Benchmarking hybrid CNN and transformer backbones with graph convolution networks (GCN) for flower growth-stage classification Authors: Aritra Das, Karib Shams, Mohammad Rifat Ahmmad Rashid, Raihan Ul Islam, Shamim H. Ripon, Ahmed Wasif Reza Published: 2026 ABSTRACT Accurate recognition of flower growth stages is important for plant phenotyping but remains challenging due to subtle visual differences and limited labeled data. This study proposes a hybrid CNN/Transformer + GCN framework for fine-grained flower growth-stage classification. A new dataset, BD Flower Growth, is introduced with 3,889 original images from eight Bangladeshi flower species, categorized into three stages (early, mid, full), forming 24 classes. The dataset is divided into training and testing sets, with augmentation applied only to the training data. Deep backbone networks are used to extract feature maps, which are transformed into graph representations and refined using Graph Convolutional Networks (GCN). A systematic ablation study is conducted by varying GCN depth (3, 5 layers), node resolution (7× 7, 14× 14), and graph construction methods (4-neighbour, 8-neighbour, and KNN with k=16). Experimental results show that performance depends strongly on both backbone and graph configuration. The best performance of 97% accuracy is achieved by EfficientNetV2, DenseNet201-based hybrid models, additionally Swin Transformer model shows the largest improvement, increasing from 84% to 97% after GCN integration. Across different settings, grid-based graphs (4- and 8-neighbour) consistently provide more stable and higher performance compared to KNN graphs, while moderate GCN depth (3–5 layers) offers the best balance accuracy. Cross-dataset evaluation on the Oxford 102 Flower dataset further demonstrates the generalization capability of the proposed approach. These findings highlight the effectiveness of hybrid graph-based learning and the importance of graph configuration in improving fine-grained classification. Introduction Flower development progresses through gradual and observable stages, beginning from early bud formation to partial bloom and eventually the full, open flower. These stages often reveal meaningful information about the plant’s growth pattern, health, and environmental response. For agriculture, horticulture, and ecological research, recognizing a flower’s growth stage can support decisions such as irrigation scheduling, harvesting time, and developmental monitoring. Despite this importance, stage identification is typically performed by manual observation, which is labor-intensive, subjective, and impractical when monitoring large numbers of plants. Deep learning has created new opportunities for automating plant-related tasks, but most existing datasets and studies focus on species classification or disease detection rather than identifying developmental stages. Fine-grained tasks like growth-stage classification remain challenging because the visual differences between early, mid, and full bloom are often subtle and depend on structural cues such as petal spread, shape, and coloration. Prior research also highlights that such fine-grained plant recognition becomes more difficult when images vary in lighting, environment, or background conditions1. To overcome these challenges, hybrid approaches that combine convolutional or transformer-based feature extraction with graph reasoning have gained interest. Graph Convolutional Networks (GCNs), in particular, can model relationships among extracted features and help distinguish nuanced visual patterns2. A major limitation in this domain is the scarcity of publicly available datasets that include multiple flower species with stage-wise class labels. Without sufficient labeled data, it is difficult to train models capable of understanding subtle developmental differences across species. To address this gap, a dataset has been developed called “BD Flower Growth”, containing 23,334 color images from eight flower species commonly found in Bangladesh3. The dataset includes 3,889 original photographs and 19,445 augmented samples, and each image is assigned a class label corresponding to one of three growth stages: early, mid, or full bloom – resulting in 24 total classes. Although simple in structure, class-labeled datasets are essential for classification-based phenotyping tasks and remain a foundation for practical deep-learning pipelines4. To classify these images, seven modern deep learning backbones has been used: ResNet50, ConvNeXt, EfficientNetV2-S, DenseNet201, MobileNetV3, ViT, and Swin Transformer. Importantly, all models were used together with a Graph Convolutional Network (GCN).The CNN/Transformer models extract visual features, and the GCN operates on these features to improve relational reasoning and fine-grained stage recognition. Initial experiments used a two-layer GCN, but this version showed clear overfitting when comparing the training and validation performance curves. After refining the design by increasing the number of GCN layers and improving regularization, the overfitting issue was significantly reduced. The final results were evaluated using accuracy, precision, recall, F1-score, and confusion matrices. Grad-CAM visualizations were also applied to understand which regions of the flowers contributed most to the model’s decisions. The contributions of this work are as follows: A large, high-quality dataset of eight Bangladeshi flower species, enriched with extensive augmentation for improved model robustness.Multi-class flower growth-stage classification that integrates convolutional and transformer backbones with Graph Convolutional Networks (GCNs) using feature-level graph construction.A comprehensive evaluation of seven backbone architectures (CNN and transformer-based) under a unified GCN framework, demonstrating performance variations across different feature representations.A systematic ablation study analyzing the impact of graph construction, including GCN depth (3 and 5 layers), node resolution (7×7 and 14×14), and connectivity patterns (4-neighbour, 8-neighbour, and KNN with k=16).A detailed comparative evaluation using standard metrics and Grad-CAM-based interpretability, providing insights into model behavior and highlighting discriminative regions across different configurations.Section “Introduction” presents the introduction, outlining the motivation, challenges, and contributions of this study. Section “Literature review” provides a comprehensive review of related literature, focusing on CNN–GCN hybrid models and graph-based learning approaches in plant phenotyping and image analysis. Section "Materials and methods" describes the materials and methods, including the proposed hybrid CNN/Transformer–GCN framework, dataset construction, preprocessing steps, model architectures, Ablation Setting and training configuration. Section “Experimental results” reports the experimental results and performance evaluation using multiple metrics, loss and accuracy analyses, confusion matrices, t-SNE and explainable AI visualizations. Section “Discussion” discusses the obtained results, compares the proposed approach with existing methods, and highlights key findings, limitations, and future research directions. Finally, Section “Conclusion” concludes the paper by summarizing the main contributions and outcomes of this work. A large, high-quality dataset of eight Bangladeshi flower species, enriched with extensive augmentation for improved model robustness. A large, high-quality dataset of eight Bangladeshi flower species, enriched with extensive augmentation for improved model robustness. Multi-class flower growth-stage classification that integrates convolutional and transformer backbones with Graph Convolutional Networks (GCNs) using feature-level graph construction. Multi-class flower growth-stage classification that integrates convolutional and transformer backbones with Graph Convolutional Networks (GCNs) using feature-level graph construction. A comprehensive evaluation of seven backbone architectures (CNN and transformer-based) under a unified GCN framework, demonstrating performance variations across different feature representations. A comprehensive evaluation of seven backbone architectures (CNN and transformer-based) under a unified GCN framework, demonstrating performance variations across different feature representations. A systematic ablation study analyzing the impact of graph construction, including GCN depth (3 and 5 layers), node resolution (7×7 and 14×14), and connectivity patterns (4-neighbour, 8-neighbour, and KNN with k=16). A systematic ablation study analyzing the impact of graph construction, including GCN depth (3 and 5 layers), node resolution (7×7 and 14×14), and connectivity patterns (4-neighbour, 8-neighbour, and KNN with k=16). A detailed comparative evaluation using standard metrics and Grad-CAM-based interpretability, providing insights into model behavior and highlighting discriminative regions across different configurations. A detailed comparative evaluation using standard metrics and Grad-CAM-based interpretability, providing insights into model behavior and highlighting discriminative regions across different configurations. Literature review GCN-based modeling in plant phenotyping and structural growth analysis Recent advances in graph-based deep learning have significantly improved the modeling of structural, spectral, and relational dependencies in plant phenotyping. In4, hyperspectral imaging is combined with a Graph Convolutional Network (GCN) to predict blueberry physiological traits under drought stress. Plants are represented as graph nodes with spectral features, while edges encode spectral similarity, enabling the Plant-GCN to capture complex inter-plant relationships and predict six physiological parameters with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values up to 0.94. However, the approach relies on fixed spectral graph construction and controlled greenhouse environments, limiting field-scale adaptability. To address genotype–environment interactions, the author5 introduces a Cross-Omics GCN framework that integrates genomics, metabolomics, and image-derived traits for maize phenotype prediction. Separate GCNs model each modality, followed by tensor fusion to capture G×E×M interactions, improving yield prediction performance from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2 = 0.32$$\end{document} to 0.43. Graph learning has also been extended to 3D plant structure analysis. The GCASSN model6 employs a GCN–attention architecture for plant organ segmentation from point clouds, achieving state-of-the-art accuracy on the Plant3D and Phone4D datasets, though requiring dense labeled 3D data and high computational cost. CNN–GCN hybrid frameworks for plant image analysis Hybrid CNN–GCN architectures combine deep visual feature extraction with relational reasoning to improve plant image understanding. For