Table D.3.

Record overview table of the studies dealing with physico‐chemical methods to distinguish fresh fish (chilled or superchilled) from frozen/thawed fish

Method subgroup	Short description of method	Conditions tested: fish species	Conditions tested: temperatures and duration	Conditions tested: freezing methods	Short description of results	Performance classification	Appraisal	Reference
Spectroscopy
Combinations of UV‐VIS/NIR: FT‐NIR, FT‐MIR	FT‐NIR spectra were recorded at 12 cm−1 resolution (64 scans) from flesh of whole fillets at room temperature with FT‐NIR MPA spectrometer, fitted with integrating sphere (12,500–3,750 cm−1) and optical fibre (11,000–3,750 cm−1). FT‐MIR spectra (or else simply FT‐IR) were recorded with VERTEX 70, at 4 cm−1 resolution (16 scans) from minced skinless samples at room temperature, in the range 4,000 to 700 cm−1. Classification algorithm: soft independent modelling of Class Analogy (SIMCA) on 30 smoothed (standard normal variate or SNV, multiplicative scatter correction (MSC, second derivatives, etc.) spectral features selected by principal component analysis (PCA). Validation set was generated by algorithm that randomly selected 30% of the spectra composing the original data. The remaining spectra were used for calibrating the model.	Atlantic mullet (Pseudupeneus prayensis)	Fresh fish (4°C) for max 48 h. Frozen fish (–18°C) for up to 2 months, thawed at 4°C for 48 h.	Not specified	The spectral differences among fresh and frozen samples were due to O‐H bond of water in the FT‐NIR data and to amines and carboxylic acids in FT‐MIR data.	SIMCA model gave high classification performance as follows: ‐ FT‐NIR integrated sphere data, pretreated with MSC: 93–95% classification ability (i.e. correct classification), 88–89% prediction ability (i.e. ability to classify in one of the two classes), 68–73% sensitivity and 75–77% specificity. ‐ FT‐NIR optical fibre smoothed data: 87–92% classification ability, 82–87% prediction ability, 85–88% sensitivity and 57–60% specificity. ‐ FT‐MIR data, pretreated with MSC: 98% classification ability, 88–98% prediction ability, 60–70% sensitivity and 95–100% specificity.	11000100111 6	(Alamprese and Casiraghi, 2015)
Combinations of UV‐VIS/NIR: NIR, VIS/NIR	NIR spectra for whole and minced sample, were acquired with NIR spectrophotometer (FOSS Nirsystems 5000 Instrument) from 1,100 to 2,500 nm and 2 nm resolution in reflectance mode. VIS/NIR spectra, from 380 to 1080 nm, for whole fish were acquired with portable MMS1 spectrophotometer (Zeiss) equipped with fibre‐optic accessory at incidence angle of 45° in reflectance mode. Classification algorithms: (a) PCA on pre=treated data with SNV, smoothing and first/second derivative; (b) partial least squares regression discriminant analysis (PLS‐DA); and (c) Multivariate Logistic Regression.	Swordfish (Xiphias gladius	Fresh vacuum packaged (polyethylene) white muscle cutlets (male and female) at 2°C for variable storage durations up to 10 days. Rapidly frozen at –18°C and slowly frozen at –10°C, both for 30 days.	Slow freezing at –10°C, within 3.5 to 5 h. Fast freezing within 110 to 180 min at –18°C. Equipment not specified.	NIR: spectral differences were detected in the regions 1,388–1,594 nm and 1,700–2,498 nm, associated with free water, O‐H functional groups and water overtones. Cross‐validation showed the highest classification accuracy (96.7%) for the whole fish samples that were rapidly frozen. VIS/NIR: Principal bands at 430 and 550–570 nm accounted for spectral differences due to changes in myoglobin. Cross‐validation accuracy was 100%.	No classification performance with additional independent samples was performed	11110001010 6	(Fasolato et al., 2012)
Combinations of UV‐VIS/NIR: FT‐MIR	MIR spectra were recorded between 3,000 and 900 cm−1, at a 4 cm−1 resolution with FTIR spectrometer OMNIC software, mounted with an Attenuated Reflectance accessory (ATR: with ZNSe crystal, at incidence angle of 45°). Classification algorithm: normalisation of MIR spectral data to a value of 1. Normalised data were analysed with PCA separately on regions 3,000–2,800 cm−1, 1,700–1,500 cm−1 and 1,500–900 cm−1. Then, Factorial Discriminant Analysis (FDA) was applied to the first 5 PCs.	Whiting fillets (Merlangius merlangus)	Fresh (not specified). Frozen slow at –20°C or fast at –35°C and each thawed slowly at 0–3°C in cold room (7–10 h) or rapidly in microwave (7 min).	Slow freezing in cold room. Fast freezing in cooling cell.	The spectral region 3,000–2,800 cm−1 corresponds to C‐H bound of methyl and methylene groups of fatty acids. The 1,700–1,500 cm−1 region corresponds to the amide I and II bands and the 1,500–900 cm−1 region is the fingerprint region referring to C‐O and C‐C stretching modes.	FDA applied on the first 5 PCs gave overall correct classification rate of 75, 37.5 and 87.5% for the 1,500–900, 1,700–1,500 and 3,000–2,800 cm−1 spectral regions, respectively. Concatenation of the 5 PCs of all three regions into a single matrix resulted in 87.5% correct classification rate, based on cross‐validation, i.e. using random spectra from the calibration set, as opposed to using spectra from independent samples, i.e. not included in the original calibration set.	11000001010 4	(Karoui et al., 2007)
Combinations of UV‐VIS/NIR: VIS/NIR	The method was based on capturing NIR spectra of minced epaxial white muscle of 4 fish species (calibration data set), with two different instruments: (a) FOSS NIRSystem 5000, at 2 nm intervals from 1,100 to 2,500 nm and (b) a UNITY Scientific SpectraStar 2500TW at 1 nm intervals from 680 to 2,500 nm. Classification by PCA (capturing the main components) followed by PLS‐DA multivariate model.	Gilthead sea bream (Sparus aurata) Red mullet (Mallus barbatus) Sole (Sole vulgaris) Sword fish (Xiphias gladius)	Not specified, but referred to samples and their spectra from databases, from published studies, so a wide range of temperatures and freezing duration must have been considered.	Not specified	Three PLS‐DA‐based strategies were applied to untreated or SNV pre‐treated spectral data to differentiate fresh from frozen/thawed fish (with one calibration and one validation set): one strategy considered all fish together, while the other two could also classify samples according to fish species.	The study presents a NIR‐based method for differentiating fresh from frozen–thawed fish also with potential to and explicitly separate fish species. It has been validated against several fish species with classification accuracy between 78 and 92%, depending on the modelling strategy (i.e. dependent or independent of the fish species) and the fish species.	11100100101 6	(Ottavian et al., 2013)
