Comparison of deep learning models in automatic classification of coffee bean species

Introduction

Coffee is cultivated in more than 70 countries and is prepared by roasting and grinding the beans of tropical trees belonging to the Coffea genus of the Rubiaceae family (Buonocore, Carratu & Lamberti, 2022). With an annual global turnover exceeding 10 billion dollars, coffee stands out as the product with the second highest economic value after crude oil (Sanders, 2019). This beverage has played a pivotal role in shaping global production and trade relations on a global scale with its economic value and rich nutritional content for centuries. In recent years, interest and demand for coffee have increased significantly and has become a beverage that all segments of society can easily access. The increasing demand for coffee has led to the development of the coffee industry.

This type of beverage has started to become a part of the daily consumption habits of billions of people due to its unique taste and aroma, positive effects on health and energizing effect (Haile, Bae & Kang, 2020). Coffee, one of the most consumed beverages after water and tea, is one of the most profitable agricultural products in the world (Buonocore, Carratu & Lamberti, 2022). Consumers’ coffee preferences vary greatly depending on the taste, aroma, quality and processing method of the coffee, as well as the origin and type of beans (Do Carmo et al., 2020; Unal et al., 2022). The type of coffee bean has become a critical issue to consider, as it plays an important role in the trade and consumption processes (Mihailova et al., 2022). Therefore, correctly identifying the type of coffee beans in particular plays a critical role in ensuring quality control in production processes and determining market prices, as well as in terms of the sustainability of trade by increasing the reliability of the product offered to the consumer (Hung, Lee & Lin, 2021). Accurate and effective coffee bean grading is essential to ensure consistency, quality control and standardization in the coffee industry. Often coffee quality is determined by expert judgement, visual inspection or traditional grading methods with limited functionality (Turi, Abebe & Goro, 2013). However, traditional sorting methods used in the coffee industry are often time-consuming, inefficient and costly, as well as prone to human error (Hung, Lee & Lin, 2021; Unal et al., 2022). To address these challenges, improve quality assessment, and achieve high-accuracy classification, automatic detection of coffee beans using digital image processing has emerged as a critical necessity. Deep learning techniques, in particular, offer rapid, non-destructive, and highly effective solutions for image classification, making them a promising approach for automating coffee bean classification. Despite their advantages, research on deep learning applications in coffee bean classification remains limited. There is a clear need for further investigation and integration of these findings into industrial production processes (Ansar et al., 2021). In this context, the aim of the research can be listed as follows:

• To compare the performance of different deep learning models in coffee bean classification.

• To identify the most successful model(s) in terms of classification accuracy.

• To evaluate the applicability of deep learning models for automatic coffee bean classification.

• To compare the performance metrics of various deep learning models using coffee bean image datasets.

This study makes a significant contribution to the literature on coffee bean classification. Most existing studies focus on classical machine learning algorithms or basic deep learning models. In contrast, this research provides a comprehensive comparison of both conventional and next-generation deep learning architectures. Specifically, the study employs five state-of-the-art convolutional neural network (CNN) models—Xception, DenseNet201, InceptionV3, InceptionResNetV2, and DenseNet121—to classify coffee beans. The deep learning architectures used in this study were chosen because they provide high accuracy and generalization capabilities in visual classification tasks. The features of these models, such as deep connections and dense information sharing between layers, allow them to perform strongly in challenging tasks such as classifying visually diverse objects such as coffee beans. The performances of the models were compared and the obtained results were compared with the existing studies in the literature, and a new approach was presented in the classification of coffee bean species. This study provides a significant contribution to the literature in terms of demonstrating the effectiveness of new generation deep learning models that have not been used before for the classification of coffee bean species. In addition, by examining the usability of these methods in industrial applications, it provides a new perspective for the improvement of automatic separation and quality control processes of coffee beans. Therefore, the study can have a significant impact not only in the academic field but also in practical and industrial applications. Comparing the performance of deep learning models and determining the most suitable model is considered an important step towards automating and improving quality control processes in the coffee industry. The main contributions of the study are:

• Comprehensive comparison of five advanced CNN models—Xception, DenseNet201, InceptionV3, InceptionResNetV2, and DenseNet121—in coffee bean classification.