fine-grained plant species classification7, combines lightweight CNN backbones with Graph Attention Networks and knowledge distillation, modeling image samples as graph nodes connected by visual similarity. This strategy enables compact models to achieve classification accuracies between 91% and 97% with low latency, making them suitable for edge deployment, although fixed graph construction limits adaptability to unseen data. Vision–language integration has also been explored to enhance representation learning. CEPrompt8 introduces cross-modal emotion-aware prompting to guide visual feature extraction using semantic textual cues, improving recognition robustness in challenging conditions. Addressing uncertainty in deep feature representations, UA-FER9 incorporates evidential deep learning and vision–language pre-training to calibrate uncertain relationships, leading to improved robustness and generalization in expression recognition tasks. Label-efficient learning is explored in10, where CNN-extracted features are used to construct graphs for semi-supervised crop and weed recognition. The framework propagates labels from limited annotated data to unlabeled samples and achieves up to 97.8% accuracy on corn datasets. Graph-based modeling has also been widely used for plant disease classification. PND-Net11integrates CNN feature extraction with GCN-based regional aggregation to represent multi-scale leaf regions as graph nodes, improving robustness across crops. Similarly12, refines CNN-extracted regional features using spatial adjacency graphs and reports accuracies up to 99%. The hybrid GCN–GAT framework in13 converts leaf images into region adjacency graphs using SLIC superpixels and achieves up to 99.73% accuracy across datasets, though static graph construction limits modeling of disease progression over time. Similarly, IncepTR14 combines Inception-based multi-scale convolution with vision transformers and attention mechanisms to capture subtle local and global variations, demonstrating strong performance in micro-expression recognition. Extended applications of graph-based learning relevant to plant growth and flower development Beyond plant imaging, CNN–GCN hybrid frameworks have demonstrated strong performance in hyperspectral image analysis. The local–global graph fusion approach in15constructs multiple graphs from local patches and global anchors and fuses them via cross-attention, achieving overall accuracies up to 99.54% on Indian Pines, University of Pavia, and Salinas datasets. To address scalability16, proposes a dual-branch CNN–MiniGCN architecture that combines spatial–spectral CNN features with batch-based graph learning, achieving 92.80% overall accuracy on Indian Pines while reducing memory overhead. HFENet17 further improves robustness using multi-scale convolution and spectral attention, achieving accuracies exceeding 99.5% across multiple hyperspectral benchmarks. Graph-based learning has also been applied to remote sensing image segmentation, where an attention-based GCN in18 refines superpixel-level CNN features for SAR segmentation, and SGA-Net19 introduces self-constructing graph attention to improve performance on ISPRS Potsdam and Vaihingen datasets. In medical imaging, graph-enhanced deep learning has achieved strong diagnostic performance. The work in7 applies transfer learning with VGG-based CNNs for brain tumor classification, achieving 99.11% accuracy, while CSHG-CervixNet20 leverages hypergraph convolution to achieve 99.31% accuracy in cervical cancer classification. The GATransformer model21 further integrates graph attention with Transformers to enhance explainability and inter-channel dependency modeling. Graph neural networks have also been extended to temporal modeling, where the hybrid ST-GCN and 1D-CNN framework in22 improves action recognition with 91.7% Top-1 accuracy, and H2GCN23 captures higher-order dependencies using hypergraph convolution. At the molecular level, graph-based network analysis provides biological insights into plant development. The gene co-expression network in24 identifies key regulatory modules controlling flowering in Hydrangea arborescens, while25 integrates transcriptomic and metabolomic networks to explain regulatory mechanisms underlying dynamic flower color changes. Table 1 illustrates the summary of the current literature review.Table 1Summary of the review.RefProblem AddressedDatasetModel/MethodResultLimitation4Blueberry physiological trait prediction under drought stressHyperspectral images + 6 physiological measuresPlant-GCN with spectral similarity graph\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} up to 0.94Controlled greenhouse only; static similarity edges5Maize phenotype prediction with multi-omics interactions (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G \times E \times M$$\end{document})156 RIL lines; SNPs, metabolites, image traitsMulti-omics GCN with cross-omics tensor fusion\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}: 0.32 → 0.43Large multi-omics data needed; cubic tensor cost6Organ-level 3D plant segmentationPlant3D and Phone4DGCN + Attention (GCASM) + Trans-net95.46% Acc., 90.41% IoUDense 3D labels; high computation7Fine-grained plant species classificationHoya flower and leaf datasetsCNN teacher → GAT student distillation91–97%Graph similarity may miss relations10Weed–crop recognition with limited labels6,000 images (corn, lettuce, radish, mixed)ResNet-101 features + GCN label propagation97.80% (Corn)Manual graph construction11Plant nutrient deficiency and disease classificationBanana, Coffee, Potato, PlantDocCNN backbone + GCN aggregation (PND-Net)96.18% (Potato)No growth-stage modeling26Crop variety suitability recommendationMulti-site phenotypic + climate dataHeterogeneous GNN (G × E)High prediction accuracyLimited traits; site tuning needed12Plant disease classificationPotato, tomato, sugarcaneCNN (ResNet) + 2–3 GCN layersUp to 99%Region detection dependent13Leaf disease classification with graphsSugarcane, Potato, AppleSLIC → RAG + GCN + GAT99.73%, 98.11%, 91.03%Static graph; no temporal modeling Table 1Summary of the review.RefProblem AddressedDatasetModel/MethodResultLimitation4Blueberry physiological trait prediction under drought stressHyperspectral images + 6 physiological measuresPlant-GCN with spectral similarity graph\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} up to 0.94Controlled greenhouse only; static similarity edges5Maize phenotype prediction with multi-omics interactions (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G \times E \times M$$\end{document})156 RIL lines; SNPs, metabolites, image traitsMulti-omics GCN with cross-omics tensor fusion\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}: 0.32 → 0.43Large multi-omics data needed; cubic tensor cost6Organ-level 3D plant segmentationPlant3D and Phone4DGCN + Attention (GCASM) + Trans-net95.46% Acc., 90.41% IoUDense 3D labels; high computation7Fine-grained plant species classificationHoya flower and leaf datasetsCNN teacher → GAT student distillation91–97%Graph similarity may miss relations10Weed–crop recognition with limited labels6,000 images (corn, lettuce, radish, mixed)ResNet-101 features + GCN label propagation97.80% (Corn)Manual graph construction11Plant nutrient deficiency and disease classificationBanana, Coffee, Potato, PlantDocCNN backbone + GCN aggregation (PND-Net)96.18% (Potato)No growth-stage modeling26Crop variety suitability recommendationMulti-site phenotypic + climate dataHeterogeneous GNN (G × E)High prediction accuracyLimited traits; site tuning needed12Plant disease classificationPotato, tomato, sugarcaneCNN (ResNet) + 2–3 GCN layersUp to 99%Region detection dependent13Leaf disease classification with graphsSugarcane, Potato, AppleSLIC → RAG + GCN + GAT99.73%, 98.11%, 91.03%Static graph; no temporal modeling Summary of the review. Materials and methods Methodology Fig. 1Workflow diagram. Workflow diagram. In this study, a hybrid CNN–GCN framework is proposed to effectively capture both local visual patterns and spatial relational information for flower growth stage classification. Fig. 1 shows that initially input images are resized and normalized, followed by standard data augmentation techniques to enhance robustness. A deep CNN backbone–selected from ConvNeXt, EfficientNetV2-S, Swin Transformer, ViT-S/16, ResNet50, DenseNet201, or MobileNetV3–is employed to extract high-level spatial feature maps from the input images. The CNN produces a feature tensor size, which is subsequently refined using adaptive average pooling to obtain a fixed spatial resolution. The pooled feature map is then reshaped such that each spatial location corresponds to a graph node, with the associated CNN feature vector serving as the node attribute. A grid-based graph is constructed by connecting neighboring spatial nodes, preserving the inherent spatial structure of the feature map. This graph representation is processed using a stack of graph convolutional layers, where successive GCN layers model spatial dependencies and contextual relationships among the CNN-extracted features. Batch normalization and ReLU activation are applied after each GCN layer, with dropout introduced in the final layer to reduce overfitting. To obtain compact representations, global mean pooling is applied independently to the CNN feature map and the GCN output, resulting in a global CNN feature vector and a graph-refined GCN feature vector, respectively. These two complementary representations are fused through feature concatenation, combining appearance-based and relational information. The fused feature vector is finally passed through a fully connected layer followed by a softmax function to perform multi-class classification of flower growth stages. Dataset description For this work, a comprehensive and custom-built dataset named BD Flower Growth was utilized3. The dataset encompasses eight common flower species found in Bangladesh: Bush Allamanda, Red Hibiscus, Yellow Bell, Pinwheel Flower, Pink Periwinkle, White Madagascar Periwinkle, Marvel of Peru, and White Hibiscus. Each species is represented across three growth stages–Early Stage (initial bud formation), Mid Stage (partial blooming), and Full Stage (complete bloom)–resulting in 24 distinct classes. A total of 3,889 original captured. Fig. 2 shows sample images from our dataset.Fig. 2Flowers name and its corresponding growth stage. Flowers name and its corresponding growth stage. Dataset collection process The dataset was collected over a period of almost four months, from June 5 to September 24, 2025, in a residential area of Aftabnagar, Dhaka, Bangladesh. Image collection was carried out on a home balcony so that the flowers could be photographed under natural outdoor conditions similar to everyday garden environments. Eight common flower species found in Bangladesh–Bush Allamanda, Red Hibiscus, Yellow Bell, Pinwheel Flower, Pink Periwinkle, White Madagascar Periwinkle, Marvel