Combinations of UV‐VIS/NIR: FT‐IR in combination with dimethylamine (DMA) estimation during storage	FT‐IR spectra were acquired with a Nicolet Magna 750 FT‐IR spectrophotometer equipped with a 40° incidence angle ZnSe crystal (referred to as ATR crystal), at 4,000 to 800 cm−1 with 4 cm−1 resolution at room temperature. Classification algorithm: discriminant analysis (DA) based on PCA scores and partial Least Square Regression correlating spectral data with changes in DMA during frozen storage.	Red hake (Urophycis chuss)	Samples comprised vacuum packaged ground white muscle of: ‐ refrigerated (fresh) samples ‐ Samples frozen for unknown duration in a domestic refrigerator at –10 to –12°C or in a Westinghouse upright freezer at –14 to –15°C. The conditions of thawing are not described.	Domestic freezer or Westinghouse upright freezer	Spectral differences caused by frozen storage, were evident in the regions 3,800 to 2,800 cm−1, associated with increase in free water, and from 1,700 to 800 cm−1, representing DMA and trimethylamine oxide (TMAO).	Randomly selected samples from the same batch (not used for calibration) were used for testing the classification performance (N=114). Percentage of correct classification was 86.0% for fresh fish, and 81.3–93.8% for frozen fish groups (i.e. with different DMA content)	11110110011 8	(Pink et al., 1998)
Combinations of UV‐VIS/NIR: UV‐VIS/NIR	Reflectance spectra were collected using a LabSpec 5000 (ASD Inc.) with a customised probe. The spectral range of the spectrophotometer was 350–2,500 nm, with spectral resolution of 3 nm at 700 nm, 10 nm at 1,400 nm and 2,100 nm. The scanning time was 10 ms and a sampling interval of 1.4 nm at the 350–1,000 nm range and 2 nm in the 1,000–2,500 nm range (wavelength accuracy ± 1 nm). Scanning took place before and after freezing/thawing. Classification by PCA (blooming detection) and double cross validation, using two calibration and 1 validation subs‐sets, coupled with PLS‐DA (classification of fresh vs. frozen/thawed).	Tuna (Thymnus thynnus)	Fresh samples: equilibrated from 4 to 16–18°C (1 h) and then scanned. Scanned samples kept at 4°C (2 h), wrapped in film and frozen at –80°C for 5, 21 and 35 day. Thawing took place at 4°C for 24 h.	Not specified	The method was capable in differentiating fresh fish from frozen for varying durations and then thawed. Discrimination was also based on the detection of (instrumentally detectable) blooming caused by myoglobin oxygenation due to initial exposure to the air. Fillets frozen below –60°C did not show detectable changes in visual characteristics when thawed, hindering discrimination from not frozen (fresh) samples.	The classification performance in the two categories, namely fresh and frozen/thawed samples were as follows: 92% probability that a fresh sample is predicted correctly as fresh and 82% that frozen/thawed is a correctly classified. The above are based on the validation subset that was not used to calibrate the model (i.e. calibration data set included two original subsets).	11000100011 5	(Reis et al., 2017)
Combinations of UV‐VIS/NIR: NIR	NIR spectra acquired by portable spectrophotometer poliSPECNIR (ITPhotonics), operated in a scanning range of 900–1,650 nm at 2 nm intervals, by direct contact of sample with the probe surface. Classification by PLS‐DA multivariate method on scaled (autoscale, SNV) variables. Model accuracy was assessed with cross‐validation and classification performance was assessed with a random validation set of independent samples.	Cuttlefish (Sepia officinalis)	Freezing at –25°C (3 h) then storage at –20°C for 5 ± 1 months. Thawing overnight (16–18 h) at temperature not specified. Fresh samples stored at 2°C for unspecified duration (but probably till end of shelf life because total viable count (TVC)/total volatile basic nitrogen (TVB‐N) were also monitored).	Not specified	20% randomly selected samples from the original data set of spectra, were left out to be used for model validation. The rest composed the calibration data set of the PLS model, which was cross‐validated. Spectral differences between fresh and frozen–thawed samples were detected in the region 960–980 nm which is associated with O‐H bonds, mainly related to the moisture content in the muscle. The discrimination between fresh and frozen was not affected by the duration of freezing.	75 to 90% of fresh samples of the fish species tested, were correctly classified as fresh.	11000110011 6	(Sannia et al., 2019)
Combinations of UV‐VIS/NIR: VIS/NIR	Scanning with NIRSystems 6500 spectrophotometer with a surface interactance fibre‐optic accessory. Spectra (optical density units in log 1/T, with T=the % of energy transmitted) recorded at wavelength range of 400–1100 nm at 2 nm intervals. Sample classification: SIMCA and LDA, using PCA. SIMCA relies on independent PCA per group, LDA uses PCA scores as input variables.	Red sea bream (Pagrus major)	Fresh fish kept in ice‐cold water for 30 min. Frozen fish (–40°C) for 30 days, thawed overnight at 5°C.	Not specified	Reflectance of fish samples was affected by freezing/thawing due to alterations in the physical structure of surface. Only one PC was enough to separate fresh from frozen–thawed samples. Original (untreated) absorbance data led to better classification than data treated with multiplicative scatter correction LDA had 100% classification efficiency as compared to 7% of SIMCA.	LDA models based on direct absorbance data may distinguish frozen–thawed sea bream from fresh with 100% accuracy. Classification accuracy was based on independent unknown samples, i.e. belonging to the prediction set, and not the calibration data set.	11000000011 4	(Uddin et al., 2005)