• Application of transfer learning techniques to improve model performance on limited datasets.

• Development of an innovative framework for classifying three distinct coffee bean species (Espresso, Kenya, and Starbucks Pike Place), addressing limitations of prior studies with small and diverse datasets.

• In-depth analysis of CNN-based models, exploring their advantages and limitations in coffee bean classification within the context of visual recognition.

These contributions aim to help better understand deep learning methods for visual classification tasks and demonstrate the effectiveness of their applications in the food and agriculture field. The following sections, in turn, will present the related works, the definitions of models, the performance evaluations.

Materials and Methods

This section provides detailed explanations of the dataset, deep learning models, and performance evaluation criteria used in the study. The flow chart of this study is presented in Fig 1.

Dataset

The dataset used in the present study consists of coffee images obtained from an external source. It is publicly available and was taken from Koklu’s (2024) website. Detailed information about the dataset, including its collection method, structure, and categorization, can be found in the source by Unal et al. (2022). This dataset includes three types of coffee beans: Espresso (Ethiopia), Kenya (Kenya) and Starbucks Pike Place (Mexico, Costa Rica, Colombia). These types of coffee are both commercially widely used and have different characteristics. Espresso is known as a high-quality coffee and is known for its unique flavor in Kenya. Starbucks Pike Place represents a widespread brand in the global market. It is expected that the classification between these types will contribute to the quality control and type identification processes in the coffee industry (Unal et al., 2022). Figure 2 shows the distribution of images in this dataset according to each class.

The dataset comprises a total of 1,554 coffee bean images, consisting of 530 Espresso, 502 Kenya, and 522 Starbucks Pike Place beans.

Convolutional neural network

CNNs are widely preferred for classifying visual data, such as coffee beans, due to their superior ability to capture spatial relationships within image data. CNN is a deep learning architecture that is frequently used in image processing, natural language processing, and computer vision (Gu et al., 2018; Acikgoz, Korkmaz & Talan, 2024). CNNs, which are effective in applications such as face recognition, automatic labeling, autonomous vehicles, and intelligent medical treatments, are designed to learn spatial relationships and features in visual data (Li et al., 2022).

CNNs typically consist of three primary types of layers. The first is the convolutional layer, which generates feature maps by applying various kernels to images, enabling the model to learn input features. This process facilitates the detection of low-level features such as edges, corners, and patterns. The second is the pooling layer, which reduces the resolution of feature maps to more compact and manageable sizes. To achieve this, techniques such as max pooling or average pooling are commonly used. The final component is the fully connected layer, responsible for high-level reasoning. This layer processes the extracted features to perform tasks such as classification or regression (Gu et al., 2018). The integration of these components enhances CNNs’ ability to learn complex and hierarchical image features, thereby improving overall performance. The structure of a classical CNN is presented in Fig. 3.

In the study, Xception, DenseNet201, InceptionV3, InceptionResNetV2 and DenseNet121 deep learning architectures were preferred for the classification of coffee beans. Each model was selected due to their strong representation capabilities and high performance that can be achieved with transfer learning. Models such as Xception, InceptionV3 and InceptionResNetV2 offer the ability to learn features at different scales thanks to their deep direct connections and inception structures. This provides a great advantage especially in terms of recognizing the visual diversity and detailed features of coffee beans. DenseNet201 and DenseNet121, on the other hand, provide more efficient information transfer by intensively sharing the information from the previous layers in each layer and facilitate the learning of deep networks. These structures are known for their strong generalization capabilities and low error rates while increasing the classification accuracy. The different structural advantages of the models, combined with transfer learning, enable us to achieve high accuracy rates in visual recognition tasks such as the classification of coffee beans. This diversity made it possible to obtain more robust and generalizable results by comparing performance between different architectures. These models represent deep learning structures that are frequently preferred for transfer learning. Transfer learning, which is widely used in fields such as speech recognition, natural language processing and computer vision, is a method where a model uses previously learned related knowledge in another similar task more quickly and effectively (Lu et al., 2015). This approach is especially useful in cases where the amount of data is limited or the training process takes a long time. In this research, the CNN network was trained using the transfer learning technique and using previously trained models.