of Peru, and White Hibiscus–were selected, and individual flowers were followed from the moment the bud appeared until the flower was fully bloomed. To capture natural changes in appearance throughout the day, images were taken three times daily–morning (Between 9:00–10:00 AM), evening (Between 5:00–6:00 PM), and night (1:00–2:00 AM)–with an interval of about eight hours between sessions. All photographs were captured using smartphone cameras. The growth of each flower was grouped into three stages (early, mid, and full) based on visible development. In total, 3,889 original images were obtained. Dataset pre-processing After the images were collected, all images were resized to a uniform resolution of 640 × 640 pixels. Standardizing the image dimensions ensures consistency across the dataset and enables more efficient processing by the model during training. The dataset was divided into 80% for training, 10% for validation, and 10% for testing using an instance-level split protocol. Images belonging to the same flower instance were assigned to only one subset, ensuring that no flower instance appeared across multiple subsets. This strategy prevents data leakage between training and evaluation sets and provides a more reliable assessment of the model’s generalization capability. To further enhance the dataset and improve the model’s generalization capability, five augmentation techniques were applied to the training set. The validation and testing sets remained unchanged, with no augmentation applied. These augmentations not only increased the size of the training dataset but also introduced realistic variations that simulate changes in lighting, position, and orientation commonly observed in real-world scenarios. Figure 3 illustrates an example of an original image alongside its corresponding augmented version.Fig. 3Augmented images. Augmented images. Baseline models To establish a strong and fair benchmark, several state-of-the-art convolutional neural networks and transformer-based architectures are used as baseline models. These architectures span conventional deep CNNs, efficiency-oriented lightweight networks, and attention-driven vision transformers, providing a broad evaluation of performance across different design philosophies. ConvNeXt ConvNeXt27 is a modern convolutional neural network that integrates design principles inspired by Vision Transformers while retaining the inductive biases and efficiency of CNNs. It utilizes large-kernel depthwise convolutions, inverted bottleneck blocks, and layer normalization to enhance feature representation. The hierarchical architecture enables effective multi-scale feature extraction, making ConvNeXt well suited for capturing both local textures and global flower structure. EfficientNetV2-S EfficientNetV2-S28 is an efficiency-optimized CNN that improves training speed and performance through fused-MBConv blocks and refined compound scaling. By jointly scaling network depth, width, and resolution, the model achieves strong accuracy with reduced computational cost. This makes EfficientNetV2-S particularly effective for scenarios where efficiency and performance must be balanced. Swin Transformer V2 Swin Transformer V229 is a hierarchical vision transformer that computes self-attention within shifted local windows, allowing scalable processing of high-resolution images. The shifted-window mechanism enables information exchange across neighboring regions while maintaining linear computational complexity. This architecture effectively captures both local patterns and long-range dependencies, which are important for modeling spatial variations in flower growth stages. Vision Transformer (ViT-B/16) Vision Transformer (ViT-B/16)30 models images as sequences of fixed-size patches processed by a standard Transformer encoder. By leveraging global self-attention, ViT captures long-range spatial relationships across the entire image. Positional embeddings preserve spatial information, enabling effective global context modeling, though the architecture typically requires larger datasets to generalize well. ResNet50 ResNet5031 is a widely used deep convolutional network that introduces residual connections to ease optimization and mitigate vanishing gradient issues. The residual learning framework allows the network to learn incremental refinements over identity mappings, facilitating stable training of deep architectures. ResNet50 serves as a strong and reliable baseline for visual feature extraction. DenseNet201 DenseNet20132 employs dense connectivity, where each layer receives feature maps from all preceding layers through concatenation. This design encourages feature reuse, strengthens gradient flow, and improves parameter efficiency. DenseNet201 is particularly effective in capturing fine-grained visual patterns, which are important for distinguishing subtle differences between flower growth stages. MobileNetV3 MobileNetV333 is a lightweight CNN architecture designed for deployment on resource-constrained devices. It combines depthwise separable convolutions, inverted residual blocks, and squeeze-and-excitation modules to enhance channel-wise feature representation with minimal computational overhead. MobileNetV3 provides an efficient baseline for evaluating performance under strict efficiency constraints. Graph convolutional network (GCN) Graph Convolutional Networks (GCNs)34 extend the idea of convolution from grid-structured data (e.g., images) to non-Euclidean graph structures. A GCN learns node representations by aggregating information from its neighbors. For a graph G = (V, E), with adjacency matrix A and degree matrix D. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{A} = A + I$$\end{document} is the adjacency matrix with self-connections, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(l)}$$\end{document} is the node embedding at layer l, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W^{(l)}$$\end{document} is the learnable weight matrix, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma (\cdot )$$\end{document} is an activation function. In this work, the feature maps extracted from the CNN backbone are reshaped into a set of spatial nodes. A 4-neighbor grid graph is constructed, enabling the GCN to operate on local receptive-field relationships. Three stacked GCN layers progressively refine the node features, allowing the model to capture structural correlations that standard CNNs may ignore. Since flower growth characteristics often depend on relative structural patterns (petal arrangements, bud expansion, and texture continuity), modeling spatial relations explicitly provides complementary information beyond pixel-wise convolutional features. Proposed hybrid model (CNN or Transformer + GCN) The principal contribution of this work is a hybrid deep learning architecture that combines visual feature extraction from CNN or Transformer backbones with relational reasoning using a Graph Convolutional Network (GCN)34. This design jointly captures appearance-based information and spatial dependencies, which are important for fine-grained flower growth-stage classification. The hybrid framework consists of five stages: (i) baseline feature extraction, (ii) graph construction, (iii) GCN-based relational reasoning, (iv) feature fusion, and (v) classification. Baseline feature extraction Given an input image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{I}$$\end{document}, a backbone network \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {B}(\cdot )$$\end{document} extracts a high-level feature tensor:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{F} = \mathcal {B}(\textbf{I}), \quad \textbf{F} \in \mathbb {R}^{C \times H \times W} \end{aligned}$$\end{document}where C denotes the number of feature channels, and H and W represent the spatial dimensions of the feature map. Graph construction from CNN features The feature tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{F}$$\end{document} is reshaped into a set of node features:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{X} = \{ \textbf{x}_i \}_{i=1}^{N}, \quad \textbf{x}_i \in \mathbb {R}^{C}, \quad N = H \times W \end{aligned}$$\end{document}A graph is then defined as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {G} = (\mathcal {V}, \mathcal {E}) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {V} = \{1, 2, \dots , N\}$$\end{document} represents the set of nodes and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {E}$$\end{document} represents the edges connecting nodes. To model spatial and feature-level relationships, multiple graph construction strategies are explored:4-neighbour graph (4N): Each node is connected to its immediate top, bottom, left, and right neighbors.8-neighbour graph (8N): Each node is connected to all surrounding neighbors, including diagonal connections.K-nearest neighbour graph (KNN): Each node is connected to its k most similar nodes in the feature space based on Euclidean distance. In this work, k = 16 is used.The graph construction strategy is treated as a hyperparameter and evaluated through ablation experiments. The “4N”, “8N”, and “KNN (k=16)” configurations reported in the experimental tables correspond to these graph definitions. 4-neighbour graph (4N): Each node is connected to its immediate top, bottom, left, and right neighbors. 4-neighbour graph (4N): Each node is connected to its immediate top, bottom, left, and right neighbors. 8-neighbour graph (8N): Each node is connected to all surrounding neighbors, including diagonal connections. 