Combinations of UV‐VIS/NIR: NIR	Dry extract spectroscopy by infrared reflection (DESIR), was performed on the meat juice or natural drip from 15 g of fresh/frozen fish or fillets. Tissue was centrifuged and aliquots (0.6 mL) of extract/drip were collected via micro‐perforated discs on glass‐fibre filter paper, that was dried (30°C, 30 min) and scanned in diffuse reflectance mode in the 1,100 to 2,500 nm range, on the deposit side in a NIR spectrophotometer (model 6500, NIRSystems Inc.). Scanning was performed on a Standard Sample Cup, manually rotated in 5 steps. NIR reflectance spectra were recorded in 2 nm steps. Sample classification: (a) PCA and (b) a dummy regression by stepwise multiple linear regression (MLR) with dummy scores: 1 = fresh; 2 = frozen/thawed.	Horse mackerel (Trachurus japoncus)	Fresh fish, soon after killing, equilibrated at 20°C. Whole fish frozen at –40°C for 10 days. Frozen fillets at –40°C for 10 days.	Not specified	NIR spectra of fresh fish clearly differed from that of frozen/thawed and wavelength‐specific changes, associated with protein denaturation, were evident in the 2nd derivative spectra. Correlation coefficient of calibration > 0.95, with both multivariate methods.	Spectra were randomly assigned to calibration set and validation set. Classification performance was estimated merging spectra used for calibration (n=54) and validation (n=54). 100% of fresh (n=54) and frozen/thawed (n=54) fish samples were correctly classified.	11000000011 4	(Uddin and Okazaki, 2004)
Combinations of UV‐VIS/NIR: FT‐NIR	Scanning with NIRFlex N500 FT‐IR, equipped with a NIRFlex solid sample cell and indium gallium arsenide detector and spinning sampling module. Spectra recorded with 32 scans in the range from 10,000 to 4,000 cm−1 with 4 cm−1 resolution. Sample classification: PCA and Mahalanobis distance discrimination analysis.	Tilapia (Oreochromis)	Samples were dorsal and belly fillets. The objective was to discriminate samples subjected to consecutive freezing‐thawing cycles: ‐ one frozen cycle: frozen (–18°C) for 12 h and thawed for 12 h. ‐ two to seven frozen cycles as above.	Not specified	Differences between the NIR reflectance of fresh and frozen/thawed samples were attributed to stretching of the OH‐NH bonds at 5,100 and 6,825 cm−1. Samples were randomly divided to calibration and validation data sets. Better discrimination was achieved among the once and repeated freezing/thawed samples in the frozen state (without or with MSC pretreatment of spectral data) than in the thawed state. Accuracy in samples classification was better in dorsal (without pre‐treatment of spectral data) than in belly samples.	The observed maximum classification accuracy ranged from 80 to 93.3%, for samples, respectively, frozen once or repeatedly, and from 86.67 to 93.33% particularly for dorsal (frozen) samples.	11100110011 7	(Wang et al., 2018)
Hyperspectral imaging	The HSI system consisted of a line‐scanning imaging spectrograph (ImSpector V10E) covering the spectral range of 308–1,105 nm with the spectroscopic resolution of 2.8 nm; a charge‐coupled device (CCD) camera (DL‐604M) with the effective resolution of 1,004 × 1,002 pixels by 12 bits and its corresponding camera lens (OLE23) with the focal length of 23 mm; an illumination source including two 150 W halogen lamps (3900‐ER).	Grass carp	Group 1: fresh fish, group. Group 2: fish cold‐stored at 4°C for 7 days. Group 3: fish frozen at –20°C for 30 day and thawed at 4°C for 12 h. Group 4: fish frozen at –40°C for 30 day and thawed at 4°C for 12 h.	Not mentioned	Visible and near infrared hyperspectral imaging in the spectral range of 400–1,000 nm in tandem with classifiers and spectral preprocessing techniques showed good discrimination of fresh, cold‐stored (4°C) and frozen (–20°C and –40°C) thawed grass carp fish fillets.	Classification performance was 92.34%	11100100001 4	(Cheng et al., 2015)
VIS/NIR with hyperspectral imaging	A line‐scanning hyperspectral imaging system in reflectance mode was employed to capture the hyperspectral images for each prawn in the wavelength range of 300–1,100 nm. The main components of the system included a spectrograph (ImSpector V10E), a 12‐bit CCD camera (DL‐604 mol/L), two 500 W halogen lamps (3900‐ER), a conveying stage operated by a stepper motor (ST‐1212‐300), and a computer supported with spectracube data.	Prawns (Metapenaeus ensis)	Four groups (N=70) of prawns with different processing were included in this study, namely unfrozen–fresh, unfrozen–soaked, frozen–fresh and frozen–soaked groups. ‘Fresh’ refers to the prawns being treated with crashed ice to a sudden death, while the ‘soaked’ ones were subsequently soaked into 10 L of sea water (25 C) for 2 h. Peeled prawns were used ‘unfrozen’ or stored at –18°C for 2 months (‘frozen’)	Not specified	The results had Charnes, Cooper, and Rhodes (1978) (CCR) of 98.33% for least‐squares support‐vector machines (LS‐SVM) model for distinguishing fresh and frozen–thawed prawns based on spectral variables. However, the average CCR only reached 73.33% for differentiating unfrozen–fresh group from unfrozen–soaked group.	The remaining (independent) 60 samples of 30 fresh and 30 soaked prawns formed the prediction set. Classification performance was 98.33%.	11000100001 4	(Dai et al., 2015)
