In recent years, Artificial Intelligence (AI) has gained significant traction in agriculture, addressing challenges like crop disease identification, yield forecasting, and precision farming [1]. By leveraging technologies like deep learning, machine learning, computer vision, and automation is transforming traditional agricultural practices into data-driven systems. This shift enables real-time decision-making, enhances sustainability, optimizes the use of resources, and boosts productivity, while also addresses critical issues such as labor shortages, environmental impacts, and connectivity in modern agricultural operations [2]. One of the major challenges in modern agriculture is crop health monitoring. Early detection of plant diseases, nutrient deficiencies, and pest infestations is crucial to provide intervention solutions to ensure high yields [3]. Another important problem is automated fruit and vegetable grading, where AI-powered systems assess the quality of produce based on color, size, and texture [4]. Additionally, weed detection is a significant application where computer vision helps differentiate crops from weeds, reducing herbicide usage [5].

Image classification plays a fundamental role in solving many agricultural problems, including above issues. Modern AI models can analyze images of crops, leaves, fruits, and soil to classify plant conditions, detect diseases, and even assess plant growth stages [6]. This approach not only enhances efficiency but also enables real-time decision-making for farmers, contributing to the modernization of agriculture. Image classification tasks in agriculture are commonly implemented following two major approaches: (1) Image-level classification refers to the task of assigning a single label to an entire image. It is a fundamental problem in computer vision, where the objective is to categorize the whole image into one of predefined classes. (2) Pixel-level or region-based classification—assigning labels to groups of pixels or specific regions within an image and it is also known as semantic segmentation, involves classifying each pixel (or group of pixels) in an image into a specific class.

By leveraging both approaches, AI-based systems can offer scalable and precise solutions for wide range of agricultural applications, from disease detection and crop monitoring to quality assessment and resource optimization.

Traditional approaches of image classification typically adopt a two-step pipeline: First, handcrafted feature is extracted from an image, then this feature is passed into classifiers such as Support Vector Machines (SVM) or K-Nearest Neighbors (KNN) for final prediction. Among popular handcrafted feature descriptors, SIFT [7] is well-known due to its robustness to scale and rotation, while HOG [8] excels at capturing gradient-based texture patterns, which makes it effective for object detection problems. Despite their success in certain tasks, these methods heavily rely on manual feature engineering, which requires domain expertise and struggles to generalize well to large, diverse datasets. Second, with the advent of deep learning, particularly Convolutional Neural Networks (CNNs), the paradigm shifted from manual feature design to automatic hierarchical feature learning directly from raw image data. Early breakthroughs such as AlexNet [9], VGGNet [10], InceptionNet [11], and ResNet [12] have demonstrated the superior performance of deep CNNs over traditional methods in large-scale image classification tasks like ImageNet [13]. These models collectively shaped the evolution of CNN architectures, from increasing network depth and using smaller convolutional kernels to incorporating multi-scale feature extraction and residual learning. Kumar et al. [14] proposed a hybrid model combining Convolutional Auto-Encoder (CAE) and DenseNet-based CNN for automatic tomato leaf disease diagnosis without preprocessing, achieving 98.35% accuracy on the PlantVillage dataset with reduced training time due to fewer trainable components. Although CNNs achieve outstanding accuracy in many classification tasks, they also have limitations. CNNs inherently operate on regular grids of pixels, which may not fully capture irregular or non-Euclidean spatial structures present in some types of images, such as medical images or biological microscopy data. Additionally, deep CNNs demand large annotated datasets and considerable computational resources for training, which can be challenging in data-scarce or resource-constrained applications.

Overall, traditional handcrafted methods are simple and easy to interpret, but they lack flexibility and scalability. In contrast, CNN-based methods achieve much higher accuracy, especially on complex and large-scale image classification tasks, but they require more data and computational resources. Therefore, choosing the right method depends on the specific context and requirement of each classification problem. This limitation has motivated researchers to explore Graph Neural Networks (GNNs) as an alternative approach for image classification. GNNs are a family of deep learning models specifically designed to process graph-structured data, where data points are represented as nodes connected by edges [15]. Unlike traditional neural networks, which work on regular grids (like images) or sequences (like text), GNNs excel at modeling relationships in non-Euclidean spaces. This flexibility makes GNNs applicable to a wide range of tasks, including social network analysis, molecular property prediction, and image classification [16].