8-neighbour graph (8N): Each node is connected to all surrounding neighbors, including diagonal connections. K-nearest neighbour graph (KNN): Each node is connected to its k most similar nodes in the feature space based on Euclidean distance. In this work, k = 16 is used. K-nearest neighbour graph (KNN): Each node is connected to its k most similar nodes in the feature space based on Euclidean distance. In this work, k = 16 is used. GCN-based relational reasoning The node features are refined through stacked graph convolution layers:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{H}^{(l)} = \textrm{GCN}_{l}(\textbf{H}^{(l-1)}, \mathcal {G}), \quad l=1,2,3 \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{H}^{(0)} = \textbf{X}$$\end{document} denotes the initial node features. Each GCN layer updates node representations by aggregating information from neighboring nodes. The graph \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {G}$$\end{document} defines the neighborhood relationships used for message passing. A graph-level representation is obtained by global average pooling over all node embeddings:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{g} = \frac{1}{N} \sum _{i=1}^{N} \textbf{H}^{(3)}_i \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{H}^{(3)}_i \in \mathbb {R}^{d}$$\end{document} denotes the feature vector of the i-th node in the final GCN layer, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g} \in \mathbb {R}^{d}$$\end{document} is the graph-level representation. Global CNN feature aggregation In parallel, the global CNN feature is computed as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{c} = \frac{1}{N} \sum _{i=1}^{N} \textbf{x}_i \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{c} \in \mathbb {R}^{C}$$\end{document} summarizes the overall visual representation of the input image. Feature fusion The CNN feature and graph feature are concatenated:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{z} = [\textbf{c} \,\Vert \, \textbf{g}] \in \mathbb {R}^{C+d} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{z}$$\end{document} combines both semantic and relational information. Classification layer The fused feature vector is passed to a linear classifier:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\textbf{y}} = \textrm{softmax}(\textbf{W}\textbf{z} + \textbf{b}) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}\in \mathbb {R}^{K\times (C+d)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}\in \mathbb {R}^{K}$$\end{document} are trainable parameters, and K is the number of output classes. By integrating backbone features with graph-based relational reasoning, the proposed hybrid model captures both visual and structural characteristics of flower growth stages, improving discrimination between visually similar classes. Algorithm 1Hybrid GCN framework for classification. Hybrid GCN framework for classification. Ablation study To better understand the contribution of different design choices in the proposed hybrid framework, a structured ablation study was conducted. The analysis focuses on the following key components:Graph construction: Three strategies were explored, including grid graphs with 4-neighbour connectivity, grid graphs with 8-neighbour connectivity, and KNN-based graphs (k=16), to evaluate how different neighborhood definitions affect spatial information modeling.GCN depth: Both 3-layer and 5-layer GCN architectures were considered to examine the impact of network depth on feature propagation and representation learning.Node resolution: Two spatial resolutions (7 × 7 and 14 × 14) were used to construct graph nodes. The 7 × 7 grid provides a lower-resolution representation with fewer nodes, while the 14 × 14 grid captures finer spatial details at the cost of increased computational complexity.Each configuration was evaluated using two random seeds (42 and 1000), and the results are reported as mean ± standard deviation. This ablation setup was consistently applied across all seven backbone models to ensure a fair comparison. Graph construction: Three strategies were explored, including grid graphs with 4-neighbour connectivity, grid graphs with 8-neighbour connectivity, and KNN-based graphs (k=16), to evaluate how different neighborhood definitions affect spatial information modeling. Graph construction: Three strategies were explored, including grid graphs with 4-neighbour connectivity, grid graphs with 8-neighbour connectivity, and KNN-based graphs (k=16), to evaluate how different neighborhood definitions affect spatial information modeling. GCN depth: Both 3-layer and 5-layer GCN architectures were considered to examine the impact of network depth on feature propagation and representation learning. GCN depth: Both 3-layer and 5-layer GCN architectures were considered to examine the impact of network depth on feature propagation and representation learning. Node resolution: Two spatial resolutions (7 × 7 and 14 × 14) were used to construct graph nodes. The 7 × 7 grid provides a lower-resolution representation with fewer nodes, while the 14 × 14 grid captures finer spatial details at the cost of increased computational complexity. Node resolution: Two spatial resolutions (7 × 7 and 14 × 14) were used to construct graph nodes. The 7 × 7 grid provides a lower-resolution representation with fewer nodes, while the 14 × 14 grid captures finer spatial details at the cost of increased computational complexity. Explainable AI (XAI) To interpret the decisions of the deep learning models in flower growth stage classification, Explainable AI (XAI) techniques were employed. Class Activation Mapping (CAM)-based methods help visualize the regions of an image that most influence the model’s predictions. In this study, Ablation-CAM was used to generate class-specific heatmaps. Unlike gradient-based methods such as Grad-CAM, Ablation-CAM evaluates the importance of feature map regions by selectively removing them and observing the effect on the prediction. This provides more reliable insights into the model’s decision-making process. For the hybrid GCN model, node-level analysis further highlights how spatial relationships between image patches contribute to classification. The results show that the model focuses on meaningful structures such as petals, buds, and leaves, improving trust and supporting the interpretability of the model for practical applications. Evaluation metrics To evaluate the performance of the proposed model, four widely used classification metrics are employed: Accuracy, Precision, Recall, and F1-score. These metrics provide a comprehensive assessment of overall correctness, class-wise relevance, and prediction balance. Accuracy measures the proportion of correctly classified samples among all samples:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} \end{aligned}$$\end{document}Precision indicates the correctness of positive predictions by measuring how many predicted positive samples are truly positive:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Precision} = \frac{TP}{TP + FP} \end{aligned}$$\end{document}Recall reflects the model’s ability to correctly identify all relevant positive samples:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Recall} = \frac{TP}{TP + FN} \end{aligned}$$\end{document}F1-score represents the harmonic mean of Precision and Recall, providing a balanced evaluation metric:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {F1-score} = 2 \times \frac{\text {Precision} \times \text {Recall}}{\text {Precision} + \text {Recall}} \end{aligned}$$\end{document}Here, TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives, respectively. Experimental setup and hyperparameters The proposed model was evaluated using the custom image dataset3, which was partitioned into training, validation, and testing subsets following an 80%/10%/10% instance-level split protocol. This ensured that images from the same flower instance were not shared across subsets. Different graph construction strategies (4-neighbour, 8-neighbour, and KNN with k=16) were evaluated as part of the ablation study. All experiments were conducted on a Kaggle environment equipped with an NVIDIA Tesla T4 GPU, which provided sufficient computational resources for efficient training and evaluation of the deep CNN or Transformer–GCN architecture. Hyper-parameter Table 2Training configuration and ablation settings.Training ParameterConfigurationData Augmentationaugmentation techniques applied to the training set: Shifting, Horizontal Flip, Random Rotation, Zooming, and Random Brightness AdjustmentBackbone ModelsConvNeXt_Base, EfficientNetV2-S, ResNet50, DenseNet201, MobileNetV3, ViT-B/16, Swin TransformerBackbone InitializationImageNet pretrained weightsBackbone Fine-TuningFull end-to-end fine-tuningGraph ConstructionGrid Graph (4-neighbour, 8-neighbour), KNN Graph (k=16)GCN Layers3-layer and 5-layer configurations (ablation study)Node Resolution7 × 7 and 14 × 14 spatial grids for graph constructionGCN Hidden DimensionsBackbone-dependent and automatically adapted to feature dimensionalityNormalizationBatch Normalization applied after each GCN layerDropout Rate0.5Feature FusionConcatenation of CNN and GCN featuresOptimizerAdamWLearning Rate SchedulerCosine Annealing with Warm RestartsWeight Decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document} Training Epochs30Evaluation ProtocolTwo random seeds (42, 1000); results reported as Mean ± StdLoss FunctionCross-Entropy Loss Table 2Training configuration and ablation settings.Training ParameterConfigurationData Augmentationaugmentation techniques applied to the training set: Shifting, Horizontal Flip, Random Rotation, Zooming, and Random Brightness AdjustmentBackbone ModelsConvNeXt_Base, EfficientNetV2-S, ResNet50, DenseNet201, MobileNetV3, ViT-B/16, Swin TransformerBackbone InitializationImageNet pretrained weightsBackbone Fine-TuningFull end-to-end fine-tuningGraph ConstructionGrid Graph (4-neighbour, 8-neighbour), KNN Graph (k=16)GCN Layers3-layer and 5-layer configurations (ablation study)Node Resolution7 × 7 and 14 × 14 spatial grids for graph constructionGCN Hidden DimensionsBackbone-dependent and automatically adapted to feature dimensionalityNormalizationBatch Normalization applied after each GCN layerDropout Rate0.5Feature FusionConcatenation of CNN and GCN featuresOptimizerAdamWLearning Rate SchedulerCosine Annealing with Warm RestartsWeight Decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document} Training Epochs30Evaluation ProtocolTwo random seeds (42, 1000); results reported as Mean ± StdLoss FunctionCross-Entropy Loss Training configuration and ablation settings. The table 2 shows that the training configuration is designed to balance model capacity and generalization by combining controlled data augmentation, deep GCN integration, and feature fusion. Full backbone training with batch normalization and dropout helps mitigate overfitting while enabling effective joint learning of CNN and graph-based representations. The use of AdamW with cosine annealing ensures stable optimization and improved convergence across all models. Class distribution and imbalance handling The dataset exhibits a moderate class imbalance across different growth stages. This imbalance arises naturally due to the non-uniform duration and occurrence frequency of each growth stage during data collection. To address this issue, several strategies were employed. First, data augmentation techniques, including rotation, flipping, zooming and brightness adjustment, were applied to increase the diversity and representation of underrepresented classes during model training only. Second, a balanced sampling strategy was used during training to ensure that each mini-batch contains a more uniform distribution of samples across classes. In addition, model performance was evaluated using Ablation-CAM, Confusion Matrix and class-wise metrics such as precision, recall, and F1-score, alongside overall accuracy. This ensures that the performance on minority classes is properly assessed and not overshadowed by majority classes. These measures help improve the robustness of the model and reduce bias toward dominant classes. Experimental results Baseline models Table 3Performance comparison of different backbone