Introduction

Horticultural crops, particularly fruits, play a crucial role in global food security, human nutrition, and agricultural economies. Fruits are rich sources of essential vitamins, minerals, antioxidants, and dietary fiber, making them indispensable components of a healthy diet. The quality and market value of fruits are highly dependent on their stage of ripeness, which influences key attributes such as color, texture, flavor, aroma, and nutritional composition [1]. Accurate determination of fruit ripeness is therefore essential for optimizing harvesting time, minimizing post-harvest losses, and ensuring consumer satisfaction, fruit ripeness has been assessed using manual and destructive methods, including visual inspection, firmness testing, and chemical analysis of parameters such as sugar content, acidity, and ethylene production. While these techniques provide valuable information, they are often subjective, labor-intensive, time-consuming, and unsuitable for large-scale commercial operations [2]. Moreover, manual grading is prone to human error and inconsistency, particularly when dealing with large volumes of produce in supply chains. With the advancement of modern technologies, there has been a growing interest in non-destructive and automated approaches for fruit quality assessment. Image processing techniques initially provided a foundation for analyzing visual characteristics of fruits; however, traditional machine learning methods required manual feature extraction and were limited in their ability to capture complex patterns. The emergence of deep learning, particularly Convolutional Neural Networks (CNNs), has significantly transformed this field by enabling automatic feature extraction and high-accuracy classification.

CNNs have demonstrated remarkable success in various image recognition tasks, including object detection, medical imaging, and agricultural applications. In the context of horticulture, CNN-based systems can analyze fruit images to identify subtle variations in color, texture, and surface patterns associated with different ripeness stages [3]. These systems offer several advantages, including high accuracy, scalability, real-time processing, and the ability to operate under non-destructive conditions. The integration of CNN-based ripeness classification into agricultural practices supports precision farming, automated grading systems, and smart supply chain management. It enables growers, processors, and retailers to make data-driven decisions regarding harvesting, storage, transportation, and marketing [4]. Furthermore, the use of CNNs contributes to reducing food waste, improving resource efficiency, and enhancing the overall quality of horticultural produce. This review aims to provide a comprehensive overview of the application of Convolutional Neural Networks in fruit ripeness classification. It explores the fundamental principles of CNNs, methodologies for image acquisition and preprocessing, various model architectures, and their applications in horticultural systems. Additionally, the review highlights key challenges, recent advancements, and future research directions in this rapidly evolving field.

2. Fundamentals of Convolutional Neural Networks

Convolutional Neural Networks (CNNs) are a class of deep learning models specifically designed for processing structured grid data, such as images. Inspired by the visual cortex of the human brain, CNNs are capable of automatically learning hierarchical feature representations from raw input data. This capability makes them highly effective for image classification tasks, including fruit ripeness detection. A typical CNN architecture consists of multiple interconnected layers, each performing a specific function in the feature extraction and classification process [5]. The primary components of a CNN include convolutional layers, activation functions, pooling layers, and fully connected layers.

Convolutional layers are the core building blocks of CNNs. These layers apply a set of learnable filters (kernels) that slide over the input image to extract local features such as edges, corners, and textures [6]. Each filter produces a feature map that highlights specific patterns within the image. As the network depth increases, the extracted features become more abstract and complex, enabling the model to distinguish subtle differences between ripeness stages. Activation functions, such as the Rectified Linear Unit (ReLU), introduce non-linearity into the model, allowing it to learn complex relationships between input features. Without activation functions, the network would behave as a linear model and would be unable to capture intricate patterns in the data. Pooling layers are used to reduce the spatial dimensions of feature maps, thereby decreasing computational complexity and preventing overfitting [7]. Common pooling techniques include max pooling and average pooling. By summarizing local regions of feature maps, pooling layers help retain important features while discarding irrelevant information. Fully connected layers are typically located at the end of the network and are responsible for performing the final classification. These layers take the high-level features extracted by previous layers and map them to output classes, such as unripe, semi-ripe, and ripe. In addition to these core components, modern CNN architectures often incorporate advanced techniques such as batch normalization, dropout, and residual connections to improve training efficiency and model performance. Several well-known CNN architectures have been widely applied in fruit ripeness classification:

• AlexNet: One of the earliest deep CNN models, known for its breakthrough performance in image classification tasks.
• VGGNet: Characterized by its simple and uniform architecture using small convolutional filters.
• ResNet (Residual Network): Introduces skip connections to enable the training of very deep networks without degradation problems.
• Inception (GoogLeNet): Utilizes parallel convolutional layers with different filter sizes to capture multi-scale features.
• MobileNet: A lightweight architecture designed for mobile and embedded devices, suitable for real-time applications.