VIS/NIR with hyperspectral imaging VIS/NIR with portable equipment	Two spectral scans: 1) Interactance imaging (hyperspectral) with imaging spectrometer (VNIR‐640) with a field of view 1 mm × 300 mm and spatial resolution 1 mm × 0.5 mm. Each pixel represents radiation in the region of 400–1,000 nm with 10 nm resolution. Scanning speed (on white diffuse conveyer) of 400 nm. 2) Interactance spectroscopy with handheld probe (XDS Optiprobe Analyser), composed of transmitting and receiving fibre bundles. Spectral data recorded as absorbance units at 400–2,500 nm and resolution of 0.5 nm. Classification algorithm: PCA applied on SNV and Savitzky–Golay second derivative pretreated spectral data. K‐nearest neighbour classifier with leave‐one‐out cross validation was used for classification. Separation line between classes was calculated using the Rosenblatt's perceptron.	Atlantic salmon (Salmo salar)	Fresh fillets (2–4°C), covered with ice. Fresh, stored in ice for 0 and 2 days, then frozen at –40°C for 3 weeks, thawed overnight at 2–4°C (fillets) or under running water (whole fish, which was then gutted and filleted).	Not specified	The spectral region 605–735 nm, represented on the 1st PC, is appropriate for discrimination of fresh from frozen–thawed fillets, based on 2nd derivative pretreated data. Differences are associated with spectral changes due to oxidation of haem proteins during freezing–thawing and chilled storage in ice.	Classification rate varied on the fillets, from 60 to the maximum of 100%, the latter in the tick loin region. Better discrimination of frozen/thawed samples from fresh samples was obtained for fillets compared to whole fish. Segmented cross validation was applied, by leaving spectral data of one day out for validation and calibrating the model with the remaining one. The procedure was applied to all days.	11000110011 6	(Kimiya et al., 2013)
UV‐VIS/NIR with hyperspectral imaging VIS/NIR hyperspectral imaging with digital camera (RGB imaging)	VIS/NIR spectra were collected without sample pretreatment using a portable instrument (VIS/NIR diode array, Hamamatsu S3904) scanning wavelength from 300 to 1,100 nm in transflection mode at 2 nm intervals. RGB images (red, green and blue colour space) were collected using a compact digital camera Kodak EasyShare M530 (4,000×3,000 pixels) within a 60×60×60 cm3 photographic box screened from the environmental light with opaque walls and equipped with four fluorescent tubes to control the illumination. Classification by PCA (capturing the main components), followed by PLS‐DA multivariate model on scaled variables	West African goatfish fillets (Pseudupeneus prayensis)	Freezing at –32°C, then storage at –18°C for up to 48 h. Thawing overnight (18 h) at 4°C, sampling immediately or after another 24 h at 4°C. Fresh samples were kept at 2°C.	Not specified	The NIR and RGB imaging differentiated (correctly classified) frozen/thawed from fresh fish with 98–100% sensitivity (Se) and 100% specificity (Sp) immediately after thawing. Se of the two methods reduced to 60–80% after 24 h of chilled storage. Sp of NIR remained high (100%) after additional storage, but Sp of RGB imaging reduced to 80%. Combining the two methods, increased Se and Sp to almost 100% even after 24 h of storage.	A combination of NIR with RGB could differentiate frozen/thawed from fresh fish both immediately after thawing and after 24 h. The method was tested against a single species and the results could be affected by seasonality. 3 data sets were used, split in 1 for calibration and 2 for validation. Calibration of model against data sets with spectra from different seasons is recommended to increase the robustness of classification.	11000010011 5	(Ottavian et al., 2014)
Hyperspectral imaging	Three in‐house developed line‐scan hyperspectral imaging systems were used to collect four types of image data from fish fillet samples: (1) reflectance images in VIS/NIR region, (2) fluorescence images by 365 nm UV excitation, (3) reflectance images in short‐wave infrared (SWIR) region, (4) Raman images by 785 nm laser excitation.	Red snapper (Lutjanus campechanus),	14 fillets including: ‐ fresh (imaged immediately at 20°C). ‐ 1st freezing cycle (the fillets were frozen in a −20°C for 24 h and then thawed in a 4°C for 24 h). ‐ 2nd freezing cycle (same freezing and thawing process).	Ordinary kitchen freezer	For VIS/NIR, Raman, and fluorescence, the fresh fillets were more easily misclassified as frozen/thawed fillets in the first cycle rather than those in the second cycle. This suggests there is a progressive change in the fish tissue associated with the freeze‐thaw process.	Independent data sets and fish samples were used for performance classification. The highest accuracies were achieved at 100% using full VNIR reflectance spectra for the species classification and 99.9% using full SWIR reflectance spectra for the freshness classification.	11000101111 7	(Qin et al., 2020)
Hyperspectral imaging	The hyperspectral images were captured in reflectance mode, which involved a spectrograph (ImSpector V10E), a high‐performance camera (DL‐604M) in a high resolution of 1,004x1,002 (spatial × spectral) pixels coupled with a camera lens, an illumination unit involving two 150 W halogen lamps (2900‐ER), a conveyer belt motivated by a stepper motor and a computer. The working spectral range was 328–1,115 nm.	Shrimps (Metapenaeus ensis)	Fresh: 79 shrimps. Chilled: 80 shrimps stored at 4°C for 80 h. Frozen–thawed: 65 shrimps frozen at –20°C for 90 h and thawed at 25°C for 2 h.	Not specified	The results indicated that the selected optimal wavelengths had a strong discriminative ability for fresh, chilled and frozen–thawed shelled shrimps. The most effective wavelengths (783, 689, 435, 416, 813, 639, 452 and 478 nm) were decided by UVE (PLS)‐SPA to be utilised for model establishment. Principal components (PCs) images explained more than 99% of variances of all spectral bands.	No performance classification	11001100000 4	(Qu et al., 2015)