models.ModelAccPrecRecF1ResNet500.940.950.930.93DenseNet2010.930.950.940.94EfficientNet_V2_S0.960.970.960.96MobileNetV30.960.970.930.94ConvNeXt_Base0.940.950.900.91ViT_B_160.930.960.920.94Swin0.840.870.820.82 Table 3Performance comparison of different backbone models.ModelAccPrecRecF1ResNet500.940.950.930.93DenseNet2010.930.950.940.94EfficientNet_V2_S0.960.970.960.96MobileNetV30.960.970.930.94ConvNeXt_Base0.940.950.900.91ViT_B_160.930.960.920.94Swin0.840.870.820.82 Performance comparison of different backbone models. Table 3 presents the performance comparison of different backbone models. Among all models, EfficientNet_V2_S achieves the best overall performance with the highest accuracy and F1-score, indicating strong feature extraction capability. In contrast, transformer-based models such as Swin show comparatively lower performance, while CNN-based models maintain more consistent results across all evaluation metrics. Ablation study Resnet50 Table 4Analysis of GCN layer and graph construction under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.4495.7795.11 ± 0.6638-neigh95.7194.9795.84 ± 0.1354-neigh95.5095.2495.37 ± 0.1358-neigh94.1295.4294.77 ± 0.923KNN (k=16)93.9294.9794.44 ± 0.535KNN (k=16)94.1893.9294.05 ± 0.13Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.2496.0395.63 ± 0.4038-neigh95.2494.4494.84 ± 0.4054-neigh95.2496.3095.77 ± 0.5358-neigh94.7194.7194.71 ± 0.003KNN (k=16)94.1895.5094.84 ± 0.665KNN (k=16)93.6595.5094.58 ± 0.93 Table 4Analysis of GCN layer and graph construction under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.4495.7795.11 ± 0.6638-neigh95.7194.9795.84 ± 0.1354-neigh95.5095.2495.37 ± 0.1358-neigh94.1295.4294.77 ± 0.923KNN (k=16)93.9294.9794.44 ± 0.535KNN (k=16)94.1893.9294.05 ± 0.13Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.2496.0395.63 ± 0.4038-neigh95.2494.4494.84 ± 0.4054-neigh95.2496.3095.77 ± 0.5358-neigh94.7194.7194.71 ± 0.003KNN (k=16)94.1895.5094.84 ± 0.665KNN (k=16)93.6595.5094.58 ± 0.93 Analysis of GCN layer and graph construction under different node resolutions. Table 4 analyzes the influence of GCN layer and graph construction at two node resolutions. At the 7 × 7 resolution, the 3-layer GCN with 8-neighbour connectivity achieves the highest mean accuracy of 95.84%, outperforming both the 4-neighbour (95.11%) and KNN graphs (94.44%). Increasing the layer to 5 layers does not consistently improve performance, as the accuracy slightly decreases or remains similar across all graph types, suggesting over-smoothing. At the higher 14 × 14 resolution, the best performance of 95.77% is obtained with a 5-layer 4-neighbour graph, indicating that deeper GCNs become more effective when spatial resolution is increased.Table 5Performance comparison of graph types under several GCN layers and resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.950.935 Layer0.940.940.92Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.930.940.935 Layer0.930.910.93 Table 5Performance comparison of graph types under several GCN layers and resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.950.935 Layer0.940.940.92Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.930.940.935 Layer0.930.910.93 Performance comparison of graph types under several GCN layers and resolutions. Table 5 compares grid-based and KNN graph structures across different GCN layers and node resolutions. For the 7 × 7 setting, both 4-neighbour and 8-neighbour graphs achieve higher accuracy (0.95) than the KNN graph (0.93) with a 3-layer GCN, while all methods experience a small drop when layer increases to 5 layers. A similar trend is observed at the 14 × 14 resolution, where the 8-neighbour graph performs best at 0.94 with 3 layers, but its performance declines noticeably to 0.91 when the layer is increased. Overall, grid-based graphs consistently provide more stable and higher performance than KNN graphs across both resolutions. DesnseNet201 Table 6Analysis of GCN layer and neighbor selection under different seeds.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.3095.5095.90 ± 0.4038-neigh95.5095.5095.50 ± 0.0054-neigh95.7795.5095.63 ± 0.1358-neigh96.7795.3096.03 ± 0.263KNN (k=16)95.2496.3095.77 ± 0.535KNN (k=16)95.7795.5095.63 ± 0.13Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.5695.2495.90 ± 0.6638-neigh96.3095.2495.77 ± 0.5354-neigh95.5096.8396.16 ± 0.6658-neigh96.5696.0396.30 ± 0.263KNN (k=16)95.2495.2495.24 ± 0.005KNN (k=16)95.5095.5095.50 ± 0.00 Table 6Analysis of GCN layer and neighbor selection under different seeds.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.3095.5095.90 ± 0.4038-neigh95.5095.5095.50 ± 0.0054-neigh95.7795.5095.63 ± 0.1358-neigh96.7795.3096.03 ± 0.263KNN (k=16)95.2496.3095.77 ± 0.535KNN (k=16)95.7795.5095.63 ± 0.13Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.5695.2495.90 ± 0.6638-neigh96.3095.2495.77 ± 0.5354-neigh95.5096.8396.16 ± 0.6658-neigh96.5696.0396.30 ± 0.263KNN (k=16)95.2495.2495.24 ± 0.005KNN (k=16)95.5095.5095.50 ± 0.00 Analysis of GCN layer and neighbor selection under different seeds. From the results presented in Table 6, it can be seen that the 3-layer GCN with 4 neighbors achieved a high mean accuracy of 95.90% in 7 × 7 setting, slightly outperforming the 3-layer configuration with 8 neighbors, which had a mean of 95.50%. Increasing the layer to 5 layers generally improved performance for larger neighbor sizes, with the 5-layer + 8-neighbor setup reaching 96.03%, indicating that deeper GCNs benefit from more connected graphs. For the KNN graph with k=16, the results were more variable, with the 3-layer model showing a mean of 95.77% and the 5-layer model slightly lower at 95.63%. In 14 × 14 setting, similar trends were observed, with the 5-layer + 8-neighbor configuration achieving the highest mean accuracy of 96.30%, demonstrating consistent gains from both layer and neighbor expansion. These comparisons suggest that GCN layer and neighbor selection significantly influence classification performance, with optimal results obtained using moderate layer and well-connected graphs.Table 7Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.930.955 Layer0.940.970.94Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.940.955 Layer0.950.940.95 Table 7Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.930.955 Layer0.940.970.94Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.940.955 Layer0.950.940.95 Performance comparison of graph types under different GCN layers and grid resolutions. The results in Table 7 indicate that the 3-layer GCN performs consistently well across different neighbor configurations, with slightly higher accuracy on the KNN graph (0.95) for both resolutions. Increasing to 5 layers improves performance for certain configurations, such as the 8-neighbour graph at 7 × 7 resolution (0.97), showing that deeper GCNs can leverage more connected neighbors. Overall, KNN graphs maintain strong performance across layers and resolutions, while the 4- and 8-neighbour setups show minor fluctuations. These comparisons suggest that both GCN layer and graph construction influence accuracy, with KNN connectivity providing more stable results. EfficientNet_V2 Table 8Ablation study on GCN layer and neighbor selection using EfficientNet_V2 backbone under different seeds.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh97.0996.8396.96 ± 0.1338-neigh96.0396.5696.30 ± 0.2654-neigh97.0996.5696.83 ± 0.2658-neigh96.8396.5696.69 ± 0.133KNN (k=16)94.8694.4094.63 ± 0.335KNN (k=16)96.8396.0396.43 ± 0.40Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.8396.3096.56 ± 0.2638-neigh96.8396.8396.83 ± 0.0054-neigh95.8396.3095.56 ± 0.2658-neigh96.2996.2996.29 ± 0.003KNN (k=16)95.7694.3194.89 ± 0.815KNN (k=16)96.5696.5696.56 ± 0.00 Table 8Ablation study on GCN layer and neighbor selection using EfficientNet_V2 backbone under different seeds.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh97.0996.8396.96 ± 0.1338-neigh96.0396.5696.30 ± 0.2654-neigh97.0996.5696.83 ± 0.2658-neigh96.8396.5696.69 ± 0.133KNN (k=16)94.8694.4094.63 ± 0.335KNN (k=16)96.8396.0396.43 ± 0.40Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.8396.3096.56 ± 0.2638-neigh96.8396.8396.83 ± 0.0054-neigh95.8396.3095.56 ± 0.2658-neigh96.2996.2996.29 ± 0.003KNN (k=16)95.7694.3194.89 ± 0.815KNN (k=16)96.5696.5696.56 ± 0.00 Ablation study on GCN layer and neighbor selection using EfficientNet_V2 backbone under different seeds. Table 8 shows that the highest mean accuracy of 96.96% occurs with the 3-layer GCN using 4 neighbors at 7 × 7, while the lowest is 94.63% for the 3-layer KNN graph at the same resolution. Increasing the layer to 5 layers generally improves KNN performance, reaching 96.56% at 14 × 14. The 8-neighbour configuration remains consistently strong across both resolutions, achieving up to 96.83%. Overall, these results indicate that both GCN layer and neighbor connectivity influence classification performance, with moderate layer and well-connected graphs providing the most stable accuracy.Table 9Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.960.970.945 Layer0.950.960.95Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.960.960.955 Layer0.950.960.96 Table 9Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.960.970.945 Layer0.950.960.95Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.960.960.955 Layer0.950.960.96 Performance comparison of graph types under different GCN layers and grid resolutions. Table 9 shows that the highest accuracy is achieved by the 3-layer GCN with 8 neighbors at 7 × 7 resolution (0.97), while the lowest occurs for the 3-layer KNN graph at the same resolution (0.94). Increasing the layer to 5 layers improves performance for KNN graphs, reaching 0.96 at 14 × 14, indicating that deeper networks benefit more from KNN connectivity. Overall, the 8-neighbour configuration consistently performs well across both layers and resolutions, while KNN graphs show slight fluctuations. These results highlight that both GCN layer and neighbor selection influence performance, with the highest stability seen in the 8-neighbour graphs. MobileNetV3 Table 10Ablation study on GCN layer and neighbor selection using MobileNetV3 backbone under different seeds.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.4494.7194.58 ± 0.1338-neigh94.1895.2494.71 ± 0.5354-neigh95.2495.7795.50 ± 0.2658-neigh95.2495.7795.50 ± 0.263KNN (k=16)95.5095.2495.37 ± 0.135KNN (k=16)95.7796.0395.90 ± 0.13Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.5094.9795.24 ± 0.2638-neigh94.4495.2494.84 ± 0.4054-neigh95.5096.3095.90 ± 0.4058-neigh95.2495.2495.24 ± 0.003KNN (k=16)95.5096.0395.77 ± 0.265KNN (k=16)95.7795.5095.63 ± 0.13 Table 10Ablation study on GCN layer and neighbor selection using MobileNetV3 backbone under different seeds.