CNNs can be trained using two primary approaches: training from scratch or transfer learning. Training from scratch requires large labelled datasets and high computational resources, while transfer learning leverages pre-trained models to achieve high performance with limited data. In horticultural applications, transfer learning is commonly used due to the scarcity of large, annotated fruit image datasets [8]. CNNs provide a robust and efficient framework for automated fruit ripeness classification by combining feature extraction and classification into a unified model. Their adaptability and high performance make them a cornerstone of modern AI-driven agricultural systems.

3. CNN-Based Approaches for Fruit Ripeness Classification

Convolutional Neural Networks (CNNs) have become a cornerstone in automated fruit ripeness classification due to their ability to learn complex visual patterns directly from image data. Unlike traditional machine learning approaches that depend on manually engineered features, CNNs automatically extract hierarchical features such as color gradients, texture variations, and surface irregularities, which are critical indicators of fruit ripeness. During the training process, CNN models learn to distinguish between different ripeness stages—such as unripe, semi-ripe, and ripe—by identifying subtle visual cues that may not be easily detectable by the human eye. This capability enables highly accurate and consistent classification, making CNNs particularly suitable for large-scale horticultural applications [9]. CNN-based ripeness classification systems typically rely on two main approaches: custom-built architectures and transfer learning. Custom CNN models are designed and trained from scratch using specific fruit datasets, allowing for flexibility and optimization according to particular use cases. However, these models require large annotated datasets and significant computational resources. In contrast, transfer learning has gained widespread popularity due to its efficiency and effectiveness. Pre-trained models such as VGGNet, ResNet, and MobileNet are fine-tuned using relatively smaller datasets, enabling faster training and improved performance even with limited data availability. This approach is especially valuable in horticulture, where collecting large, labeled datasets can be challenging.

The performance of CNN models in ripeness classification is influenced by several factors, including image quality, dataset diversity, and preprocessing techniques. Controlled imaging conditions, such as consistent lighting and uniform backgrounds, often result in higher accuracy. However, real-world applications require models to perform reliably under variable conditions, including changes in illumination, occlusion, and background complexity. To address this, data augmentation techniques such as rotation, flipping, scaling, and color adjustments are commonly applied to improve model robustness and generalization capability. Evaluation of CNN-based ripeness classification models is typically conducted using standard performance metrics such as accuracy, precision, recall, F1-score, and confusion matrices. High-performing models frequently achieve accuracy levels exceeding 90% under controlled conditions [10]. Despite these promising results, challenges remain in distinguishing closely related ripeness stages and ensuring consistent performance across different fruit varieties and environmental conditions. Nevertheless, CNN-based approaches continue to evolve rapidly, offering significant potential for improving automation and decision-making in horticultural systems.

4. Applications of CNNs in Horticulture

The application of Convolutional Neural Networks in horticulture has significantly improved the efficiency and accuracy of fruit ripeness assessment across various stages of the agricultural value chain. One of the most critical applications is in determining optimal harvesting time. By accurately classifying fruit ripeness in real time, CNN-based systems enable growers to harvest crops at the ideal stage, ensuring maximum quality, shelf life, and market value. This is particularly important for climacteric fruits, where ripening continues after harvest and timing is crucial for maintaining quality during transportation and storage. In post-harvest management, CNN-based systems are widely used for automated sorting and grading of fruits. These systems analyze visual characteristics such as color, size, shape, and surface defects to classify fruits into different quality grades. Automated grading not only ensures consistency and reduces human error but also increases processing speed and efficiency in packing facilities [11]. An integrating CNNs with conveyor belt systems and robotic sorting mechanisms, large volumes of fruits can be processed rapidly with minimal manual intervention.

CNNs also play a vital role in supply chain optimization by enabling accurate prediction of fruit ripeness and shelf life. This information helps stakeholders make informed decisions regarding storage conditions, transportation schedules, and market distribution. For instance, fruits classified at earlier ripeness stages can be directed toward distant markets, while fully ripe fruits can be supplied to local markets for immediate consumption. Such intelligent distribution strategies help reduce post-harvest losses and food waste, which are major concerns in the horticultural sector, the integration of CNNs with emerging technologies such as the Internet of Things (IoT), drones, and robotic systems has led to the development of smart agriculture solutions. Drones equipped with imaging sensors can capture large-scale orchard data, which is then analyzed using CNN models to monitor crop health and ripeness. Similarly, robotic harvesters equipped with vision systems can identify and pick fruits at the appropriate ripeness stage. These innovations contribute to precision agriculture by enabling real-time monitoring, automated decision-making, and efficient resource utilization, ultimately enhancing productivity and sustainability.