Hyperspectral imaging	A Micro‐Hyperspec VNIR camera with 17 mm lens was used to acquire hyperspectral images. Spectral data was collected from 400 to 1,000 nm, with spectral resolution of 5 nm. Firstly, hyperspectral images were obtained from both sides of intact fish (with scales). Then, fish scales were removed totally, and hyperspectral imaging measurement was applied on both sides of the scaled intact fish. After that, hyperspectral images were measured on both sides of each fish fillet, including flesh and skin sides.	Crucian carp (Carassius auratus)	1st group: 6 fresh fish. 2nd group: 18 fish were stored at 4 °C for 1, 2, and 3 days. 3rd group: 16 fish kept at −20°C for 7, 14, and 21 days.	Ordinary freezer	The PLS‐DA classification results for fresh and frozen/thawed fish showed that PLS‐DA models can identify fresh and frozen/thawed intact fish well. Cross‐validation of the models developed showed an accuracy of: 100% for intact scaled fish; 92.5% for intact fish with scales; 97.5% for fish fillets (flesh side); and 97.5% for fish fillets (skin side).	No performance classification	11000101000 4	(Shan et al., 2018)
VIS/NIR with hyperspectral imaging	Interactance imaging: using an imaging spectrometer (VNIR‐640), two custom made fibre optic light lines (200 mm long, powered by 150 W aluminium coated halogen lamps), with focusing acrylic rod lenses and two black painted aluminium screens for light baffling. Illumination adjacent and parallel to the detectors field of view. Focal distance 1,000 mm and field depth 25 mm. Handheld interactance probe: single beam spectrometer XDS Optiprobe analyser (FOSS) with transmitting and receiving fibre bundles. Measurements on middle of the loin and in the tail region. Data were expressed as absorbance units from 400 to 2,500 nm with 0.5 nm resolution. Classification algorithm: spectral data (32 pixels/spectra) were pre‐treated with SNV and 2nd derivative using Savitzky–Golay with a 2nd order polynomial and 50 nm wide smoothing window. Classification was made by PCA and Rosen‐blatts perceptron coupled with a sigmoid response function was used to separate classes in the PCA space. K‐Nearest neighbour classifier using the Euclidean distance was applied for discrimination of fresh from frozen–thawed samples.	Atlantic cod (Gadus morhua)	Fresh gutted iced fish sent by plane and stored in ice for up to 13 days. Fillets with skin, without skin and whole fish were frozen at –40°C on day 0 and after 2 days of storage. Frozen samples were kept for 2 weeks and then thawed overnight at 4°C.	Not specified	Spectral differences between fresh and frozen/thawed samples were evident around 507 and 636 nm, as well as the PC loading of wavelengths 487 and 646 nm, all associated with the absorption by the oxidised haem pigments (e.g. methaemoglobin and metmyoglobin). The average correct classification rate was 87.8, maximum classification (96–100%) obtained in the loin and close to the centre line of the fillet.	Classification algorithm was cross‐validated, i.e. without independent samples	11010010010 5	(Sivertsen et al., 2011)
Hyperspectral imaging	The illumination source was a pair of fibre‐optic line lights, each 200 mm wide and powered by three 150 W halogen lamps. Cylindrical lenses mounted in front of the line lights focused the light into two 10 mm thick parallel lines, 40 mm apart. The fibre optic line lights and detector were mounted at heights of 150 mm and 1,030 mm, respectively, above the conveyor belt. The hyperspectral camera operated in the VIS‐NIR range from 430 to 1,000 nm. Imaging of the samples was performed on a conveyor belt travelling at 40 cm/s.	Atlantic cod	200 loin pieces with two modes of sample freezing and sample thawing: fast freezing at –40°C and slow freezing at –20°C followed by thawing in 4°C circulating water (fast) and at 4°C by gently circulating air in a climate controlled cabinet (slow).	Blast freezing at –40°C (fast) and freezing in still air at –20°C (slow).	The two freezing methods showed clear separation in the first principal component. The two groups could be classified with 100% accuracy. Using the 450–600 nm region, the –40°C samples were predicted with 98% accuracy and the –20°C samples with 99% accuracy. The samples could be classified into once‐frozen and twice‐frozen with good accuracy, with success rates ranging from 97% to 100%.	The performance classification for fresh samples achieved 100%, for once‐thawed 98% and for twice‐thawed 93% for full spectrum.	11100001001 5	(Washburn et al., 2017)
VIS/NIR with hyperspectral imaging	The VIS/NIR hyperspectral imaging system was employed to capture hyperspectral images of fillets in reflectance mode. The system consists of a spectrograph, a 12‐bit CCD camera with a C‐mount 23‐mm lens having 672×512 (spatial × spectral) pixels, two 150 W tungsten halogen lamps for illumination, a conveyer belt driven by a stepping motor, and a computer with data acquisition and preprocessing software. The spectral resolution is 2.8 nm in 380–1,030 nm.	Halibut (Psetta maxima)	108 fillets, of which: 48 fresh and frozen fillets; 30 fillets stored at the constant temperature of −70 and −20°C for fast freezing (FF) and slow freezing (SF); 72 samples of 32 fresh and 40 frozen–thawed after 9 days (20 FF‐T and 20 SF‐T).	Slow and fast freezing	The general trends of the spectral curves for three categories of fish were similar. However, the overall absorbance level was found to decrease in FF‐T samples. Spectral characteristics of FF‐T were more similar with fresh fish, and the total absorbance level of FF‐T was higher than SF‐T fish. The spectral changes at 970, 729, 836, 928, 552, 512 and 620 nm among fresh, FF‐T, and SF‐T fish made the differentiation more well founded.	From the same batch of fish samples, 16 fresh samples and 20 frozen–thawed samples (10 FF‐T and 10 SF‐T) formed the prediction set. Correct classification rate was 93.75% for fresh and 90% for frozen–thawed fish (average 91.637%).	10100101001 5	(Zhu et al., 2013)