Results

The networks used for transfer learning were selected among the models trained on ImageNet dataset: Xception, DenseNet201, InceptionV3, InceptionResNetV2 and DenseNet121. In this study, the pre-trained weights of these networks were taken and applied to the classification of coffee beans. The performance of the models was evaluated with accuracy, precision, recall and F-1 score metrics obtained from the confusion matrix of each CNN model. The cross-validation method applied to Xception, DenseNet201, InceptionV3, InceptionResNetV2 and DenseNet121 models analyzed the classification performance between Espresso, Kenya and Starbucks Pike Place coffee bean species. Table 4 presents the performance analysis performed based on the confusion matrix of each model.

The Xception model demonstrated high classification accuracy for the Espresso and Starbucks Pike Place classes, with low misclassification rates. However, it exhibited a higher misclassification rate for the Kenya class compared to the other classes. The DenseNet201 model, on the other hand, achieved the highest accuracy especially in the Espresso class and provided the lowest misclassification rate. This model, which also performed well in the Kenya class, made more misclassifications in the Starbucks Pike Place class, and therefore its performance seems lower than other classes. The InceptionV3 model generally exhibited high accuracy rates in all classes. This model attracted attention by making less misclassifications, especially in the Kenya class, compared to other models. In the Starbucks Pike Place class, although its performance was lower than other models, the number of misclassifications was lower. The InceptionResNetV2 model stands out with its low misclassification rates in the Starbucks Pike Place class. However, this model showed a moderate performance in the Espresso and Kenya classes. On the other hand, the DenseNet121 model stands out as one of the models that provide the highest accuracy rates in the Espresso and Kenya classes. However, the correct classification rate of this model is low and the misclassification rate is high in the Starbucks Pike Place class.

In general, the DenseNet201, InceptionV3 and DenseNet121 models showed the best performance and had the lowest misclassification rates in the Espresso class. In the Kenya class, the DenseNet121 model showed the highest performance and had the least misclassifications. In the Starbucks Pike Place class, the Xception and InceptionResNetV2 models showed better performance than the other models and achieved lower misclassification rates.

The high misclassification rate in the Starbucks Pike Place class indicates that this class may have features that are harder to distinguish from others or that there may be more confusion elements for this class in the dataset. These findings help us understand the strengths and weaknesses of the models in certain classes and provide clues about improvements that can be made to improve the classification performance.

Accuracy, precision, recall and F-1 score values of five different models (Xception, DenseNet201, InceptionV3, InceptionResNetV2 and DenseNet121) for Espresso, Kenya and Starbucks classes are shown in Table 5.

The Xception model achieved high performance in the Espresso class, with an accuracy of 0.96, precision of 0.98, and sensitivity of 0.96. The F1-score is 0.97, indicating that the model exhibited a well-balanced performance. In the Kenya class, the accuracy value is 0.83, the precision value is 0.97, and the sensitivity value is 0.83. This shows that the model keeps the rate of false positives low, while making more errors in false negatives. The F-1 score of 0.89 indicates that the model has reasonable performance in this class. In the Starbucks Pike Place class, with high accuracy (0.96) and sensitivity (0.96), but lower precision (0.82), the model generally made correct classifications in this class, but produced false positives to a certain extent. The F-1 score is 0.88. The DenseNet201 model performed well overall, with very high accuracy (0.98) and sensitivity (0.98) in the Espresso class, but slightly lower precision (0.90). The F-1 score was 0.94. In the Kenya class, accuracy was 0.89, precision and sensitivity were 0.93 and 0.89, respectively. This indicates a balanced performance with slightly higher false negatives. The F-1 score was 0.91. In the Starbucks Pike Place class, accuracy was 0.83, precision and sensitivity were 0.87 and 0.83, respectively. The F-1 score was 0.85, indicating that the model performed moderately in this class. The InceptionV3 model performed very strongly in the Espresso class, with very high accuracy (0.98), precision (0.95) and sensitivity (0.97). The F-1 score was 0.96. In the Kenya class, accuracy, precision and sensitivity rates are all 0.93, indicating that the model has a balanced and high performance in this class. The F-1 score is also 0.93. In the Starbucks Pike Place class, accuracy (0.88), precision (0.91) and sensitivity (0.88) rates have shown a good performance in this class. The F-1 score is 0.89. In the Espresso class, the InceptionResNetV2 model has high accuracy (0.95), precision (0.97) and sensitivity (0.95) values, and has generally shown a successful performance. The F-1 score is 0.96. In the Kenya class, accuracy (0.85) and sensitivity (0.85) values are slightly lower than the precision (0.96) value. The model reduced false positives in this class, but produced some false negatives. The F-1 score is 0.90. In the Starbucks Pike Place class, accuracy (0.93) and sensitivity (0.93) are high, while precision (0.83) is slightly lower. The F-1 score is 0.88, indicating good performance in general.