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.4494.7194.58 ± 0.1338-neigh94.1895.2494.71 ± 0.5354-neigh95.2495.7795.50 ± 0.2658-neigh95.2495.7795.50 ± 0.263KNN (k=16)95.5095.2495.37 ± 0.135KNN (k=16)95.7796.0395.90 ± 0.13Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.5094.9795.24 ± 0.2638-neigh94.4495.2494.84 ± 0.4054-neigh95.5096.3095.90 ± 0.4058-neigh95.2495.2495.24 ± 0.003KNN (k=16)95.5096.0395.77 ± 0.265KNN (k=16)95.7795.5095.63 ± 0.13 Ablation study on GCN layer and neighbor selection using MobileNetV3 backbone under different seeds. Table 10 shows that the highest mean accuracy of 95.90% is achieved by the 5-layer KNN graph at 7 × 7, while the lowest occurs for the 3-layer 4-neighbour graph at the same resolution (94.58%). The 5-layer configurations generally outperform 3-layer ones, particularly for KNN and 4-neighbour graphs, which reach up to 95.77–95.90%. The 8-neighbour graphs show consistent performance, with mean accuracies ranging from 94.71% to 95.50%, slightly lower than the KNN graphs in most cases. At 14 × 14, the 3-layer KNN graph achieves 95.77%, and the 5-layer 4-neighbour graph reaches 95.90%, indicating that layer contributes more to connected graphs. Overall, both GCN layer and neighbor selection influence classification accuracy, with KNN connectivity and deeper layers providing the most stable and highest performance.Table 11Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.930.940.955 Layer0.950.940.95Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.930.945 Layer0.950.950.95 Table 11Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.930.940.955 Layer0.950.940.95Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.930.945 Layer0.950.950.95 Performance comparison of graph types under different GCN layers and grid resolutions. Table 11 shows that the highest mean accuracy of 0.95 is achieved by both the 3-layer and 5-layer KNN graphs at 7 × 7 as well as 14 × 14 resolutions, while the lowest occurs for the 3-layer 4-neighbour graph at 7 × 7 (0.93) and the 3-layer 8-neighbour graph at 14 × 14 (0.93). Increasing the layer to 5 layers slightly improves 4-neighbour performance from 0.93–0.94 to 0.95. The 8-neighbour graphs remain consistent, with mean accuracies between 0.94–0.95, showing minor fluctuations across resolutions. Overall, KNN graphs provide the most stable and highest performance across all layers and resolutions, while 4- and 8-neighbour configurations show small variations. These results suggest that both layer and graph connectivity influence performance, with KNN connectivity ensuring the most reliable outcomes. ConvNeXt Table 12Ablation study on GCN layer and neighbor selection using ConvNeXt backbone under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.9796.3095.63 ± 0.6638-neigh94.5895.1694.87 ± 0.4154-neigh95.7796.8396.30 ± 0.5358-neigh96.0396.5696.30 ± 0.263KNN (k=16)92.5690.5492.05 ± 0.725KNN (k=16)91.2391.1091.17 ± 0.09Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.7894.1294.95 ± 1.1738-neigh96.5695.7796.16 ± 0.4054-neigh96.0396.0396.03 ± 0.0058-neigh96.3095.2495.77 ± 0.533KNN (k=16)89.8888.7889.33 ± 0.785KNN (k=16)92.4593.1592.80 ± 0.50 Table 12Ablation study on GCN layer and neighbor selection using ConvNeXt backbone under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.9796.3095.63 ± 0.6638-neigh94.5895.1694.87 ± 0.4154-neigh95.7796.8396.30 ± 0.5358-neigh96.0396.5696.30 ± 0.263KNN (k=16)92.5690.5492.05 ± 0.725KNN (k=16)91.2391.1091.17 ± 0.09Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.7894.1294.95 ± 1.1738-neigh96.5695.7796.16 ± 0.4054-neigh96.0396.0396.03 ± 0.0058-neigh96.3095.2495.77 ± 0.533KNN (k=16)89.8888.7889.33 ± 0.785KNN (k=16)92.4593.1592.80 ± 0.50 Ablation study on GCN layer and neighbor selection using ConvNeXt backbone under different node resolutions. Table 12 shows that the highest mean accuracy of 96.30% is achieved by the 5-layer models with both 4- and 8-neighbour graphs at 7 × 7, while the lowest value of 89.33% occurs for the 3-layer KNN graph at 14 × 14. At 7 × 7, 4- and 8-neighbour configurations perform strongly (94.87–96.30%), whereas KNN graphs drop to 91.17–92.05%. Increasing the layer improves performance for structured graphs, such as 4-neighbour (95.63% to 96.30%), but KNN remains consistently lower across both resolutions. Overall, 8-neighbour and 4-neighbour graphs outperform KNN, especially at higher resolution, indicating that structured connectivity is more effective than KNN for ConvNeXt features.Table 13Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.940.925 Layer0.930.940.91Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.950.895 Layer0.960.960.93 Table 13Performance comparison of graph types under different GCN layers and grid resolutions.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.940.925 Layer0.930.940.91Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.950.895 Layer0.960.960.93 Performance comparison of graph types under different GCN layers and grid resolutions. Table 13 shows that the highest accuracy of 0.96 is achieved by the 5-layer model with both 4- and 8-neighbour graphs at 14 × 14, while the lowest value of 0.89 occurs for the 3-layer KNN graph at the same resolution. At 7 × 7, the 3-layer 4-neighbour configuration performs best with 0.95, whereas KNN graphs show lower performance, dropping to 0.91–0.92. Increasing layer from 3 to 5 layers improves results at 14 × 14 (e.g., 0.94 to 0.96 for 4-neighbour), but slightly reduces performance at 7 × 7. Overall, 4- and 8-neighbour graphs consistently outperform KNN, especially at higher resolution, indicating that structured connectivity is more effective than KNN in this setting. Swin Table 14Ablation study on GCN layer and neighbor selection using Swin backbone under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.0396.3096.16 ± 0.1338-neigh95.5095.2495.37 ± 0.1354-neigh94.7193.3994.05 ± 0.6658-neigh95.5095.5095.50 ± 0.003KNN (k=16)80.4276.9878.70 ± 1.725KNN (k=16)82.4580.1481.30 ± 1.63Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.2495.5095.37 ± 0.1338-neigh94.4495.7795.11 ± 0.6654-neigh95.5094.9795.24 ± 0.2658-neigh96.3095.2495.77 ± 0.533KNN (k=16)73.5481.4877.51 ± 3.975KNN (k=16)79.1080.9580.03 ± 0.93 Table 14Ablation study on GCN layer and neighbor selection using Swin backbone under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh96.0396.3096.16 ± 0.1338-neigh95.5095.2495.37 ± 0.1354-neigh94.7193.3994.05 ± 0.6658-neigh95.5095.5095.50 ± 0.003KNN (k=16)80.4276.9878.70 ± 1.725KNN (k=16)82.4580.1481.30 ± 1.63Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh95.2495.5095.37 ± 0.1338-neigh94.4495.7795.11 ± 0.6654-neigh95.5094.9795.24 ± 0.2658-neigh96.3095.2495.77 ± 0.533KNN (k=16)73.5481.4877.51 ± 3.975KNN (k=16)79.1080.9580.03 ± 0.93 Ablation study on GCN layer and neighbor selection using Swin backbone under different node resolutions. Table 14 shows that the highest mean accuracy of 96.16% is achieved by the 3-layer 4-neighbour graph at 7 × 7, while the lowest value of 77.51% occurs for the 3-layer KNN graph at 14 × 14. At 7 × 7, structured graphs (4- and 8-neighbour) perform strongly (95.37–96.16%), whereas KNN graphs drop significantly to 78.70–81.30%. At 14 × 14, performance remains stable for 4- and 8-neighbour configurations (95.11–95.77%), but KNN again shows much lower accuracy (77.51–80.03%). Overall, structured neighbor connections consistently outperform KNN across both resolutions, indicating that Swin features are less effective with KNN-based graph construction.Table 15Performance comparison of graph types under different GCN layers using Swin backbone.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.950.805 Layer0.950.930.82Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.930.745 Layer0.960.970.81 Table 15Performance comparison of graph types under different GCN layers using Swin backbone.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.950.805 Layer0.950.930.82Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.950.930.745 Layer0.960.970.81 Performance comparison of graph types under different GCN layers using Swin backbone. Table 15 shows that the highest accuracy of 0.97 is achieved by the 5-layer 8-neighbour graph at 14 × 14, while the lowest value of 0.74 occurs for the 3-layer KNN graph at the same resolution. At 7 × 7, 4- and 8-neighbour graphs perform consistently well (0.93–0.95), whereas KNN graphs drop to 0.80–0.82. At 14 × 14, increasing layer improves performance for structured graphs, with 4-neighbour rising from 0.95 to 0.96 and 8-neighbour from 0.93 to 0.97. Overall, KNN graphs remain significantly lower across both resolutions, indicating that structured neighbor connections are more effective for Swin-based representations. ViT Table 16Ablation study on GCN layer and neighbor selection using ViT backbone under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.7194.1894.44 ± 0.2638-neigh94.9794.9794.97 ± 0.0054-neigh93.1294.1893.65 ± 0.5358-neigh93.9293.9293.92 ± 0.003KNN (k=16)94.1894.9794.58 ± 0.405KNN (k=16)95.7595.5095.62 ± 0.17Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh92.5992.8692.72 ± 0.1338-neigh92.5992.3392.46 ± 0.1354-neigh94.1894.4494.31 ± 0.1358-neigh94.7194.7194.71 ± 0.003KNN (k=16)92.8693.1292.99 ± 0.135KNN (k=16)94.1894.9794.58 ± 0.40 Table 16Ablation study on GCN layer and neighbor selection using ViT backbone under different node resolutions.Node Resolution: 7 × 7LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh94.7194.1894.44 ± 0.2638-neigh94.9794.9794.97 ± 0.0054-neigh93.1294.1893.65 ± 0.5358-neigh93.9293.9293.92 ± 0.003KNN (k=16)94.1894.9794.58 ± 0.405KNN (k=16)95.7595.5095.62 ± 0.17Node Resolution: 14 × 14LayerNeighborSeed1 (42)Seed2 (1000)Mean ± Std34-neigh92.5992.8692.72 ± 0.1338-neigh92.5992.3392.46 ± 0.1354-neigh94.1894.4494.31 ± 0.1358-neigh94.7194.7194.71 ± 0.003KNN (k=16)92.8693.1292.99 ± 0.135KNN (k=16)94.1894.9794.58 ± 0.40 Ablation study on GCN layer and neighbor selection using ViT backbone under different node resolutions. Table 16 shows that the highest mean accuracy of 95.62% is achieved by the 5-layer KNN graph at 7 × 7, while the lowest value of 92.46% occurs for the 3-layer 8-neighbour graph at 14 × 14. At 7 × 7, most configurations perform closely (93.65–95.62%), with KNN slightly outperforming structured graphs in deeper settings. In contrast, at 14 × 14, increasing layer significantly improves performance, with 5-layer 8-neighbour reaching 94.71% compared to 92.46% for its 3-layer counterpart. Overall, both layer and graph type influence performance, with deeper models consistently providing better results across both resolutions.Table 17Performance comparison of graph types under different GCN layers using ViT backbone.