5. Challenges and Limitations

Despite the significant advancements in CNN-based fruit ripeness classification, several challenges and limitations hinder their widespread adoption in real-world horticultural systems. One of the primary challenges is the variability in environmental conditions under which images are captured. Factors such as lighting intensity, shadows, background complexity, and camera angles can significantly affect image quality and, consequently, model performance. CNN models trained under controlled laboratory conditions often struggle to maintain accuracy when deployed in open-field environments, where conditions are highly dynamic and unpredictable. Another major limitation is the availability of large, high-quality, and annotated datasets required for training robust CNN models. In many cases, datasets are limited in size and lack diversity in terms of fruit varieties, ripeness stages, and environmental conditions. This can lead to overfitting, where the model performs well on training data but poorly on unseen data. Data collection and annotation are also time-consuming and resource-intensive processes, particularly when expert knowledge is required to accurately label ripeness stages. Computational complexity and resource requirements present additional challenges, especially for real-time applications and deployment on edge devices such as smartphones and embedded systems. Deep CNN models often require high processing power and memory, which may not be readily available in field conditions. Although lightweight architectures such as MobileNet have been developed to address this issue, achieving a balance between model efficiency and accuracy remains a key research challenge.

Generalization across different fruit types and varieties is another critical concern. Models trained on a specific fruit dataset may not perform well when applied to other fruits due to differences in color, texture, and ripening patterns. This limits the scalability of CNN-based systems and necessitates the development of more generalized or adaptable models. Additionally, distinguishing between closely related ripeness stages remains difficult, as visual differences can be subtle and influenced by external factors. Finally, practical deployment of CNN-based systems requires integration with hardware, user interfaces, and existing agricultural workflows. Issues related to cost, technical expertise, and user acceptance can further restrict adoption, particularly among small-scale farmers. Addressing these challenges will require advancements in data collection, model design, and system integration, as well as interdisciplinary collaboration between researchers, engineers, and agricultural practitioners.

5. Applications in Horticulture

The application of Convolutional Neural Networks (CNNs) in horticulture has significantly enhanced the efficiency, accuracy, and automation of fruit ripeness assessment across different stages of the agricultural value chain. One of the most important applications is in determining optimal harvesting time. CNN-based systems analyze visual features such as color changes, surface texture, and morphological characteristics to accurately classify the ripeness stage of fruits. This enables growers to make precise harvesting decisions, thereby reducing the risks associated with premature or delayed harvesting. Harvesting at the correct stage not only improves fruit quality and shelf life but also maximizes market value and consumer satisfaction. In post-harvest operations, CNN-based technologies are widely utilized for automated sorting and grading of fruits. These systems are capable of evaluating multiple quality parameters, including ripeness level, size, shape, color uniformity, and surface defects [12]. An integrating CNN models with conveyor-based sorting systems and robotic mechanisms, large volumes of fruits can be processed rapidly and consistently. This reduces human error, enhances operational efficiency, and ensures uniform quality standards in packaging and distribution. As a result, producers and suppliers can meet stringent market requirements and improve overall supply chain performance.

CNN-based ripeness classification also plays a crucial role in optimizing supply chain management. Accurate prediction of fruit ripeness and shelf life enables stakeholders to make informed decisions regarding storage conditions, transportation logistics, and market distribution strategies. For example, fruits identified at earlier ripeness stages can be transported to distant markets, while fully ripe fruits can be directed toward local markets for immediate consumption. Such data-driven decision-making helps minimize post-harvest losses, reduce food waste, and improve profitability across the supply chain, the integration of CNNs with smart agriculture systems has opened new avenues for precision horticulture [13]. When combined with Internet of Things (IoT) devices, sensors, drones, and robotic platforms, CNN-based models enable real-time monitoring of crop conditions and automated decision-making. Drones equipped with imaging systems can capture orchard-level data, which is analyzed using CNNs to assess fruit ripeness and overall crop health. Similarly, robotic harvesters can use CNN-based vision systems to identify and selectively harvest fruits at the optimal stage of maturity [14-16]. These advancements contribute to increased efficiency, reduced labor dependency, and improved resource utilization, ultimately supporting sustainable and technologically advanced horticultural practices.