The DenseNet121 model is very successful in the Espresso class with very high accuracy (0.98), sensitivity (0.98) and precision (0.93) values. The F-1 score is 0.96. In the Kenya class, accuracy (0.96), precision (0.90) and sensitivity (0.96) rates indicate that the model is very successful in this class. The F-1 score is 0.93. In the Starbucks Pike Place class, accuracy (0.83), precision (0.94) and sensitivity (0.83) rates indicate that the model performs well in this class, but false negatives are more common. The F-1 score is 0.88.

All models showed high performance in the Espresso class in general. DenseNet201, InceptionV3 and DenseNet121 models were particularly prominent in the Espresso class. In the Kenya class, the InceptionV3 model exhibited the most balanced and highest performance. In the Starbucks Pike Place class, the accuracy rates were generally lower than in the other classes. Xception and InceptionResNetV2 stand out among the models with the highest performance. These differences reveal the capacity of the models to cope with different difficulty levels for each class. The low precision rates in the Starbucks Pike Place class indicate that false positives are more common in this class than in the other classes.

Table 6 presents the training times for the five CNN architectures—Xception, DenseNet201, InceptionV3, InceptionResNetV2, and DenseNet121—measured in minutes. The results highlight notable differences in the computational efficiency of these models. Among the tested architectures, InceptionV3 demonstrated the shortest training time, completing the process in 70.4 min, which reflects its lightweight design and efficient architecture. This is a significant advantage, particularly for real-time or resource-constrained applications.

On the other hand, DenseNet201 required the longest training time at 317.6 min, reflecting the model’s complexity and the increased number of parameters that need optimization. Similarly, InceptionResNetV2 and Xception exhibited longer training times of 225.9 min and 174.8 min, respectively, indicating their computationally intensive nature compared to InceptionV3. The DenseNet121 model, with a training time of 157.5 min, balances computational efficiency and model complexity, making it a viable alternative for scenarios requiring moderate computational resources.

These findings underscore the trade-off between computational cost and performance, as observed in the study. While InceptionV3 provides the most computationally efficient solution with excellent classification performance (93% accuracy), models such as DenseNet201 and InceptionResNetV2 require significantly more training time. This highlights the importance of selecting models based on the specific needs of the application, balancing computational efficiency with accuracy and other performance metrics.

In conclusion, the comparison of training times reaffirms InceptionV3 as the most efficient and practical model for real-world applications, particularly when computational resources are limited. However, for use cases where high resource availability is not a constraint, models like DenseNet201 and InceptionResNetV2 may be explored for their potential advantages in handling more complex feature representations.

The comparison of performance metrics of all classification models used in the study is presented in Fig. 4.