One of the main strengths of GNNs lies in their iterative aggregation of information from neighbor nodes, which allows the model to effectively captures both local structures and global dependencies within the graph. This capability is particularly beneficial in image classification, where understanding relationships between regions can enhance recognition accuracy. Among the various GNN architectures, Graph Convolutional Networks (GCNs) have gained widespread adoption due to their simplicity and effectiveness. GCNs extend the concept of convolution from CNNs to graph data by aggregating features from a node’s neighbors through a weighted sum [17]. In image classification tasks, this can be applied by first converting an image into a graph, where each node represents a meaningful region (e.g., pixel, superpixel) and edges encode spatial or semantic relationships between regions. This graph-based representation allows GCNs to integrate both local features and global context, offering a more holistic view of the image and improving classification performance. These two GCN-based approaches align closely with the two image classification paradigms discussed earlier: the first approach—graph-level classification—corresponds to assigning a single label to each image, while the second—node-level classification—corresponds to pixel-level or region-based classification, where each part of the image is individually labeled.

Although more recent architectures such as GraphSAGE [18] and GIN [19] have been proposed to address certain limitations of GCNs, GCN remains a preferred choice in many image classification applications due to its computational efficiency and stable performance on moderately sized graphs. GraphSAGE includes sampling strategies and more complex aggregation functions that are especially useful in large-scale or dynamic graphs, but may add unnecessary complexity when dealing with relatively static and structured graphs derived from images. Meanwhile, GIN achieves strong discriminative power by using multilayer perceptrons (MLPs) for aggregation, but often requires more extensive hyperparameter tuning and training resources. In contrast, GCNs provide a favorable trade-off between model expressiveness and computational simplicity, which makes them well-suited for tasks involving structured data such as images represented as graphs, especially when interpretability and efficiency are important [15,16].

Several studies applied GCNs to graph-level classification tasks, where each image was treated as a single graph to be labeled. Rodrigues et al. [20] conducted experiments on various image-to-graph transformation strategies including segmentation granularity, superpixel feature selection, and edge construction methods. Their experiments showed that these design choices substantially influence GCN performance. The best configuration yielded over 91.3 ± 0.4% accuracy across various datasets, confirming the feasibility and effectiveness of GCN-based models for image classification through graph representations. Han et al. [21] proposed Vision GNN (ViG), a novel graph-based architecture that represents images as graphs and achieved comparable accuracy to CNN-based ResNet models on CIFAR-10, while offering better interpretability due to its explicit graph structure. On the ImageNet dataset, ViG achieved 83.7% top-1 accuracy, and it also improved performance in object detection tasks on the COCO dataset, demonstrating strong representational capacity for both classification and detection tasks. Moreover, this study also demonstrated that the proposed method outperforms several other graph models, such as GraphSAGE and GIN.

In contrast, node-level classification is particularly effective for large-scale images (e.g., UAV (Unmanned Aerial Vehicle) imagery or hyperspectral images), which are divided into superpixels or patches. Each superpixel is modeled as a node, with edges capturing spatial or spectral relationships between them. GCNs are then used to label each node individually, enabling high-resolution, region-level interpretation. To support this, the miniGCN model proposed in [22] introduced an efficient mini-batch training scheme tailored for large-scale hyperspectral datasets. It also supports out-of-sample inference without retraining and integrates features from both CNNs and GCNs via various fusion strategies. Experiments on Indian Pines and Pavia University datasets showed significant improvements, with overall accuracy (OA) reaching 75.11% and 79.79%, respectively.

Another study [23] also leveraged multi-view features from hyperspectral images and applied GCNs to enhance classification performance by introducing an adaptive multi-feature fusion framework that integrates spectral and textural information through a multi-branch architecture with attention-based fusion. On the Salinas dataset, this method achieved an OA of 98.03 ± 1.02% and showed greater robustness to spectral variability, outperforming baseline CNNs by 4%–6% in overall accuracy. Liu et al. [24] introduced a CNN-enhanced GCN (CEGCN) that jointly learns from pixel- and superpixel-level features, integrating adaptive graph learning into the model. The proposed CEGCN was evaluated on three widely used hyperspectral image (HSI) datasets: Indian Pines, University of Pavia, and Salinas. Experimental results demonstrate that CEGCN consistently outperforms state-of-the-art deep learning methods (e.g., DCNN, HybridSN, and DBDA) in terms of OA, average accuracy (AA), and kappa coefficient (KPP). It effectively smooths predictions in homogeneous regions while preserving fine details in small or complex areas, and shows greater robustness under limited training data conditions. These studies collectively demonstrate the case-dependent performance of different GCN variants in image classification. While hybrid or enhanced models often outperform standard GCNs, the degree of improvement varies significantly across datasets—highlighting the importance of aligning model architecture with the structure and scale of the image data. Furthermore, GCNs remain a competitive baseline in many graph-based vision tasks due to their simplicity and reliability, particularly when interpretability and computational efficiency are key considerations.