5. Challenges and Limitations

The remarkable progress in CNN-based fruit ripeness classification, several challenges continue to limit their widespread implementation in real-world horticultural systems. One of the most significant challenges is the variability in environmental conditions during image acquisition. Factors such as inconsistent lighting, shadows, background clutter, and varying camera angles can significantly affect image quality and reduce model accuracy. CNN models trained under controlled laboratory conditions often struggle to generalize effectively in open-field environments, where conditions are dynamic and unpredictable. This highlights the need for robust models capable of handling real-world variability. Another major limitation is the lack of large, diverse, and well-annotated datasets. High-quality datasets are essential for training deep learning models; however, collecting and labeling such data is time-consuming, labor-intensive, and often requires expert knowledge to accurately define ripeness stages. Limited datasets can lead to overfitting, where models perform well on training data but fail to generalize to new or unseen data. [15], variability in fruit varieties, geographical conditions, and cultivation practices further complicates dataset standardization. Computational complexity and resource requirements also pose significant challenges. Deep CNN architectures require substantial processing power, memory, and energy consumption, which may not be feasible for deployment in field conditions or on edge devices such as smartphones and embedded systems. Although lightweight models such as MobileNet and EfficientNet have been developed to address these constraints, achieving an optimal balance between computational efficiency and classification accuracy remains a critical research issue. Generalization across different fruit types and ripeness stages is another limitation. CNN models trained on specific fruit datasets may not perform well when applied to other fruits due to differences in color, texture, and ripening patterns, distinguishing between closely related ripeness stages can be difficult, as visual differences are often subtle and influenced by external factors such as environmental stress or post-harvest handling. Practical deployment also requires integration with hardware systems, user-friendly interfaces, and existing agricultural workflows, which can present technical and economic barriers, particularly for small-scale farmers.

6. Future Perspectives and Research Directions

The future of CNN-based fruit ripeness classification lies in the development of more robust, efficient, and scalable systems that can operate effectively under real-world conditions. One promising direction is the integration of CNNs with advanced imaging technologies such as hyperspectral and multispectral imaging. These technologies capture information beyond the visible spectrum, enabling the detection of internal and biochemical changes associated with ripening that are not visible to the human eye. Combining spectral data with CNN models can significantly enhance classification accuracy and reliability. Another important research direction is the development of lightweight and energy-efficient models for deployment on edge devices. With the increasing availability of smartphones and portable devices, there is growing potential for on-field ripeness assessment using mobile applications. Edge computing can enable real-time processing without relying on cloud infrastructure, reducing latency and improving accessibility for farmers in remote areas. The integration of CNNs with other emerging technologies such as the Internet of Things (IoT), robotics, and autonomous systems is also expected to drive innovation in horticulture. IoT-enabled sensors can provide continuous environmental and crop data, which, when combined with CNN-based image analysis, can support predictive and adaptive decision-making. Robotic systems equipped with vision-based CNN models can automate harvesting and sorting processes, improving efficiency and reducing labor dependency, the development of large, standardized, and publicly available datasets will be essential for advancing research in this field. Collaborative efforts among research institutions, industry stakeholders, and policymakers can facilitate data sharing and benchmarking, enabling the development of more generalized and transferable models. The incorporation of explainable AI techniques will also be important to improve model transparency and user trust, particularly in critical decision-making processes.

7. Conclusion

Convolutional Neural Networks have emerged as powerful tools for automated fruit ripeness classification, offering significant advantages over traditional methods in terms of accuracy, efficiency, and scalability. By enabling non-destructive, real-time analysis of visual features, CNN-based systems have the potential to transform various aspects of horticultural production, including harvesting, post-harvest management, and supply chain optimization. Their ability to learn complex patterns from image data allows for consistent and objective classification, reducing reliance on manual labor and minimizing human error. Despite these advancements, several challenges remain, including variability in environmental conditions, limited dataset availability, computational constraints, and difficulties in generalizing across different fruit types. Addressing these challenges will require continued research and innovation in model development, data collection, and system integration. Interdisciplinary collaboration among computer scientists, agricultural experts, and engineers will be essential to bridge the gap between research and practical implementation, the integration of CNN-based ripeness classification systems into horticultural practices represents a significant step toward precision agriculture and smart farming. As technologies continue to evolve, these systems are expected to play a central role in enhancing productivity, reducing post-harvest losses, and ensuring high-quality produce for global markets.