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.950.935 Layer0.950.940.94Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.920.935 Layer0.940.950.93 Table 17Performance comparison of graph types under different GCN layers using ViT backbone.Node Resolution: 7 × 7Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.950.935 Layer0.950.940.94Node Resolution: 14 × 14Layer4-neighbour8-neighbourKNN (k=16)3 Layer0.940.920.935 Layer0.940.950.93 Performance comparison of graph types under different GCN layers using ViT backbone. Table 17 shows that the highest accuracy of 0.95 is achieved by the 3-layer 8-neighbour graph at 7 × 7 and the 5-layer 8-neighbour graph at 14 × 14, while the lowest value of 0.92 occurs for the 3-layer 8-neighbour graph at 14 × 14. At 7 × 7, all configurations perform closely (0.93–0.95), with slight improvements for 5-layer models in 4-neighbour and KNN graphs. At 14 × 14, increasing layer improves the 8-neighbour graph from 0.92 to 0.95, while 4-neighbour and KNN remain relatively stable around 0.93–0.94. Overall, the results indicate that ViT features are less sensitive to graph construction, with only minor variations across layer and neighbor types. Loss curve analysis Fig. 4Loss curve. Loss curve. Figure 4 shows the training and validation loss curves for the seven baseline backbone models. In all cases, the training loss decreases rapidly during the first few epochs and then gradually stabilizes, indicating that the models capture the dominant data patterns early in training and fit the training set effectively. The validation loss curves also exhibit an overall decreasing trend, suggesting reasonable generalization to unseen samples. However, differences emerge in stability. ConvNeXt and ViT-S/16 display relatively larger fluctuations after convergence, indicating mild overfitting despite achieving accuracies of 0.94 and 0.93, respectively (Table 3). In contrast, EfficientNetV2-S, DenseNet201, and MobileNetV3 show smoother and more stable validation curves, reflecting stronger generalization; this is consistent with their competitive accuracies of 0.96, 0.93, and 0.96. ResNet50 demonstrates a balanced behavior, with steady convergence and moderate fluctuation in validation loss, aligning with its solid accuracy of 0.94, indicating reliable but slightly less stable generalization compared to the best-performing models. he Swin Transformer shows a consistent reduction in training loss along with a relatively stable validation loss trend, indicating good convergence behavior and controlled learning dynamics. Although minor fluctuations are present. However, despite this stable training pattern, its performance remains comparatively lower, with an accuracy of 0.84 (Table 3), indicating that good loss convergence does not necessarily translate to superior classification performance in this case. Overall, while all models converge successfully, the degree of validation stability clearly differentiates their performance. Architectures such as MobileNetV3 and EfficientNetV2-S, both achieving 0.96 accuracy, combine smooth convergence with strong predictive performance, making them more robust choices. In contrast, models with higher fluctuations, such as ConvNeXt and ViT-S/16, show signs of overfitting, while ResNet50 occupies a middle ground with consistent yet slightly less stable generalization. This comparative behavior highlights the importance of validation stability as an indicator of model reliability beyond accuracy alone.Fig. 5Accuracy curve. Accuracy curve. Figure 5 illustrates the training and validation accuracy curves for the seven baseline backbone models. In all cases, the training accuracy increases rapidly during the initial epochs and then gradually levels off, indicating that the models are able to quickly learn the key patterns present in the dataset. The validation accuracy curves generally follow a similar upward trend, suggesting that the learned representations transfer well to unseen data. However, differences become apparent in terms of stability. ConvNeXt and ViT-S/16 exhibit more noticeable fluctuations in validation accuracy after convergence, pointing to mild overfitting and less consistent generalization, even though they achieve respectable final accuracies of 0.94 and 0.93, respectively (Table 3). In contrast, EfficientNetV2-S, DenseNet201, and MobileNetV3 present smoother and more stable validation curves, indicating more reliable learning behavior; this is reflected in their strong accuracies of 0.96, 0.93, and 0.96. ResNet50 shows a well-balanced trend, with steady convergence and moderate variation in validation accuracy, aligning with its solid performance of 0.94. The Swin Transformer, on the other hand, demonstrates a steady and well-aligned progression between training and validation accuracy, suggesting stable learning dynamics. However, despite this visually consistent behavior, its overall performance remains lower, with an accuracy of 0.84 (Table 3), implying that stable curves alone do not guarantee superior predictive capability. Overall, while all models successfully learn the task, the stability of the validation accuracy plays a key role in distinguishing their effectiveness. EfficientNetV2-S and MobileNetV3 stand out by combining smooth and closely aligned training-validation curves with the highest accuracy of 0.96, making them the most reliable performers. In comparison, ConvNeXt and ViT-S/16 show greater variability, indicating sensitivity during training and a tendency toward overfitting. ResNet50 lies in between, offering consistent and dependable performance without being the most stable. These observations highlight that beyond final accuracy values, the smoothness and alignment of accuracy curves provide important insights into model robustness and generalization ability. t-SNE analysis Fig. 6t-SNE. t-SNE. The t-SNE visualizations in Fig. 6 shows the feature embeddings of seven hybrid models, highlighting how each variant organizes the data in the latent space. Most of the models produce reasonably well-formed clusters, indicating that the hybrid architecture is effective in learning discriminative features. In several cases, the clusters are compact and clearly separated, reflecting strong intra-class consistency and reduced overlap between classes. However, a few models still show slight dispersion or minor overlaps, suggesting differences in how effectively each configuration captures relationships among samples. Overall, the distributions are structured and distinguishable across all variants, demonstrating the robustness of the hybrid design. Confusion matrix Fig. 7Confusion Matrix of GCN + EfficientNet (3L, 8 N, 7×7). Confusion Matrix of GCN + EfficientNet (3L, 8 N, 7×7). The confusion matrix is shown in Fig. 7 that the GCN + EfficientNet model performs very strongly, with most predictions concentrated along the diagonal, indicating high classification accuracy across flower types and growth stages. Classes such as Bush Allamanda (Early/Mid), Red Hibiscus (Early), and White Hibiscus (Early) achieve near-perfect predictions with minimal or no misclassification. A few minor confusions appear between closely related growth stages, for example within White Hibiscus and Madagascar Periwinkle, suggesting the model sometimes struggles to distinguish subtle stage differences. Similarly, slight misclassifications are visible in classes like Yellow Bell and Pinwheel Flower, mostly between adjacent stages (e.g., Early vs Mid). Overall, the model demonstrates high robustness and effective feature extraction for multi-class, multi-stage flower classification (Fig. 8). Explainable AI To further interpret the behavior of the trained CNN–GCN hybrid model, Ablation-CAM was applied to representative samples from different flower categories and growth stages. The generated heatmaps highlight the regions that have the most significant impact on the model’s predictions by measuring the effect of removing feature activations. As shown in Fig. 9, the model consistently focuses on semantically meaningful regions such as petals, buds, and blooming structures, rather than background elements. In early-stage samples, the model primarily attends to compact bud regions, indicating sensitivity to initial growth structures. For mid-stage flowers, attention gradually expands toward partially opened petals, capturing transitional morphological features. In fully bloomed samples, the activation spreads across the entire petal region, reflecting the model’s ability to capture complex textures and color variations. Compared to traditional attention visualization, Ablation-CAM produces more localized and reliable importance regions, suggesting that the model relies on discriminative and biologically relevant features. This confirms that the hybrid CNN–GCN framework effectively captures both visual and structural characteristics necessary for fine-grained growth-stage classification.Fig. 8Grad cam of sample images. Grad cam of sample images. Failure case analysis To further analyze the limitations of the proposed framework, a failure case example is presented in Fig. 9. In this example, an image from the Bush Allamanda Early Stage class is incorrectly predicted as Bush Allamanda Mid Stage. The corresponding Ablation-CAM visualization shows that the model attention is not fully concentrated on the flower region and partially focuses on surrounding background areas. This misclassification may occur due to factors such as illumination variation, image quality degradation, and high visual similarity between adjacent growth stages. In particular, subtle differences between early and mid growth stages can make class separation difficult when the discriminative regions are not properly highlighted by the model. Although the proposed model achieves strong overall performance, this failure case demonstrates that incorrect attention localization can still affect classification accuracy in visually challenging samples. Such analysis provides further insight into the strengths and limitations of the proposed hybrid framework.Fig. 9Failure case. Failure case. Complexity analysis Table 18Complexity analysis of hybrid models in terms of parameters and FLOPs.ModelParam (M)FLOPs (G)GCN + ResNet5023.564.13GCN + DenseNet20118.144.39GCN + ConvNeXt87.5715.37GCN + EfficientNet20.212.9GCN + MobileNet2.990.23GCN + ViT57.3211.28GCN + Swin18.962.99 Table 18Complexity analysis of hybrid models in terms of parameters and FLOPs.ModelParam (M)FLOPs (G)GCN + ResNet5023.564.13GCN + DenseNet20118.144.39GCN + ConvNeXt87.5715.37GCN + EfficientNet20.212.9GCN + MobileNet2.990.23GCN + ViT57.3211.28GCN + Swin18.962.99 Complexity analysis of hybrid models in terms of parameters and FLOPs. Table 18 presents the computational complexity of different GCN-backbone models in terms of the number of parameters and floating point operations (FLOPs). It can be observed that lightweight models like GCN + MobileNet have very low parameters and FLOPs, making them suitable for resource-constrained environments, whereas heavier models such as GCN + ConvNeXt and GCN + ViT require significantly more computation. This comparison highlights the trade-off between model complexity and potential accuracy performance. Discussion This study addresses the underexplored problem of fine-grained flower growth-stage classification across multiple species using RGB images. Unlike prior works that focus on species identification, disease detection, or rely on hyperspectral and