This approach assesses yield, marbling distribution, and the segmentation of rib–eye and bone. Gonçalves et al. (2021) employed a convolutional neural network (CNN)-based segmentation approach to identify carcass regions in smartphone-captured images, by first partitioning each image into superpixels and then classifying each segment as either carcass or background with 96.0% accuracy.

In the context of regression-based prediction, AI models have proven highly effective at estimating key carcass traits directly from image data. For example, Matthews et al. (2022) demonstrated the effectiveness of a 2-dimensional CNN model for accurately estimating the weight of specific cuts from a video analysis system. In a similar application for quality assessment, Lee et al. (2022b) developed a specialized CNN, mean and standard deviation-based ensemble network (or MSENet), to successfully predict marbling scores from red-green-blue (RGB) images taken at the intersection of the 12th and 13th ribs.

Research focusing on segmenting specific RoI within entire carcass structures has led to the development of specialized architectures such as Marbling-Net, which have demonstrated superior performance compared to standard baseline models like U-Net in precisely segmenting marbling regions (Zhang et al., 2023). Complementing these quality-focused applications, the versatility of segmentation techniques has been effectively demonstrated in anatomical analysis across multiple imaging platforms. For instance, CNN-based architectures have achieved successful detection of loin eye areas in ultrasound imaging (de Melo et al., 2022), while efficient encoder-decoder algorithms have been developed for real-time bone segmentation from X-ray images (Xu et al., 2023).

Beyond segmentation tasks, keypoint detection has emerged as another critical computer vision technique for meat processing applications, particularly from the perspective of meat automation systems. Manko et al. (2022) advanced robotic manipulation capabilities in meat processing by developing keypoint detection algorithms to identify critical anatomical landmarks on pig carcasses, enabling accurate prediction of limb orientation and identification of optimal gripping points for automated handling. Their methodology involved collecting RGB images from 6 different viewpoints to ensure comprehensive spatial coverage and robust keypoint localization across varying perspectives.

Specific cut analysis frequently focuses on IMF content and back-fat thickness prediction. Kvam and Kongsro (2017) employed ultrasound imaging to successfully predict IMF in pigs using a custom CNN model trained on preprocessed images that were normalized to unit standard deviation and augmented through flipping to enhance model robustness. Similarly, Liu et al. (2018) developed an approach to predict IMF categories in pork using a support vector machine (SVM) model applied to RGB images. To segment the RoI from the images, they employed a gray-level histogram combined with the Otsu method for background removal and binarization. Eighteen color features were subsequently extracted and applied to the SVM model, achieving a classification accuracy of 75.0% for IMF levels (grades 1–3), demonstrating the feasibility of this approach for practical implementation.

For IMF prediction, several studies have employed diverse data sources and AI methodologies with notable success. Shahinfar et al. (2019) utilized a comprehensive dataset including animal weights recorded from birth through slaughter (130–700 d) and 3 500 ultrasound images collected concurrently, with the RF model achieving superior predictive performance among the evaluated AI approaches. Kucha et al. (2022) employed near-infrared (NIR) spectra from vacuum-packed pork samples, preprocessing the spectral data using SD transformation and the correlated optimized warping (or COW) method to eliminate polyethylene film interference, with the SVM model subsequently demonstrating robust performance in IMF content estimation. Chen et al. (2022b) integrated computer vision scores derived from RGB images with conventional pork quality traits such as meat color, marbling score, pH value, and drip loss. The datasets underwent background removal, grayscale conversion, and RoI segmentation to derive the computer vision score, with the GB model achieving a Pearson correlation coefficient of 0.81, highlighting the potential of combining computer vision features with conventional quality metrics.

For back-fat thickness classification and prediction, several studies have developed AI-driven approaches with varying methodologies. Masferrer et al. (2018) systematically varied both the number and type of predictor variables to improve SVM model performance for categorizing ham into 4 thickness groups: thin (0–10 mm), standard (11–15 mm), semi-fat (16–20 mm), and fat (>20 mm), with the SVM model successfully classifying ham categories using variables including lean meat percentage, sex, and breed. Masferrer et al. (2019) employed the AutoFomIII system to collect characteristic data from 400 pig carcasses and successfully classified them into 4 categories (HC1: <9 mm, HC2: 9–12 mm, HC3: 13–19 mm, and HC4: >19 mm) based on subcutaneous fat thickness using an SVM model. Lee et al. (2022a) proposed a CNN-based architecture called back-fat thickness estimation network (or BTENet) for predicting back-fat thickness in slaughtered pigs. The model successfully performed both segmentation and thickness estimation simultaneously by cropping the back-fat region from original images.