In various studies on applying Graph Convolutional Networks (GCNs) for image classification problems, the first essential step is converting images from regular pixel grids into graph structures. One common approach represents each individual pixel as a node, with edges connecting neighboring pixel [25]. While this method preserves fine-grained spatial information, it results in extremely large and computationally expensive graphs, especially for high-resolution images. Another approach divides the image into uniform square patches, where each patch is treated as a node and edges link neighboring patches [21]. This reduces the graph size significantly, but it comes at the cost of losing important boundary details between objects, especially when object edges do not align with the patch boundaries.

To overcome these limitations, superpixel-based graph construction emerged as a more effective alternative. In this approach, the image is segmented into coherent regions—superpixels—that group together pixels with similar color, texture, or other low-level properties. Each superpixel becomes a node, and edges capture spatial relationships between neighboring superpixels [20,26]. This technique strikes a balance between reducing graph size and preserving important object boundaries and structural details, making it particularly suitable for image classification tasks where both local texture and global structure matter.

A key advantage of the superpixel-based graph construction method is the ability to balance graph complexity and structural preservation, which directly influences classification performance. Building on this foundation, several studies explored different strategies to optimize this approach, evaluated how various factors in graph construction impact the effectiveness of GCN models. One notable study, Rodrigue et al. [20] investigated different strategies for converting images into graphs, particularly by using superpixel segmentation. This research evaluates how various graph construction techniques impact classification performance, considering multiple aspects: the method of transforming an image into a graph, the number of nodes in the graph, the representation of features for each node, and the definition of edges between nodes. These factors play a crucial role in determining how well the graph structure preserves spatial and semantic relationships within the image, ultimately affecting the effectiveness of the GCN model. Similarly, Tang et al. [27] integrated the gSLIC superpixel algorithm with attention mechanisms, aiming to enhance GCNs’ ability to capture structural information within images. This study demonstrates that the proposed approach reduces graph construction time while also improving classification accuracy, making it a more efficient and effective method for image representation in GCN-based classification tasks.

Even though representing image as graph using superpixel-based segmentation proves to be a more suitable approach compared to the other two methods, thanks to it effectively balances computational efficiency and structural preservation. However, the next crucial challenge lies in determining the most effective way to define and represent node features for optimal classification performance. Rodrigue et al. [20] conducted experiments on seven different feature extraction methods and compared their effectiveness in GCN-based classification. These methods include spatial features such as the geometric centroid and pixel position distribution, structural features like the number of pixels per superpixel, and color-based features, including average RGB and HSV values along with their standard deviations. Their study emphasizes the significance of selecting appropriate node descriptors that not only capture local texture details but also preserve broader semantic information, ensuring that the graph representation retains essential characteristics for accurate image classification.

Another critical aspect of graph construction is constructing the connectivity between nodes, i.e., how edges are formed. In study [20], Rodrigue et al. explored three different strategies for establishing edges in superpixel-based graphs: Region Adjacency Graphs (RAGs), which connect nodes based on direct superpixel neighborhood relationships; K-Nearest Neighbors with spatial distance (KNN-Spatial), where nodes are linked to their closest neighbors based on spatial proximity; and K-Nearest Neighbors with combined spatial and color distances (KNN-Combined), which considers both spatial closeness and color similarity to form edges. Each of these approaches impacts how information propagates through the graph and influences the final performance of GCN models in classification tasks. Interestingly, their findings suggest that while more complex edge relationships, such as those incorporating both spatial and color distances, might intuitively capture richer structural information, they do not always lead to better classification performance. In some cases, simpler connectivity strategies, like RAGs or purely spatial KNN, are sufficient or even superior, highlighting the importance of carefully selecting edge construction methods based on the specific dataset and task requirements.

Building upon this foundation, this part of literature review focuses on how different published studies have implemented and applied graph-based models such as Graph Convolutional Networks (GCNs) to image-related tasks. We aim to examine the effectiveness of these approaches, analyzing the specific adaptations made to GCN architectures, the datasets used, and the overall impact on classification performance. By reviewing these implementations, we seek to gain deeper insights into the practical applications of GCNs in image classification and identify key factors that contribute to their success. In the study [27], Tang conducted a comprehensive analysis of graph-based classification methods using the GCN model. This study discussed key challenges, such as graph construction, over-smoothing, and scalability, while also explored strategies to overcome these issues. These studies highlight the importance of effective image-to-graph transformation and advanced graph processing techniques in leveraging GCNs for image classification.

Knyazev et al. [28] represented images as collections of superpixels and formulated the image classification task as a multigraph classification problem. This study extended graph convolution techniques and adopted relation type fusion methods to enhance the expressiveness of GCNs. To improve classification accuracy, they proposed a multigraph representation incorporating learnable and hierarchical relation types for a richer image representation. Through experiments were conducted on the MNIST, CIFAR-10, and PASCAL datasets, and their proposed model achieved notable improvements in accuracy, surpassing CNNs in specific classification tasks. Another significant work, Wharton et al. [29] proposed an innovative multi-scale hierarchical representation learning approach to enhance visual recognition. Additionally, they introduced a coarser hierarchical representation of multiple graphs using spectral clustering-based region aggregation. A novel gated attention mechanism was designed to aggregate cluster-level class-specific confidence. Their model was evaluated on five datasets covering fine-grained visual classification (FGVC) and generic visual classification, achieving competitive results.