Figure 4 provides a comprehensive comparison of the performance of five deep learning models—Xception, DenseNet201, InceptionV3, InceptionResNetV2, and DenseNet121—across key metrics, including accuracy, precision, recall, F1-score, and MCC. Among these, InceptionV3 demonstrates the highest performance, achieving the best scores across all metrics, with an accuracy, precision, recall, and F1-Score of 93%, and an MCC of 89%. DenseNet121 closely follows, also achieving an MCC of 89%, making it a competitive option for applications requiring high accuracy with lower computational costs. Xception and InceptionResNetV2 exhibit slightly lower but consistent performance across all metrics, while DenseNet201 shows relatively weaker performance, particularly in MCC (85%), suggesting it may be less effective in handling class imbalances. These results highlight InceptionV3 as the most reliable model for coffee bean classification while showcasing DenseNet121 as an efficient alternative for resource-constrained applications. The figure underscores the importance of balancing classification accuracy and computational efficiency in selecting the optimal model for specific use cases.

ROC curves allow you to evaluate the discriminative properties of models. Figure 5 shows the ROC curves of the models used.

Table 7 presents the pairwise comparisons of the classification models using McNemar’s test to assess the statistical significance of their performance differences. The test was conducted by analyzing discrepancies in predictions across all classes (Espresso, Kenya, and Starbucks Pike Place). For each model pair, the test statistic and p-value were calculated, as shown in the table. The results indicate that none of the pairwise comparisons yielded statistically significant differences in performance, as all p-values exceed the commonly used significance threshold (p < 0.05). For instance, the comparison between InceptionV3 and DenseNet121, two of the top-performing models, resulted in a p-value of 1.000, demonstrating that the observed differences in accuracy and other metrics between these models are not statistically significant. Similarly, comparisons involving Xception and other models, such as DenseNet201 and InceptionResNetV2, also revealed no significant differences, with p-values consistently above 0.05.

The lack of statistical significance across all comparisons suggests that while the models exhibit slight numerical variations in performance metrics (e.g., accuracy, precision, recall), these differences are not large enough to conclusively favor one model over another. This finding aligns with the overall trend observed in the evaluation metrics, where all models demonstrated strong and comparable classification capabilities.

From a practical standpoint, the absence of significant differences implies that the choice of model can be influenced by other factors, such as computational efficiency, resource constraints, or ease of deployment, rather than purely numerical performance. For example, InceptionV3, with its highest accuracy (93%), may still be the preferred choice for applications prioritizing classification performance, whereas DenseNet121, with its computational efficiency, might be better suited for resource-limited environments.

Overall, Table 7 reinforces the robustness of all tested models and highlights the importance of considering additional factors beyond performance metrics when selecting a model for real-world applications. Future studies with larger datasets or additional statistical analyses may provide further insights into potential performance distinctions. Table 8 presents the comparison of experimental studies.

The comparison table illustrates the varying degrees of success achieved by different machine learning models in the classification of coffee beans. EfficientNet-B0, as demonstrated by Balakrishnan Jayakumari et al. (2024), achieved a remarkable accuracy of 100%, suggesting that even lightweight models can excel when applied to smaller, well-structured datasets. However, in contrast, models such as Rivalto, Pranowo & Santoso (2020), which reported an accuracy of 74.26%, highlight the challenges of using smaller datasets and less sophisticated CNN architectures.

Our study stands out in this comparative analysis, especially with InceptionV3, which achieved an impressive accuracy of 93%. This performance places our model among the top-performing architectures, rivaling those used in other studies, such as Buonocore, Carratu & Lamberti (2022) with YOLO (94%) and Hsia, Lee & Lai (2022) with LDCNN (98.38%). The strength of our study lies in the successful application of transfer learning using pre-trained models on the ImageNet dataset, which has proven to be highly effective in tackling the complex task of coffee bean classification across multiple classes. Moreover, DenseNet121 and Xception, which also exhibited strong performance in our study, further validate the robustness of our model selection.

Another important observation is the application of transformer-based models, such as the Swin Transformer used by Izza & Kusuma (2024), which achieved an accuracy of 84.75%. Although transformer architectures are increasingly being adopted in image classification tasks, our findings suggest that more traditional CNNs like InceptionV3 and DenseNet201 still hold a competitive edge, particularly in complex, multi-class classification problems such as ours.