multi-omics data, the proposed approach enables practical and cost-effective stage recognition suitable for real-world agricultural applications. To support this task, we introduce the BD Flower Growth dataset of eight species across three developmental stages. Methodologically, we propose a CNN/Transformer + GCN hybrid framework that enhances conventional appearance-based feature learning by modeling structural and relational dependencies among flower parts. By converting spatial features into graph nodes and applying multi-layer GCNs, the model captures subtle morphological transitions between growth stages. Experiments with seven strong backbones demonstrate the robustness of the approach. CNN-based hybrids consistently outperform transformer-based variants, indicating the continued effectiveness of convolutional inductive biases for fine-grained biological datasets. A comparison with existing studies is provided in Table 19.Table 19Comparison of the proposed framework with recent related studies.Ref.DatasetModelResult35Strawberry Growth Stage DatasetEfficientNetB783.7% accuracy36Early-stage Apple FlowersYOLOvPrecision: 73.5%, Recall: 83.5%37Chrysanthemum Flower Growth DatasetFusion-YOLOAverage Precision: 68.87%38Wheat Phenological Stage ClassificationEMS-GCN86.2% accuracy39Apple Flower and Flower BudsMask R-CNNPrecision: 78.2%40Sunflower Growth Stage DatasetDual-Branch YOLOv10nPrecision: 90.0%, Recall: 89.4%41RoseBlooming datasetYOLOv5sF1-score = 0.9503BDFlower Growth Stage (Our Dataset)GCN + EfficientNetV2-S97.0% accuracy3BDFlower Growth Stage (Our Dataset)GCN + Swin Transformer97.0% accuracy Table 19Comparison of the proposed framework with recent related studies.Ref.DatasetModelResult35Strawberry Growth Stage DatasetEfficientNetB783.7% accuracy36Early-stage Apple FlowersYOLOvPrecision: 73.5%, Recall: 83.5%37Chrysanthemum Flower Growth DatasetFusion-YOLOAverage Precision: 68.87%38Wheat Phenological Stage ClassificationEMS-GCN86.2% accuracy39Apple Flower and Flower BudsMask R-CNNPrecision: 78.2%40Sunflower Growth Stage DatasetDual-Branch YOLOv10nPrecision: 90.0%, Recall: 89.4%41RoseBlooming datasetYOLOv5sF1-score = 0.9503BDFlower Growth Stage (Our Dataset)GCN + EfficientNetV2-S97.0% accuracy3BDFlower Growth Stage (Our Dataset)GCN + Swin Transformer97.0% accuracy Comparison of the proposed framework with recent related studies. Comaprison between baseline and hybrid models Table 20 shows that the GCN-based hybrid models consistently improve performance over their baseline counterparts across most architectures. For example, DenseNet201 improves from 0.93 accuracy and 0.94 F1-score to 0.97 for both metrics when integrated with GCN, while ConvNeXt_Base increases from 0.91 to 0.96 in F1-score. GCN + DenseNet201, GCN + EfficientNetV2-S and GCN + Swin achieved the highest overall accuracy of 0.97. However, the Swin backbone showed the largest relative improvement, increasing from 0.84 to 0.97 after integration with the GCN module. Here, in the Table 20, L denotes the number of GCN layers, N represents the number of neighbors in the graph (“4N” and “8N” denote 4-neighbour and 8-neighbour graph connectivity), and node resolutions such as 7× 7 and 14× 14 indicate the spatial granularity used for graph construction.Table 20Performance comparison between baseline and GCN-based hybrid models.ModelAccPrecRecF1ResNet500.940.950.930.93GCN + ResNet50 (3L, 8 N, 7×7)0.950.960.930.94DenseNet2010.930.950.940.94GCN + DenseNet201 (5L, 8 N, 7×7)0.970.960.970.97EfficientNet_V2_S0.960.970.960.97GCN + EfficientNet_V2_S (3L, 8 N, 7×7)0.970.980.970.97MobileNetV30.960.970.930.94GCN + MobileNetV3 (5L, 4 N, 7×7)0.950.950.950.95ConvNeXt_Base0.940.950.900.91GCN + ConvNeXt_Base (5L, 8 N, 14×14)0.960.970.950.96ViT_B_160.930.960.920.94GCN + ViT_B_16 (3L, 8 N, 7×7)0.950.960.950.95Swin0.840.870.820.82GCN + Swin (5L, 8 N, 14×14)0.970.980.950.96 Table 20Performance comparison between baseline and GCN-based hybrid models.ModelAccPrecRecF1ResNet500.940.950.930.93GCN + ResNet50 (3L, 8 N, 7×7)0.950.960.930.94DenseNet2010.930.950.940.94GCN + DenseNet201 (5L, 8 N, 7×7)0.970.960.970.97EfficientNet_V2_S0.960.970.960.97GCN + EfficientNet_V2_S (3L, 8 N, 7×7)0.970.980.970.97MobileNetV30.960.970.930.94GCN + MobileNetV3 (5L, 4 N, 7×7)0.950.950.950.95ConvNeXt_Base0.940.950.900.91GCN + ConvNeXt_Base (5L, 8 N, 14×14)0.960.970.950.96ViT_B_160.930.960.920.94GCN + ViT_B_16 (3L, 8 N, 7×7)0.950.960.950.95Swin0.840.870.820.82GCN + Swin (5L, 8 N, 14×14)0.970.980.950.96 Performance comparison between baseline and GCN-based hybrid models. Although multiple hybrid configurations achieved an accuracy of 0.97, their relative improvements over the baseline models differed, with the Swin backbone showing the largest gain. Table 19 compares the proposed method with recent competitive approaches, including graph-based and transformer-related models such as EMS-GCN and Dual-Branch YOLOv10n. In addition, Table 20 presents direct comparisons between the proposed hybrid framework and strong transformer baselines, including ViT-B/16 and Swin Transformer. The proposed GCN-enhanced models generally improve performance over baseline models, although the degree of improvement varies across different backbones, demonstrating that the integration of graph-based relational reasoning improves classification performance over both conventional CNN and transformer architectures. Cross-domain validation on external flower datasets Table 21 presents the performance of the best GCN configurations selected from the ablation study, applied across three datasets: two Sunflower Growth datasets4243, and the publicly available Oxford 10244 dataset. These datasets differ in image characteristics, class distribution, and visual complexity, making them suitable for cross-domain validation. As shown in Table 21, the proposed hybrid framework maintains strong performance across different external datasets, with results varying depending on the backbone model and dataset characteristics. The results show that these configurations, typically deeper GCNs with sufficient neighborhood connections, achieve high accuracy, precision, recall, and F1 scores, across different datasets, although the exact values vary depending on the backbone model and dataset characteristics.Table 21Performance comparison of the GCN-based proposed framework across different external datasets.ModelLayerNodeNeigh.DatasetAccPrecRecF1GCN + ResNet5037× 78420.980.990.980.98430.970.970.970.97440.950.950.940.94GCN + DenseNet57× 78421.001.001.001.00430.990.990.990.99440.970.970.960.96GCN + MobileNet57× 74420.990.990.990.99430.980.980.980.98440.930.930.920.92GCN + EfficientNet37× 78421.001.001.001.00430.980.980.970.98440.980.980.980.98GCN + ConvNeXt514× 148421.001.001.001.00430.980.980.980.98440.960.970.950.96GCN + ViT37× 78420.980.970.980.98430.960.970.970.96440.960.960.950.96GCN + Swin514× 148420.980.980.990.99430.970.970.970.97440.980.980.970.98 Table 21Performance comparison of the GCN-based proposed framework across different external datasets.ModelLayerNodeNeigh.DatasetAccPrecRecF1GCN + ResNet5037× 78420.980.990.980.98430.970.970.970.97440.950.950.940.94GCN + DenseNet57× 78421.001.001.001.00430.990.990.990.99440.970.970.960.96GCN + MobileNet57× 74420.990.990.990.99430.980.980.980.98440.930.930.920.92GCN + EfficientNet37× 78421.001.001.001.00430.980.980.970.98440.980.980.980.98GCN + ConvNeXt514× 148421.001.001.001.00430.980.980.980.98440.960.970.950.96GCN + ViT37× 78420.980.970.980.98430.960.970.970.96440.960.960.950.96GCN + Swin514× 148420.980.980.990.99430.970.970.970.97440.980.980.970.98 Performance comparison of the GCN-based proposed framework across different external datasets. Limitation and future work The dataset was collected from a single location, specifically a balcony environment, under relatively controlled conditions. As a result, variations in background complexity, lighting conditions, weather, and plant diversity are limited compared to real-world agricultural or field environments. Despite the strong performance of the proposed CNN–GCN hybrid framework, several methodological limitations remain. The current graph construction is based on a fixed spatial grid derived from CNN feature maps, which preserves local structure but limits the model’s ability to capture long-range or semantically adaptive relationships between distant regions. In addition, the use of standard GCN layers introduces the risk of oversmoothing when deeper graph stacks are applied, which can reduce feature diversity and weaken class separability. The feature fusion strategy also relies on simple concatenation, which does not explicitly model interactions between convolutional and graph-based representations. Future work will focus on extending the dataset by collecting images from multiple locations and environmental conditions, including outdoor field settings. Additionally, future work will aim to improve the flexibility and expressiveness of the graph learning pipeline. Adaptive or attention-based graph construction could allow the network to learn task-specific spatial relationships instead of relying on predefined neighborhoods. Incorporating advanced graph architectures, such as graph attention or residual graph connections, may help mitigate oversmoothing and enable deeper relational reasoning. More sophisticated fusion strategies, including gated or attention-based integration, could also be explored to better combine visual and structural features and further enhance fine-grained growth-stage classification performance. The reported results are presented as mean ± standard deviation over two runs with different random seeds. Formal statistical significance testing was not performed because the limited number of repeated runs does not provide sufficient statistical power for reliable hypothesis testing. More extensive statistical validation with additional runs will be considered in future work. Conclusion This paper presents a hybrid CNN/Transformer–GCN framework for fine-grained flower growth-stage classification and introduces the BD Flower Growth dataset with 3,889 original images. The proposed method integrates deep feature extraction with graph-based refinement by constructing graphs from feature maps and applying Graph Convolutional Networks. A systematic ablation study evaluates the effects of GCN depth (3, 5 layers), node resolution (7×7, 14×14), and graph construction (4-neighbour, 8-neighbour, KNN with k=16). The results show that performance depends on graph design, where structured grid-based graphs provide more stable outcomes. The best performance of 97% is achieved by multiple hybrid configurations, including EfficientNetV2-S, DenseNet201, and Swin-based models. Cross-dataset evaluation on the Oxford 102 Flower dataset further demonstrates the generalization capability of the proposed framework. Overall, the study highlights the importance of graph configuration in improving fine-grained visual classification for plant phenotyping applications.