As mentioned above, image classification is a core task in computer vision, serving as the foundation for a wide range of real-world applications. While Convolutional Neural Networks (CNNs) have achieved remarkable success in this domain, they have notable limitations. CNNs primarily operate on raw pixel grids, which restricts their ability to capture structural relationships between different image regions. Moreover, their reliance on dense pixel-level processing leads to high computational costs, making them inefficient for large-scale classification tasks.

To address these challenges, researchers have explored Graph Convolutional Networks (GCNs) as an alternative approach, leveraging graph structures to model spatial relationships within an image. In recent years, node-level classification using Graph Convolutional Networks (GCNs) has gained significant attention in the field of image classification, especially in applications involving high-resolution or structured data such as hyperspectral and UAV imagery. In this approach, each region or superpixel in an image is treated as a node in a graph, and the task is to assign a class label to each node. Numerous studies have demonstrated the effectiveness of GCNs in capturing spatial or spectral relationships between regions, thereby improving classification accuracy at the pixel or region level.

Despite these promising results, graph-level classification—where each entire image is represented as a graph and assigned a single label—remains relatively underexplored, particularly in the context of agricultural image analysis. This setting presents different challenges, such as how to construct graph representations that accurately reflect the holistic structure of an image and how to design models that can learn robust global features from such representations.

Motivated by the success of GCNs in node classification and the need for more scalable and generalizable models in agricultural tasks, this study shifts the focus toward graph-level classification. By representing each agricultural image as a graph and learning from the collective properties of its regions and their interrelations, we aim to explore the potential of GCNs in capturing global patterns and semantics that are critical for tasks such as crop type recognition or disease identification at the image level.

This study investigates the application of Graph Convolutional Networks (GCNs) for graph-level classification of agricultural images. While most existing works concentrate on node-level classification, this research focuses on the more challenging and less explored task of classifying entire images represented as graphs. The main contributions are as follows: 1.

Graph-Based Representation for Whole-Image Classification: We propose a framework that converts each image into a graph, enabling the application of GCNs to model spatial structures and semantic regions for comprehensive image classification.

In this section, the paper formally defines the problem addressed in this study, which is cast as a graph classification task—each agricultural image is treated as an individual graph, and the goal is to assign a corresponding class label to the entire graph. To enable this formulation, we first describe a graph construction methodology that transforms input images into graph representations by segmenting them into superpixels, defining nodes, extracting discriminative features, and establishing meaningful edge connections. The final subsection presents the proposed classification model, which employs a three-layer Graph Convolutional Network (GCN) to iteratively aggregate local node features and produce a robust graph-level representation for accurate classification.

In this study, we explored the effectiveness of Graph Convolutional Networks (GCNs) for image classification by leveraging superpixel-based graph representations. This task is crucial for many AI-based applications in modern agricultural models. While traditional CNNs have shown great success in this domain, they often require large computational resources due to their reliance on raw pixel data. By transforming images into graph structures, our proposed approach significantly reduces input size while maintaining high classification accuracy. Experimental results across multiple datasets demonstrate that the proposed method performs competitively with CNNs, achieving high accuracy even with reduced input complexity. The findings suggest that superpixel-based graph representations effectively capture spatial relationships within an image, enabling GCNs to extract meaningful features while minimizing computational costs. Moreover, our comparative analysis highlights the trade-offs between raw pixel-based CNNs and graph-based learning. While CNNs operate directly on pixel grids, GCNs benefit from structured feature learning, which enhances classification performance without excessive computational overhead. The success of our model across different datasets indicates its adaptability to various image resolutions and domains.

Although this study evaluated the proposed model on diverse set of agricultural image datasets, we acknowledge that the current evaluation does not fully validate the model’s generalization ability under more complex real-world conditions. Specifically, scenarios involving occlusions, lighting variations, and cluttered or dynamic backgrounds are underrepresented in the existing datasets. Therefore, an important future research direction is to collect more challenging datasets that better reflect such complexities. This would enable a more rigorous assessment of the model’s robustness and adaptability in practical agricultural applications.

Overall, our study underscores the potential of graph-based learning as a promising alternative to conventional deep learning methods in image classification. Future work can focus on optimizing graph construction techniques, exploring hybrid models that integrate CNN-based feature extraction with GCN-based relational learning, and extending the approach to more complex real-world applications.