The significance of our study becomes more apparent when considering the broader context of coffee bean classification research. The diversity of models and datasets in the literature highlights the challenges inherent in achieving high accuracy, particularly when classifying multiple coffee bean types. Our study not only demonstrates the efficacy of cutting-edge CNN architectures but also offers valuable insights into the classification of challenging datasets, such as Starbucks Pike Place, which presented notable classification difficulties across various studies.

In conclusion, our study contributes meaningfully to the field by demonstrating the high potential of advanced CNN models like InceptionV3 and DenseNet121 in achieving robust classification accuracy, even in challenging multi-class scenarios. These further cements the importance of model selection and dataset management in the pursuit of optimal classification performance. The findings from our study align with the broader trends observed in existing literature yet underscore the unique contributions our approach makes in pushing the boundaries of coffee bean classification research.

Discussion

This study explores the classification of three different coffee bean species—Starbucks Pike Place, Espresso, and Kenya—using deep learning models trained with transfer learning. The results highlight the significant impact of model selection on classification performance, particularly in the context of visually similar coffee species.

The InceptionV3 model achieved the highest classification success rate, with an overall accuracy of 93%. This model outperformed DenseNet121 (92%) and other models such as Xception and InceptionResNetV2 (both at 91%). InceptionV3 demonstrated superior performance across multiple evaluation metrics, including accuracy, precision, recall, F-1 score, and MCC. Furthermore, the ROC curve analysis reinforced its excellent discriminative ability, making InceptionV3 the most reliable model for coffee bean classification in this study.

One of the most challenging aspects of this study was the classification of Espresso and Starbucks Pike Place beans, which exhibited considerable visual similarities. As evidenced by the confusion matrix, these two coffee species were more frequently misclassified compared to Kenya beans. This challenge is likely due to overlapping visual features, such as similar color tones, textures, and bean sizes, which may have created ambiguity in the models’ learned feature representations. Additionally, variations in lighting conditions, bean orientations, or subtle intra-class differences may have further contributed to these misclassifications. These findings highlight a common limitation in visual-based classification tasks, where models relying solely on image features may struggle to differentiate between classes with high visual resemblance. Addressing this issue could involve incorporating non-visual features, such as chemical composition data from spectroscopy, to complement image data and enhance feature differentiation. Furthermore, advanced data augmentation techniques, such as domain-specific transformations, could increase training data diversity, enabling the models to better generalize to subtle inter-class variations. These approaches, combined with further refinement of model architectures, could significantly improve classification accuracy for visually similar coffee types and overcome the challenges observed in this study.

Another notable finding is the overall effectiveness of CNN-based models, particularly when trained via transfer learning. All five models tested in the study provided strong classification results, with relatively low misclassification rates. This demonstrates the potential of deep learning architectures to handle complex classification tasks in the coffee industry. The success of these models suggests that combining CNN architectures with accessible commercial digital scanners could pave the way for more efficient and cost-effective methods of coffee bean classification. These holds promise for non-destructive, real-time identification of coffee bean species, benefiting both producers and consumers.

This study highlights the potential of artificial intelligence technologies in agriculture by examining the use of deep learning methods in the automatic classification of coffee bean species. The findings revealed the performance differences of various model architectures and identified the strengths of each model in terms of classification accuracy. However, transferring these findings to real-world applications could significantly increase efficiency and quality in the sector. Automatic classification of coffee bean species can minimize human errors and increase overall efficiency in the sector, especially by speeding up quality control processes. For example, such systems can increase final product quality by allowing coffee producers to quickly distinguish and classify high-quality beans according to quality. Furthermore, considering the scalability of these systems, high accuracy can be achieved at lower costs by digitizing quality control processes in agriculture. However, there are some challenges to the successful application of this model in practice. In particular, the generalizability and accuracy of the model should be increased under different environmental conditions and for different coffee bean species. In order to be applied to wider areas, it is necessary to test it with datasets from different geographical regions and to evaluate the sensitivity of the model to environmental factors. In addition, by integrating automatic classification systems, it will be possible for small-scale producers in the sector to benefit from these technologies. In future studies, testing this model under various conditions and training it with larger, more diverse data sets will further strengthen its success in the field of application. Thus, quality control processes in coffee production can be made more efficient and competitive advantage can be achieved in international markets.