Figure 4 summarizes the research trends in meat production studies. Overall carcass analysis has primarily focused on predicting carcass yield and segmenting loin eye areas and bone structures, while specific cut analysis has concentrated on IMF content and back-fat thickness prediction through regression and classification tasks. Most studies employ image-based approaches, utilizing segmented RoI from original images followed by various preprocessing techniques such as resizing and simple linear iterative clustering (SLIC) before applying CNN-based models for enhanced accuracy in specific anatomical characteristic analysis. Nevertheless, such approaches utilizing superpixel-based partitioning in carcass detection studies may result in computational limitations and reduced accuracy. Therefore, segmentation-based methodologies should be prioritized as they demonstrate significant advantages in both processing efficiency and classification precision.

Various meat quality parameters and other relevant chemical traits have been determined using AI-based ML approaches. Ndob and Lebert (2018) successfully predicted pH and water activity in pork by integrating data from a pH meter and an a_w meter with a deep neural network (DNN) model. Building on this foundation, Dixit et al. (2021) utilized NIR technology to predict pH and IMF levels in red meat using hyperspectral imaging (HSI) datasets from beef, lamb, and venison, with a custom CNN model achieving an R² of 0.89 for both parameters. In a related study focusing on specific fatty acid analysis, Wang et al. (2021) utilized NIR spectroscopy to quantify stearic acid content in lamb meat from 3 different cuts, with the least-squares (LS)-SVM achieving an RMSE of 0.18 and R² of 0.76.

El Karam et al. (2023) investigated NaCl concentrations ranging from 1.1% to 1.9% in pork muscle using ultraviolet spectrometry, with a SVM model achieving 97.6% classification accuracy. Advancing beyond single-parameter analysis, Cheng et al. (2023b) developed a multitask CNN model to predict lipid and protein oxidation in frozen-thawed pork, achieving R² values of 0.97 for thiobarbituric acid reactive substances (TBARS) and 0.96 for carbonyl content. Complementing this work, Cheng et al. (2023a) introduced a lightweight 3-dimensional (3D) CNN model specifically for TBARS prediction, achieving an RMSE of 0.03 and R² of 0.92. Furthermore, Cui et al. (2024) explored long-chain saturated fatty acid content prediction in lamb, beef, and pork using visible (Vis) NIR spectral data, with the CNN model achieving improved performance on generative interference adversarial autoencoder enhanced datasets (RMSE of 0.55 and R² of 0.72) compared to raw datasets.

Moving to visual and textural quality assessments, Sun et al. (2018) employed a computer vision system combined with AI models to classify color and marbling in pork, with the SVM model achieving 92.5% accuracy in color classification and 75.0% in marbling classification. Expanding on this multi-attribute approach, Tang et al. (2023) predicted 14 meat quality traits using Vis-NIR imaging combined with a DNN model, demonstrating high predictive accuracy for fat (R² = 0.72) and chemical lean (R² = 0.73), with moderate performance for other quality attributes including moisture, collagen, and color parameters.

In the realm of sensory evaluation, Wang et al. (2019) utilized an electronic (E) tongue sensor array to detect ions in beef samples, with the SVM model achieving 90.0% accuracy in flavor scoring. Similarly, Cui et al. (2022) employed comprehensive two-dimensional gas chromatography-mass spectrometry (GC × GC – MS) and a DNN to predict beef flavor based on Maillard reaction products, demonstrating accuracy exceeding 90.0% in predicting flavor scores. For texture analysis, Zhang et al. (2022b) applied NIR spectroscopy to predict texture properties of lamb loins, with the SVM model achieving exceptionally high predictive accuracy for hardness (R² =0.98), gumminess (R² = 0.98), and chewiness (R² = 0.98).

Quality defect detection has shown considerable progress through AI-driven approaches. Wold et al. (2017) investigated wooden breast syndrome in chicken breast fillets using NIR spectroscopy, with a linear discriminant analysis (LDA) model achieving 99.5% accuracy in defect detection. Building on this success, Geronimo et al. (2019) combined computer vision with NIR spectroscopy to classify woody breast abnormalities in chicken, with an SVM model achieving 97.5% accuracy using NIR spectral data. Beyond biological defects, Ahn et al. (2020) detected needles in meat using X-ray images, with their CNN model reaching 99.8% classification accuracy. Additionally, breed and production method classification has been explored by Alaiz-Rodríguez and Parnell (2020) who employed Fourier-transform infrared (FTIR) spectroscopy to differentiate suckling lamb carcasses raised on milk replacer from those raised on ewe milk, with the DNN-PCA approach achieving 93.8% accuracy. García-Infante et al. (2024) further demonstrated the potential of multidata integration by classifying lamb carcasses from 3 native Spanish breeds using organoleptic, sensorial, and nutritional data, with DNN achieving 88.0% accuracy when combining both datasets.