(Front‐face) fluorescence spectroscopy	Fluorescence spectra acquired with FluoroMax2 spectrofluorimeter, with a variable angle front‐surface accessory and incidence angle of the excitation radiation at 56°. Emission spectra of tryptophan residues (305–400 nm with 191 measurements at 0.5 nm increment) was recorded at excitation wavelength 290 nm. Emission spectra of nicotinamide adenine dinucleotide (NADH) (360–570 nm with 211 measurements at 1 nm increment) was recorded at excitation wavelengths 340 nm. Classification algorithm: PCA followed by FDA on the first 5 PCs.	Whiting fillets (Merlangius merlangus)	Fresh fillets (unspecified storage temperature and duration). Frozen at slow or fast rate to –20°C, kept for 8 days, then thawed fast (0°C in the centre after 7 min) in microwave, or slowly in cold room at 0–3°C for 7–10 h.	Slow freezing in cold room at –20°C Fast freezing in cooling cell at –35°C	Classification rate was calculated based on both calibration and independent validation spectra Tryptophane fluorescence spectra had a maximum for fresh fish at 326 nm and at 330 nm for frozen–thawed. NADH emission spectra of fresh had a maximum at 455 nm and a shoulder at 403 nm. The corresponding maximum emissions for frozen–thawed fish were observed at 379 and 455 nm. The correct classification rate based on NADH was 100%.	Correct classification rate between fresh and frozen fish was 70.8% based on tryptophane, and 100% based on NADH.	11010011011 7	(Karoui et al., 2006)
(Front‐face) fluorescence spectroscopy	Fluorescence spectra acquired directly on the fish muscle using a Fluoromax‐4 spectrofluorometer (Jobin Yvon). The incidence angle of the excitation radiation was set at 60° to ensure that reflected light, scattered radiation, and depolarisation phenomena were minimised at 20°C. The emission spectra of tryptophan residues (305–450 nm), NADH (360–600 nm) and riboflavin and other unknown fluorescent compounds (405–650 nm) were recorded with the excitation wavelengths sets at 290, 340 and 380 nm, respectively. The excitation spectra of vitamin A (250–390 nm) were obtained after emission wavelength set at 410 nm. Classification by canonical correlation analysis, coupled with FDA.	Sea bass (Dicentrarchus labrax)	Fresh fillets: up to 13 days at 4°C. Sequence of freezing, thawing and refrigerated storage: i) frozen–thawed (3 months at –18°C), then refrigerated up to 9 days at 4°C, and ii) refrigerated (up to 9 days at 4°C), then frozen–thawed (3 months at –18°C) fillets.	Not specified	The discrimination of fresh from frozen samples was primarily based on changes in the emission of tryptophane and NADH at ~368 and 468 nm, respectively, although the choice of multiple wavelengths was overall shown to affect classification performance.	Depending on the excitation set (290, 340, 380, 410 nm), correct classification of groups corresponding to different sequences of freezing/thawing refrigerated storage duration ranged from 83.33 to 94.87%.	10000110011 5	(Karoui et al., 2017)
NMR/MRI spectroscopy	All 1H MRI images were acquired using a Bruker BMT imaging console connected to a 2.35 Tesla, 31 cm horizontal bore super‐conducting magnet. A cylindrical eight strut bird‐cage radio‐frequency (RF) probe, internal diameter 9.4 cm, was used in quadrature mode to transmit and receive the magnetic resonance (MR) signal from the probe.	Cod and mackerel	Fresh fish. Frozen at –18°C in a domestic freezer for periods of 2, 4, 8 or 12 weeks and then thawed to room temperature 4 h.	Not specified	The MR parameters of both cod and mackerel were sensitive towards freeze‐thawing and it was possible to differentiate between fresh and frozen–thawed fish for both species. There was a significant difference at the 1% level (p < 0.01) between the fresh sample and that after freeze–thaw treatment.	No performance classification	10000100100 3	(Nott et al., 1999a)
NMR/MRI spectroscopy	All 1H MRI images were acquired using a Bruker BMT imaging console connected to a 2.35 Tesla, 31 cm horizontal bore super‐conducting magnet. A 20 cm gradient set built ‘in‐house’ with each axis powered by a Techron gradient amplifier provided gradient strengths up to 100 mT m21. An ‘in‐house’ built, cylindrical, eight strut, bird‐cage RF probe, internal diameter 9.4 cm, was used in the quadrature mode to transmit and receive the MR signal.	Trout	Fast frozen fish (4) at –18°C for 4 weeks and then thawed. Slow frozen fish (4) at –18°C for 4 weeks and then thawed.	Fast freezing in liquid nitrogen followed by storage in a domestic freezer at –18°C	Quantitative MRI provided parameters that were sensitive to the effects of freeze/thawing and to the method of freezing and duration of frozen‐storage. For fish which had been previously frozen, the change in the MR parameters after repeat freeze–thawing were smaller than those observed for the initial 2 day freeze–thaw of fresh fish, indicating a method which could be used to allow for the variability observed for fresh trout.	No performance classification	10000101000 3	(Nott et al., 1999b)
NMR/MRI spectroscopy (NMR)	TCA extracts of the salmon fillets and the NMR samples were analysed by NMR. 1D 1H NMR spectra of all TCA extracts were acquired at 300 K on a Bruker Avance 600‐MHz spectrometer equipped with a 5‐mm z‐gradient TXI (H/C/N) cryoprobe. The NMR data were acquired with the Bruker pulse sequences noesygppr 1d, NS = 48 and RG = 144.	(Farmed) Atlantic salmon	Fresh fish kept at 4°C and sampled on day 5, 6, 7, 9, 11, 14, 18. Frozen at −40°C for 16 h and thawed before sampling on day 5, 6, 7, 9, 11, 14 and 18. Fish frozen at −20°C and kept chilled at 4°C and both stored in sealed plastic bags at 4°C for 14 days. Sampling on day 10, 11, 12 and 14.	Not specified	The score plot showed a distinct fresh – thawed samples grouping according to NMR data upon performing PCA (74% PC1). Based on spectral regions, inferences can be made about changes in certain metabolites/compounds, e.g. aspartate, fumartate and phenylalanine. Thawing induces the formation of aspartase (new metabolite), due to mitochondrial enzymatic activity. Identification of aspartase could be a reliable indicator of thawing.	No performance classification Extended incubation at 4°C (e.g. > 3 day) seems to enhance differentiation between thawed and never frozen (fresh) sample. Study with a single species and under unspecified freezing conditions.	11100100000 4	(Shumilina et al., 2020)
