & Sparjan Babu, D. S. (2013). Surface color based pre-
bunches. 4th International Conference on Smart Sensors and diction of oil content in oil palm (Elaeis guineensis Jacq.) fresh
Application: Digitalization for Societal Well-Being, ICSSA 2022, fruit bunch. African Journal of Agricultural Research , 8(6).
108–113. [Link] Syazwani, N., Nawi, M., Deros, B. M., Nizam, M., Rahman, A., &
Sabri, N., Ibrahim, Z., Syahlan, S., Jamil, N., & Mangshor, N. N. A. Sukadarin, E. H. (2016). Malaysian oil palm workers are in pain:
(2017). Palm oil fresh fruit bunch ripeness grading identifica- Hazards identification and ergonomics related problems. Malay-
tion using color features. Journal of Fundamental and Applied sian Journal of Public Health Medicine, 16(1), 50–57. [Link]
Sciences, 9(4S), 563–579. [Link] www.​resea​rchga​te.​net/​publi​cation/​29776​5530
Sabzi, S., Nadimi, M., Abbaspour-Gilandeh, Y., & Paliwal, J. (2022). Thang, Y. M., Ariffin, A. A., Appleton, D. R., Asis, A. J., Mokhtar,
Non-destructive estimation of physicochemical properties and M. N., & Yunus, R. (2017). Determination of sugars composi-
detection of ripeness level of apples using machine vision. Inter- tion in abscission zone of oil palm fruit. IOP Conference Series:
national Journal of Fruit Science, 22(1), 628–645. [Link] Materials Science and Engineering, 206(1). [Link]
org/​10.​1080/​15538​362.​2022.​20925​80 1088/​1757-​899X/​206/1/​012034
Sari, N., Shiddiq, M., Fitra, R. H., & Yasmin, N. Z. (2019). Ripeness Tzuan, G. T. H., Hashim, F. H., Raj, T., Baseri Huddin, A., & Sajab,
classification of oil palm fresh fruit bunch using an optical probe. M. S. (2022). Oil palm fruits ripeness classification based on the
Journal of Aceh Physics Society, 8(3), 72–77. [Link] characteristics of protein, lipid, carotene, and guanine/cytosine
24815/​jacps.​v8i3.​14122 from the Raman spectra. Plants, 11(15). [Link] 390/​
Septiarini, A., Sunyoto, A., Hamdani, H., Kasim, A. A., Utaminin- plant​s1115​1936
grum, F., & Hatta, H. R. (2021). Machine vision for the maturity United States Department of Agriculture. (2023a). Oilseeds: World
classification of oil palm fresh fruit bunches based on color and markets and trade. [Link]
texture features. Scientia Horticulturae, 286. [Link] USDAF​AS/​subsc​riber/​new
1016/j.​scien​ta.​2021.​110245 United States Department of Agriculture. (2023b). World agricultural
Setiawan, A. W., Mengko, R., Putri, A. P. H., Danudirdjo, D., & production.
Ananda, A. R. (2019). Classification of palm oil fresh fruit bunch Utom, S. L., Mohamad, E. J., Ameran, H. L. M., Kadir, H. A., Muji, S.
using multiband optical sensors. International Journal of Electri- Z. M., Rahim, R. A., & Pusppanathan, J. (2018). Non-destructive
cal and Computer Engineering, 9(4), 2386–2393. [Link] oi.o​ rg/​ oil palm fresh fruit bunch (FFB) grading technique using optical
10.​11591/​ijece.​v9i4.​pp2386-​2393 sensor. International Journal of Integrated Engineering, 10(1),
Setiawan, A. W., & Prasetya, O. E. (2020). Palm oil fresh fruit bunch 35–39. [Link]
grading system using multispectral image analysis in HSV. Vo, H.-T., Chau Mui, K., Nguyen Thien, N., & Pham Tien, P. (2024).
2020 IEEE International Conference on Informatics, IoT, and Automating tomato ripeness classification and counting with
206 Food and Bioprocess Technology (2025) 18:183–206

YOLOv9. In IJACSA) International Journal of Advanced Computer Measurement Science Review (Vol. 10, Issue 1, pp. 7–14). [Link]
Science and Applications (Vol. 15, Issue 4). www.​ijacsa.​thesai.​org doi.​org/​10.​2478/​v10048-​010-​0002-x
Wang, A., Qian, W., Li, A., Xu, Y., Hu, J., Xie, Y., & Zhang, L. (2024). You, K. Y., Wee, F. H., Lee, Y. S., Abbas, Z., Lee, K. Y., Cheng, E. M.,
NVW-YOLOv8s: An improved YOLOv8s network for real-time Khe, C. S., & Jamlos, M. F. (2020). A review of oil palm fruit
detection and segmentation of tomato fruits at different ripeness ripeness monitoring using microwave techniques in Malaysia.
stages. Computers and Electronics in Agriculture, 219. [Link] IOP Conference Series: Materials Science and Engineering,
doi.​org/​10.​1016/j.​compag.​2024.​108833 767(1). [Link]
Wong, Z. Y., Chew, W. J., & Phang, S. K. (2020). Computer vision Zolfagharnassab, S., Shariff, A. R. B. M., Ehsani, R., Jaafar, H. Z., &
algorithm development for classification of palm fruit ripe- Aris, I. Bin. (2022). Classification of oil palm fresh fruit bunches
ness. AIP Conference Proceedings, 2233. [Link] based on their maturity using thermal imaging technique. Agri-
1063/5.​00021​88 culture (Switzerland), 12(11). [Link]
Xiao, B., Nguyen, M., & Yan, W. Q. (2024). Fruit ripeness iden- e1211​1779
tification using YOLOv8 model. Multimedia Tools and Zulkifli, N., Hashim, N., Abdan, K., & Hanafi, M. (2019). Applica-
Applications, 83(9), 28039–28056. [Link] tion of laser-induced backscattering imaging for predicting and
s11042-​023-​16570-9 classifying ripening stages of “Berangan” bananas. Computers
Xing, Z., Ren, J., Fan, X., & Zhang, Y. (2023). S-DETR: A transformer model and Electronics in Agriculture, 160, 100–107. [Link]
for real-time detection of marine ships. Journal of Marine Science and 1016/J.​COMPAG.​2019.​02.​031
Engineering, 11(4). [Link]
Yee, C., Hashim, N., Maringgal, B., & Hafizz, M. (2020). A review Publisher's Note Springer Nature remains neutral with regard to
of non-destructive techniques applied for measuring quality of jurisdictional claims in published maps and institutional affiliations.
oil palm fresh fruit bunches. Journal of Agricultural and Food
Engineering, 1(1), 1–6. [Link]
Yeow, Y. K., Abbas, Z., & Khalid, K. (2010). Application of microwave
moisture sensor for determination of oil palm fruit ripeness. In