Storage condition assessment has become increasingly important, with studies focusing on distinguishing between frozen and thawed meat. Huang et al. (2016a) classified pork samples into 4 categories based on freezing history using NIR spectroscopy, with a DNN model achieving 93.3% accuracy. Similarly, Xu and Sun (2017) employed HSI to identify freezer burn in frozen salmon fillets, with a TreeBagger ensemble classifier, achieving 97.8% accuracy. E-nose technology has also proven effective in this area, as demonstrated by Górska-Horczyczak et al. (2017), who introduced an E-nose to differentiate fresh pork from frozen, thawed, and spoiled counterparts, with a DNN model achieving 85.4% average precision across 7 categories.

Complementing spectroscopic approaches, Abie et al. (2021) leveraged bioimpedance to assess different thawing methods on meat quality, with a long short-term memory (LSTM) model achieving classification accuracies of 91.6% for fast thawing and 95.0% for slow thawing conditions. Recent portable technology applications have shown promising results, with Swanson and Gowen (2022) examining poultry samples using portable Vis-NIR with multivariate AI methods, achieving 88.0% accuracy for detecting thawed poultry with an SVM. Park et al. (2023) combined HSI with AI to classify beef samples into 3 storage states, with an SVM model achieving 97.6% classification accuracy, demonstrating the potential for rapid, noninvasive assessments of meat storage conditions.

Figure 5 provides a concise overview of spectroscopic techniques for meat–quality assessment and distinguishes workflows by data format. Tabular spectral measurements are smoothed and feature–selected before predictive models such as DNN or SVM are trained. Conversely, HSI data undergo RoI extraction followed by analysis with CNN. Both pipelines support regression of physicochemical attributes (pH, moisture, acidity) and classification of product quality and storage conditions (normal vs. defective, frozen vs. thawed), underscoring the versatility of AI–enabled spectroscopy.

Freshness evaluation can be broadly categorized into freshness status analysis and microbe analysis. In freshness status analysis, the primary objective is to classify meat into discrete categories such as fresh, semi-fresh, and spoiled, with classification criteria varying considerably among studies. In contrast, microbe analysis focuses on predicting microbial counts generated during spoilage, enabling more precise evaluation of overall freshness.

A variety of sensor- and imaging-based approaches have been explored to categorize meat freshness. Hasan et al. (2012) employed an E-nose system to classify meat and fish samples into 4 categories, with SVM and K-nearest neighbors (KNN) models achieving accuracies of 94.5% and 96.2%, respectively. Building on this foundation, Haddi et al. (2015) combined an E-nose and E-tongue with AI models to classify beef, goat, and sheep freshness over 5 storage d, achieving 81.3% accuracy for E-nose data and 100% for E-tongue data. Vajdi et al. (2019) monitored fish headspace over a 15-d period using an E-nose device to classify samples into freshness categories, with a DNN model achieving 96.8% classification accuracy.

Expanding beyond electronic sensing, several studies have employed HSI techniques for freshness analysis. Huang et al. (2016b) compared NIR spectroscopy with computer vision systems to assess fish freshness, with a DNN model achieving 90.0% accuracy for image-based data and 93.3% for spectral data. Yang et al. (2017) utilized HSI data to classify cooked beef samples as fresh, medium fresh, or spoiled, achieving 97.1% accuracy with a LS-SVM. Ropodi et al. (2017) employed multispectral imaging (MSI) to distinguish beef from horse meat and detect adulteration, achieving 95.3% accuracy using an SVM. Alshejari and Kodogiannis (2017) investigated packaging conditions and storage temperatures on beef using multispectral images, with a DNN achieving 90.1% accuracy. Zhang et al. (2022a) advanced lamb freshness prediction using Vis-NIR imaging with chemical measurements, with a RF model achieving 91.0% accuracy for classifying lamb into premium, subfresh, and spoiled categories. Huang et al. (2023) integrated MOF@SnS₂ sensors with AI to determine beef storage duration, with a DNN model achieving 78.5% accuracy.