An important consideration for future development involves the integration of transformer-based approaches, such as Vision Transformer (ViT) and Swin Transformer, which have shown promise in image classification tasks through their ability to capture global features. While these models may require larger datasets and higher computational resources compared to CNNs, they offer scalability and adaptability for complex classification problems. Incorporating these advanced architectures alongside CNNs could potentially enhance model performance, especially for visually similar classes like Espresso and Starbucks Pike Place, where traditional CNNs may encounter challenges. Additionally, comparing the efficiency and accuracy trade-offs between CNNs and transformer-based models under real-world conditions could provide deeper insights into their practical applications. These advancements would further solidify the role of artificial intelligence in modernizing agricultural processes, ensuring the scalability and robustness of automatic classification systems for diverse coffee bean types and global production scenarios.

Conclusions

The results of this study underscore the significant potential of deep learning models for the automatic classification of coffee bean species. By evaluating five popular CNN architectures—Xception, DenseNet201, InceptionV3, InceptionResNetV2, and DenseNet121—this research identified InceptionV3 as the top-performing model, achieving an average accuracy of 93%. The model consistently outperformed others across all major metrics, including precision, recall, and F-1 score, while demonstrating superior discriminative capabilities in ROC curve analysis.

A pivotal contribution of this study is the statistical validation of model performance through McNemar’s test, as detailed in Table 7. Although InceptionV3 exhibited the highest numerical performance across key metrics, the test results revealed no statistically significant differences between its performance and those of DenseNet121 (92%) and Xception (91%). This finding highlights the robustness of all tested CNN architectures for classifying coffee bean species. Furthermore, the absence of significant differences emphasizes the importance of practical considerations, such as computational efficiency and ease of deployment, in selecting models for real-world applications.

An important consideration in this study is the computational efficiency of the tested models, as summarized in Table 6. Among the five CNN architectures, InceptionV3 demonstrated the shortest training time (70.4 min) while achieving the highest classification performance (93% accuracy), highlighting its suitability for real-world applications requiring both efficiency and accuracy. In contrast, models like DenseNet201 (317.6 min) and InceptionResNetV2 (225.9 min) required significantly longer training times, reflecting their higher computational complexity. These findings underscore the importance of balancing computational cost with performance metrics when selecting models for practical use. For applications with limited computational resources, InceptionV3 offers a highly efficient solution, whereas more complex models may be preferred in scenarios where resource constraints are not a concern and additional performance gains are sought. This analysis further emphasizes the need to consider both accuracy and computational efficiency when optimizing classification systems for scalable and practical deployment in the coffee industry.

An additional insight is the challenge posed by visually similar coffee bean types, such as Espresso and Starbucks Pike Place, which resulted in higher misclassification rates. This underscores the need for further refinement in model architectures and the incorporation of additional features, such as chemical composition data, to improve classification accuracy. Addressing these challenges would enable more precise differentiation between visually similar types, ultimately enhancing the practical applicability of the models.

The broader implications of this research lie in its potential to transform quality control processes within the coffee industry. The study demonstrates that CNN models, particularly InceptionV3, can provide a non-destructive, cost-effective, and accurate method for real-time coffee bean classification. This advancement could minimize human errors, improve operational efficiency, and ensure the delivery of more reliable and consistent products to consumers and producers. Moreover, accurate classification can play a critical role in combating food fraud by ensuring authenticity and maintaining product integrity throughout the supply chain.

To further enhance the predictive power and applicability of these models, future work should focus on expanding the dataset with a larger variety of images, representing diverse coffee types and geographical regions. This would improve model generalizability and classification accuracy. Additionally, the integration of advanced architectures, such as transformers or hybrid models, could offer enhanced performance, particularly in differentiating visually similar coffee types. By addressing these areas, future research could further optimize classification systems and contribute to more scalable, robust, and innovative solutions for quality control in the coffee industry.

In conclusion, this study highlights the strong potential of deep learning models for coffee bean classification and lays a foundation for future research aimed at optimizing classification methodologies and advancing agricultural quality control practices.

Supplemental Information