Raman spectroscopy	Raman spectra of fat samples (preheated at 50°C in waterbath) were obtained with a DeltaNu Examiner Raman Microscopy with a 785 nm laser source and charge‐coupled device (0°C). Integration time of 15 s and 100 mW laser power. Spectra ranged from 200 to 2,000 cm−1 with 2 cm−1 resolution. Gas Chromatography (GC) was applied for estimating changes in fatty acid compositions. Classification model: Chemometrics with PCA.	Horse mackerel (Trachurus trachurus) European anchovy (Engraulis encrasicolus) Red mullet (Mullus surmuletus) Bluefish (Pomatamus saltatrix) Atlantic salmon (Salmo salar) Flying gurnard (Trigla lucerna)	Separate models for discrimination of fish species and differentiating fresh from once or twice frozen–thawed fish. Fish were washed and filleted. Fresh fish, kept refrigerated at 4°C for 12 h. Frozen fish at –18°C for 24 h, thawed for 12 h at 4°C. Frozen–thawed at –18°C for 24 h and thawed for 12 h at 4°C then re‐frozen/thawed again.	Deep freezer	PCA was efficient in discriminating fish species and fresh from once or twice frozen–thawed fish of either species, based on Raman spectra data. Changes in of spectra data were attributed to differences in the content of fatty acids and the structure of lipids.	Classification and discrimination efficiency of the proposed method was evidently high, but illustrated only graphically, without providing performance indicators. Accurate quantitative expression of classification performance is hindered, or not feasible.	10000110110 5	(Velioglu et al., 2015)
Electrical parameters
Electrical impedance	The system and sensor for measuring impedance in fish samples consisted of a software application an electronic equipment. For each one of the frequencies the electronic equipment generated the corresponding sinusoidal voltage waveform and applied it to the electrode.	Atlantic salmon	3 fillets on each day: ‐ Fresh (day 0). ‐ F1 (frozen group 1) Frozen at –18°C and thawed at 4°C for 24 h: on day 15, 30, and 60. ‐ F2 (frozen group 2) Fish samples submitted to 2 freezing cycles.	Shelf‐freezer equipped with T‐type thermocouples (0.1 mm diameter).	Four discriminant functions were obtained, the two first functions (F1 and F2) explaining more than 97% variance (F1 90.80% and F2 6.51%). 71.93% of cases were correctly classified. F1 determined the separation of fresh (day 0) from frozen–thawed samples. No separation between the frozen–thawed samples depending on the storage time or freezing cycles was observed.	No performance classification. Just a PCA, followed by a DA were conducted with existing samples.	10000111000 4	(Fernandez‐Segovia et al., 2012)
Electrical impedance	The impedance system consisted of an electronic equipment and a software application that runs on a computer. The software application performed a frequency sweep, obtaining the impedance modulus and the phase of the sample for a configured range of frequencies (50 frequencies between 1 Hz and 1 MHz). The impedance measurements were carried out by inserting the sensors into the sample perpendicular to the muscular fibres of the fish.	Sea bream (Sparus aurata)	Fresh samples (3). Frozen samples at −18°C and analysed on day 15, 30, and 60 after thawing at 4°C for 24 h (27). Samples subjected to a 2nd freeze/thaw cycle: on day 15, 30 and 60 after repeated freezing thawing at 4°C for 24 h (3).	Ordinary freezer	The module and phase impedance spectra of fresh samples and frozen/thawed sea bream obtained with the electrode AH were similar, which means that this electrode could not differentiate the different types of samples. No differences in the spectra of samples submitted to 2 freezing cycles compared with only 1 cycle were observed.	No performance classification. Just principal component analysis (PCA), followed by a discriminant analysis (DA), was conducted with data obtained with the electrodes used (internal samples).	00000111000 3	(Fuentes et al., 2013)
Electrical impedance	The impedance equipment consisted of a computer, the impedance instrument (CHI660E), and an electrode (10 mm apart, 1 mm in diameter) composed of four gold‐plated copper needles. The sinusoidal voltage of 10–1–105 Hz is released by the workstation as the excitation signal. The impedance measurements were made by inserting the electrodes into the fish muscle fibres at an angle of 90° and ensuring that the electrodes were fully introduced into the sample. Fish fillets were placed in small refrigerators and connected to external instruments by wires.	Salmon fillets and rainbow trout	Fresh salmon and trout to distinguish the fish species. Chilled salmon stored at 4 ± 0.5°C during 12 days and taken each 3 days. Frozen–thawed salmon frozen at –18 ± 0.5°C for 3 days and thawed at 4 ± 0.5°C.	Not specified	Atlantic salmon/rainbow trout, chilled/frozen–thawed salmon and fresh/stale salmon could be distinguished quickly with a 100% recognition accuracy achieved in training set and prediction set.	60 salmon fillets were selected for freshness detection and the performance classification displayed 100% accuracy for the discrimination between fresh and frozen–thawed fish samples.	10000000111 4	(Sun et al., 2020)
Electrical impedance	The HP 4284A Precise LCR meter was used to measure the electrical impedance of fish samples. Impedance magnitude (|Z|) and impedance phase (u) were measured at twenty‐seven frequencies from 100 to 1 MHz (0.1, 0.2, 0.5, 0.8, 1, 2, 5, 8, 10, 15, 20, 30, 50, 60, 71.4, 80, 85.7, 100, 150, 200, 300, 500, 600, 666.66, 800, 960 and 1,000 kHz) under constant current of 0.2 mA.	Sea bass (Dicentrarchus labrax)	50 fish samples: ‐ 1st group – fresh (control) after 48 h after the catch. ‐ chilled and stored for 3 and 6 days at 8°C. ‐ frozen 1 month at –18°C with/without temperature fluctuations and thawed at +4 °C. ‐ frozen 4 months at –18°C with/without temperature fluctuations and thawed at +4°C. ‐ 2nd freeze‐thawed cycle.	Freezing chamber	The results showed that there was a difference between the control and frozen–thawed groups in |Z| measured at low and medium frequencies, but there was no difference among the frozen–thawed groups in |Z| at none of the frequencies measured.	No performance classification	00000100000 1	(Vidacek et al., 2012a)