The accessibility of smartphone technology has led to extensive research utilizing RGB cameras for freshness analysis. Arsalane et al. (2018) evaluated beef freshness using RGB images, extracting color space features with an SVM model achieving 100% classification accuracy. Guo et al. (2020) employed a ResNet-101 model on smartphone images of chicken, fish, and beef to classify samples as fresh, less fresh, or spoiled, achieving 98.5% accuracy. Similarly, Amani and Sarkodie (2022) utilized a CNN model on meat images to differentiate fresh from spoiled meat, achieving 100% classification accuracy. Hewawasam et al. (2023) developed an integrated poultry management system using RGB cameras, with a MobileNet model achieving 100% accuracy in identifying chicken slaughter timing. Kim et al. (2023b) fused RGB images and air quality measurements using MobileNet and 1-dimensional (1D) CNN, achieving 99.4% classification accuracy for categorizing pork freshness. Kim et al. (2023c) employed a color sensor to measure hue-saturation-value (or HSV) values for pork classification, achieving 98.7% accuracy with a 1D-CNN.

Recent advances in computer vision have further enhanced freshness assessment capabilities. Elangovan et al. (2024) classified red meat freshness into 2 or 3 categories using RGB cameras, with ConvNet-18 achieving 99.4% accuracy for 2-category classification and ConvNet-24 achieving 96.6% for 3-category classification. Arsalane et al. (2024) predicted beef freshness by analyzing camera-acquired images using fast wavelet transform feature extraction, with KNN achieving 92.5% accuracy and SVM achieving 90.1%. Zheng et al. (2024) developed detect cells rapidly network (or DCRNet) for detecting and counting stained cells to assess meat quality, achieving an average precision score of 81.2% while reducing manual cell counting workload to less than 0.5%.

Some research efforts combine freshness classification with microbial activity estimation. Papadopoulou et al. (2013) employed a portable E-nose to assess beef fillet samples, with an SVM model achieving 90.4% classification accuracy and a correlation coefficient of 0.86 for predicting microbial counts. Kodogiannis et al. (2014) integrated FTIR spectroscopy with neuro-fuzzy modeling to classify beef samples while simultaneously estimating microbial counts, with their adaptive fuzzy logic system achieving 95.9% classification accuracy and a regression RMSE of 0.37. Gu et al. (2017) focused on lipid oxidation in Chinese style sausages by classifying 3 quality levels and performing regression analyses for acid value using an E-nose, with both DNN and SVM models achieving 100% classification accuracy and the SVM model attaining an R² of 0.98 for regression analysis. Kaswati et al. (2020) used Vis-NIR imaging to classify chicken samples as fresh or spoiled while predicting pH levels, with a RF model achieving 85.5% classification accuracy and partial least-squares regression model yielding R² values between 0.80 and 0.84 for pH prediction.

A substantial group of investigations has focused on estimating total volatile basic nitrogen (TVB-N), which is closely linked to microbial spoilage. Huang et al. (2014) employed NIR spectroscopy, computer vision, and E-nose techniques to predict TVB-N content in pork samples, with a DNN model combining all 3 data sources achieving an RMSE of 2.73 and R² of 0.95. Huang et al. (2015) proposed a NIR spectroscopy system with a BP-AdaBoost algorithm, achieving an RMSE of 6.94 and correlation coefficient of 0.83. Khulal et al. (2016) quantified TVB-N content in chicken using HSI, with a DNN model incorporating the ant colony optimization method achieving an RMSE of 6.38 and correlation coefficient of 0.75. Dai et al. (2016) predicted TVB-N content in prawns during cold storage using HSI data, with an LS-SVM model achieving an RMSE of 0.72 and correlation coefficient of 0.95.

Beyond TVB-N analysis, some approaches target additional spoilage indicators. Liang et al. (2019) employed terahertz spectroscopy to predict the K value in pork samples, with a BP-AdaBoost model achieving an RMSE of 9.89 and R² of 0.84. For total viable count (TVC) prediction. Fengou et al. (2020) combined FTIR spectroscopy and MSI data to evaluate TVC in meat samples, with integrated MSI and FTIR data yielding an RMSE of 0.88 and correlation coefficient of 0.83. Dourou et al. (2021) employed FTIR spectroscopy on chicken liver samples under various temperature conditions, with an SVM model achieving an RMSE of 0.72 and R² of 0.76 for noninoculated samples. Kolosov et al. (2023) integrated MSI data with a deep CNN model to measure TVC in pork samples, with ResNet-34 achieving MAE values of 0.05 for aerobically stored samples and 0.06 for modified atmosphere packaged samples. Lakehal and Lakehal (2023) collected beef samples across 6 freezing periods to predict storage time using Commission Internationale de l’Eclairage color space (or CIELAB) transformation, with a DNN model achieving an RMSE of 0.06 and R² of 0.97.