References

1. Rendón-Vargas, A., Luna-Álvarez, A., Mújica-Vargas, D., Castro-Bello, M., & Marianito-Cuahuitic, I. (2024, November). Application of Convolutional Neural Networks for the Classification and Evaluation of Fruit Ripeness. In International Congress of Telematics and Computing (pp. 150-163). Cham: Springer Nature Switzerland.
2. Rodriguez, M., Pastor, F., & Ugarte, W. (2021, January). Classification of fruit ripeness grades using a convolutional neural network and data augmentation. In 2021 28th Conference of Open Innovations Association (FRUCT) (pp. 374-380). IEEE.
3. Naranjo-Torres, J., Mora, M., Hernández-García, R., Barrientos, R. J., Fredes, C., & Valenzuela, A. (2020). A review of convolutional neural network applied to fruit image processing. Applied Sciences, 10(10), 3443.
4. Olisah, Chollette C., Ben Trewhella, Bo Li, Melvyn L. Smith, Benjamin Winstone, E. Charles Whitfield, Felicidad Fernández Fernández, and Harriet Duncalfe. “Convolutional neural network ensemble learning for hyperspectral imaging-based blackberry fruit ripeness detection in uncontrolled farm environment.” Engineering applications of artificial intelligence 132 (2024): 107945.
5. Mazen, F. M., & Nashat, A. A. (2019). Ripeness classification of bananas using an artificial neural network. Arabian Journal for Science and Engineering, 44(8), 6901-6910.
6. Tapia-Mendez, E., Cruz-Albarran, I. A., Tovar-Arriaga, S., & Morales-Hernandez, L. A. (2023). Deep learning-based method for classification and ripeness assessment of fruits and vegetables. Applied Sciences, 13(22), 12504.
7. Kangune, K., Kulkarni, V., & Kosamkar, P. (2019, October). Grapes ripeness estimation using convolutional neural network and support vector machine. In 2019 Global Conference for Advancement in Technology (GCAT) (pp. 1-5). IEEE.
8. Almomen, M., Al-Saeed, M., & Ahmad, H. F. (2023). Date fruit classification based on surface quality using convolutional neural network models. Applied Sciences, 13(13), 7821.
9. Wang, C., Liu, S., Wang, Y., Xiong, J., Zhang, Z., Zhao, B., … & He, P. (2022). Application of convolutional neural network-based detection methods in fresh fruit production: a comprehensive review. Frontiers in plant science, 13, 868745.
10. Miragaia, R., Chávez, F., Díaz, J., Vivas, A., Prieto, M. H., & Moñino, M. J. (2021). Plum ripeness analysis in real environments using deep learning with convolutional neural networks. Agronomy, 11(11), 2353.
11. Kutyrev, A., Kiktev, N. A., Kalivoshko, O., & Rakhmedov, R. S. (2022, November). Recognition and Classification Apple Fruits Based on a Convolutional Neural Network Model. In IT&I (pp. 90-101).
12. Ibrahim, Z., Sabri, N., & Isa, D. (2018). Palm oil fresh fruit bunch ripeness grading recognition using convolutional neural network. Journal of Telecommunication, Electronic and Computer Engineering (JTEC), 10(3-2), 109-113.
13. Gururaj, N., Vinod, V., & Vijayakumar, K. (2023). Deep grading of mangoes using convolutional neural network and computer vision. Multimedia Tools and Applications, 82(25), 39525-39550.
14. Mommoh, J. S., Obetta, J. L., John, S. N., Okokpujie, K., Omoruyi, O. N., & Awelewa, A. A. (2024). Detection of fruit ripeness and defectiveness using convolutional neural networks. Inf. Dyn. Appl, 3(3), 184-199.
15. Benmouna, B., García-Mateos, G., Sabzi, S., Fernandez-Beltran, R., Parras-Burgos, D., & Molina-Martínez, J. M. (2022). Convolutional neural networks for estimating the ripening state of fuji apples using visible and near-infrared spectroscopy. Food and Bioprocess Technology, 15(10), 2226-2236.
16. Steinbrener, J., Posch, K., & Leitner, R. (2019). Hyperspectral fruit and vegetable classification using convolutional neural networks. Computers and Electronics in Agriculture, 162, 364-372.