Electrical impedance	The HP LCR‐Meter‐4284A was used to measure resistance and reactance. The electronic device operated on 19 frequencies from 1 Hz to 1 MHz (0.1, 0.2, 0.5, 1, 2, 5, 10, 15, 20, 50, 80, 100, 200, 300, 400, 500, 800 and 1,000 kHz), measuring electrical properties by the constant current method (0.2 mA). Measurement was configured in a two‐electrode format. The fish samples were placed on a wooden isolator, and the electrodes were inserted in the muscle tissue.	Atlantic chub mackerel (Scomber colias)	Fresh fish samples. Samples on first freeze‐thaw cycle: one group was frozen at –20°C (slow freezing) and the other group by immersion in liquid nitrogen at temperature –196°C (fast freezing). Both groups were stored for 14 days at –20°C and then thawed by air at 4°C. Samples also underwent a second freeze‐thaw cycle using the same parameters as above.	Freezer with a natural airflow at –20°C	Thawed Atlantic chub mackerel samples previously frozen with different freezing rates had different reactance at frequencies higher than 150 kHz. Thus, the different reactance allowed discrimination of fresh/frozen samples.	No performance classification	10100101000 4	(Vidacek et al., 2012b)
Dielectric properties (Microwave)	Measuring average muscle anisotropy and electrical impedance anisotropy (microwave reflection coefficient) of the muscle surface with a 10 GHz microwave probe (active area 2 cm), placed parallel and perpendicularly to the muscle at 9 points Classification method: based on the average and variance of anisotropy distribution muscle fibres.	Salmon	Fresh fillets stored for 3, 4 and 10 days at 5°C. Fresh samples after storage as above subjected to fast (in alcohol) and slow (in air) freezing at –20°C, then thawed at 5°C for 12 h.	In alcohol (fast) In air (slow)	Fresh muscle is more anisotropic than frozen and lean more than fat tissue. Freezing and fat tissue reduce the average muscle anisotropy and increase the variation of anisotropy along the muscle surface, shifting the overall anisotropy distribution to the left of the fresh muscle anisotropy distribution.	No classification performance data are presented. Differentiation of samples is illustrated only visually though the partially (and not entire) overlapping anisotropy distributions. As such, accurate (and reproducible) quantitative classification performance indicators cannot be deduced, neither a classification threshold for extent of distribution overlapping, e.g. specific percentile, or else.	00000111010 4	(Clerjon and Damez, 2007)
Dielectric properties (Torry meter and k‐values)	Torry meter readings taken with GR Torry meter above the lateral line (skin side) and on the dorsal muscle (bone side) in the individual mode. K‐values determined for 5 g of dorsal muscle extracted with 19% perchloric acid, targeting ATP and its degradation products in perchloric acid. Classification algorithm: no statistical analysis, but only threshold values for the output of the two methodologies.	Yellowtail (Seriola quinqueradiata)	Fresh fish in ice for up to 18 days. Frozen at –20°C for 18 h, in freezer or dipped in liquid nitrogen for 30–40 min followed or not by storage in freezer.	Not specified	Fresh samples had higher total mixed ratio (TMR) (10–11 on skin and approx. 15 on bone side) and very low (close to 0) K‐values. Chill or frozen storage reduced TMR and increases K‐values. Liquid‐Nitrogen frozen samples behaved similarly to frozen samples kept in the freezer. Freezing affects the dielectric properties of the muscle.	Accurate classification performance cannot be deduced from graphs or table of the paper.	10001011010 5	(Kim et al., 1987)
Differential Scanning Calorimetry (DSC)
DSC	A heat‐flux differential scanning calorimeter (DSC‐60, Shimadzu) was used to perform Differential Scanning Calorimetry (DSC) analysis. Muscle samples (~30 mg) were sealed in an aluminium capsule and scanned from 30 to 90°C at a rate of 5°C per min, against an empty reference capsule. After baseline subtraction in each thermogram, total denaturation enthalpy, myosin and actin peak denaturation enthalpy, were estimated.	Seabream (Sparus aurata)	Samples covered with ice and stored on ice for 7 days at 4–5°C. Single‐frozen samples at –20°C and stored for 40 days. Double‐frozen samples at –20°C for an additional 7 days after thawing for 8 h.	Not specified	PCA explained 53.7% of the total observed variation and clearly separated the samples in three clusters without overlap: the first group contained the fresh fish samples, the second referred to the single‐frozen fillets and the third to the double‐frozen treatment. In this work, seabream single frozen fillets had significantly higher total denaturation enthalpy than double frozen fillets.	No classification performance	00000100000 1	(Matos et al., 2011)

CCD = charge‐coupled device; CCR = Charnes, Cooper, and Rhodes; DA = Discriminant Analysis; DESIR = dry extract spectroscopy by infrared reflection; DMA = dimethylamine; DSC = differential scanning calorimetry; FDA = Factorial Discriminant Analysis; FF = fast freezing; FT = Fourier transform; GC = Gas Chromatography; IR = infrared; LDA = Linear Discriminant Analysis; LS‐SVM = least‐squares support‐vector machines; MLR = multiple linear regression; NADH = nicotinamide adenine dinucleotide; PC = principal component; PCA = principal component analysis; PLS‐DA = partial least squares regression discriminant analysis; Se = sensitivity; SIMCA = Soft independent modelling of Class Analogy; Sp = specificity; SWIR = short‐wave infrared; SNV = standard normal variate; MSC = multiplicative scatter correction; MIR = middle‐infrared; MR = magnetic resonance; MRI = magnetic resonance imaging; NIR = near‐infrared; NMR = Nuclear magnetic resonance; RGB = red, green, and blue colour space; RF = radio‐frequency; SF = slow freezing; TMAO = trimethylamine oxide; TMR = total mixed ratio; TVB‐N = total volatile basic nitrogen; TVC = total viable count; u = impedance phase; UV = ultraviolet; UV‐VIS = ultraviolet–visible; UV‐VIS/NIR = ultraviolet–visible/near‐infrared; VIS/NIR = visible/near‐infrared; (|Z|) = impedance magnitude.