Figure 6 provides an overview of meat freshness evaluation using various equipment. In summary, meat freshness evaluation utilizes 2 main types of data collection: tabular data obtained from E-nose and spectroscopic sensors and image data captured through RGB imaging and HSI. For tabular data, the collected information was preprocessed through smoothing techniques (including standard normal variate [SNV], Savitzky-Golay smoothing [SG], and normalization) and dimension reduction using PCA. In addition, visual information of the meat was captured using RGB imaging. The images underwent further preprocessing through region of interest segmentation, augmentation (using rotations and flips), and resizing. These preprocessed data enabled the AI model to perform dual analytical tasks: classification of meat status (fresh, semi-fresh, and spoiled) and regression analysis of microbial content (TVC and TVB-N levels).

Meat authentication analysis typically involves predicting the extent of adulteration across different meat types or identifying specific muscle cuts. For meat type classification, the objective is to differentiate among pork, lamb, and beef based on their composition or to determine the level of admixture. For meat cut classification, the goal is to classify samples according to the distinctive features. Table 6 provides a summary of studies that apply AI to meat evaluation, with a specific focus on meat authentication.

Early research in this field utilized electronic sensing approaches. Tian et al. (2013) employed an E-nose to predict pork percentage in minced lamb and pork mixtures, with a DNN model achieving R² of 0.97 and RMSE of 5.26, outperforming multiple linear regression. Güney and Atasoy (2015) identified fish species using an E-nose with metal oxide gas sensors, with a 2-level binary decision trees (DT) model achieving 96.1% classification accuracy, surpassing naive Bayes (84.7%) and KNN (80.0%).

Expanding beyond E-nose approaches, several studies have employed spectral techniques with remarkable success. Alfar et al. (2016) authenticated and classified fats derived from beef, chicken, and lard using micro-NIR spectrometry, with a SVM model achieving 98.3% 3-class classification accuracy. Acquarelli et al. (2017) discriminated chicken, pork, and turkey products using FTIR spectroscopy, with a CNN model achieving 100% accuracy. Al-Sarayreh et al. (2018) classified lamb, beef, pork, and fat using HSI data, with a 3D-CNN achieving 96.1% accuracy and operating 4.7 times faster than SVM.

Adulteration detection has become increasingly important in meat authentication. Zhao et al. (2019) employed Vis-NIR techniques to detect adulteration in spoiled beef, with a LS-SVM model achieving an RMSE of 5.67 and correlation coefficient of 0.97. Al-Sarayreh et al. (2020) investigated snapshot HSI coupled with deep learning to classify red meat, with a 3D-CNN model achieving classification accuracies of 98.6%, 96.9%, and 97.1% on line-scanning, NIR, and visible snapshot datasets, respectively. Ayaz et al. (2020) used HSI to classify minced beef, chicken, and lamb, with an SVM model achieving 88.8% accuracy. Zhang et al. (2022c) introduced a CNN to predict pork content in adulterated lamb at various levels, achieving classification accuracies of 100%, 100%, and 99.9% for fresh, frozen-thawed, and mixed datasets, respectively.

Advanced spectroscopic techniques have further enhanced meat species identification capabilities. Robert et al. (2021) employed Raman spectroscopy to distinguish among beef, venison, and lamb, with an SVM model achieving 93.0% accuracy with a radial basis function kernel. Sun et al. (2022) integrated laser-induced breakdown spectroscopy and Raman spectroscopy to classify meat species (pork, beef, and lamb), with a DNN model achieving F₁-scores of 99.8% for beef, 99.4% for lamb, and 99.1% for pork.

Research focusing on meat cut classification has demonstrated the potential for precise muscle-specific identification. Sanz et al. (2016) classified 4 types of lamb muscles using HSI data, with a least-squares mean classifier achieving 96.6% accuracy. Li et al. (2021a) developed a classification model for 3 beef cuts (sirloin, shank, and flank) by integrating Vis-NIR imaging with AI, with a LDA model reaching 90.9% accuracy. Recent developments in computer vision have opened new avenues for meat authentication. Prakash et al. (2021) employed an ensemble AI approach that merged CNN, multinomial logistic regression, and DT to categorize 5 muscle types of meat using RGB images, with the ensemble model achieving 95.0% accuracy using grayscale images and 99.1% with original colored images. Huang et al. (2022) investigated the classification of 4 pork primal cuts (ham, loin, belly, and neck) using mobile phone images, with a ResNet-50 model enhanced with convolutional block attention module achieving 94.4% classification accuracy.

Figure 7 synthesizes spectroscopic meat–authentication studies that use tabular signals from E–nose and other sensors, along with HSI and RGB imagery. Tabular spectra are enhanced by smoothing (SNV, SG, normalization) and feature–selection methods such as PCA and improved whale optimization, while HSI/RGB images undergo superpixel– and SLIC–based segmentation to capture spatial context. These pipelines enable AI models to distinguish meat species (beef, pork, lamb) and cuts (neck, loin, ham) with high accuracy, underscoring the strength of AI–enabled spectroscopy for comprehensive meat authentication.

Chen et al. (2022a) forecasted monthly pork prices using a bi-recurrent neural network (RNN) LSTM model on historical price records, achieving a RMSE of 0.69 and MAPE of 3.36, outperforming RNN and standard LSTM models. Recognizing the importance of external factors, some researchers have incorporated additional variables to enhance predictive accuracy. Chuluunsaikhan et al. (2020) integrated pig-related news with historical price data in an LSTM model, reducing RMSE from 1 928 to 1 155 and MAPE from 8.4 to 5.1 compared to autoregressive integrated moving average with exogenous variables (or ARIMAX). Chen et al. (2021) embedded online news articles using a bidirectional encoder representations from transformer (BERT) model within a multi-BERT-LSTM framework, achieving an RMSE of 0.96 and MAPE of 0.17, significantly outperforming standard LSTM.

Building on this multifactor approach, Ma et al. (2019) integrated external features including soybean meal prices and macroeconomic indicators into AI algorithms, with a deep belief network model achieving an RMSE of 1.20 and MAPE of 7.13, outperforming autoregressive integrated moving average (ARIMA). Yang et al. (2021) compared various models using price data and macroeconomic indicators, with a hybrid ARIMA-LSTM model achieving the best performance (RMSE of 2.038). Suaza-Medina et al. (2023) focused on the Lleida market by incorporating regional market data, with ARIMA achieving an R² of 0.98, outperforming RNN and LSTM. Rahmani et al. (2024) expanded the dataset to include multiple economic indicators, with RF and AdaBoost achieving lower RMSE values of 0.16 and 0.16, respectively, compared to ARIMA.

Ko et al. (2023) proposed a deep learning-based framework to predict flavor and appearance preferences in pork. The study collected 1 767 flavor preference data points and 5 150 appearance preference data points, as well as pork characteristics measured using ultrasound equipment (i.e., AutoFomIII) and additional sex and breed information. The collected data were preprocessed by applying label encoding to the sex and breed features and normalizing each feature before being used in a stacking ensemble of linear models, RF, and a DNN. The final system achieved a MAPE of 11.17 for flavor and 5.02 for appearance on a scale of 1 to 5. Jeong et al. (2024b) further expanded this concept by integrating an AutoFomIII device to collect pork characteristics and conducting a taste preference survey, yielding 2 321 preference scores. Their deep learning-based stacking model reached a MAPE of 5.34, outperforming baseline DNN (7.53) and extreme GB (7.93) models. Additionally, SHapley Additive exPlanations values revealed that consumer age, gender, and pork fat properties significantly influence taste preferences. Overall, these studies underscore the value of combining historical data, external market indicators, consumer demographics, and advanced AI techniques to improve meat price forecasting and optimize product offerings according to consumer preferences.

These studies underscore the value of combining historical data, external market indicators, consumer demographics, and advanced AI techniques to improve meat price forecasting and optimize product offerings according to consumer preferences. Figure 8 presents an overview of the methodology for predicting consumer preferences using ultrasound measurement equipment. The study utilized the AutoFomIII device to collect comprehensive carcass measurements for this purpose. The resulting tabular data underwent several preprocessing steps, including feature selection to identify relevant measurements, label encoding for categorical variables (such as sex and breed), and normalization to smooth the data. The processed data were then analyzed using an ensemble approach that stacked multiple models and combined them through a meta-regressor that successfully predicted consumer preference scores for both flavor and appearance attributes, demonstrating the effectiveness of integrated measurement and AI approaches in understanding consumer preferences for meat products.