Hybrid CNN-GNN architectures with distributed training for heathland plant classification.

Accurate predictions of heathland plant species are crucial for ecological monitoring and assessing biodiversity. Previous research has predominantly utilised convolutional neural networks (CNNs), which process images arranged on a regular grid. Although CNNs are effective at extracting local visual features, they frequently do not capture the irregular spatial relationships characteristic of heathland vegetation. Earlier approaches employing graph neural networks (GNNs) for plant classification have generally relied on simple or dataset-wide graph constructions, with limited consideration for edges that represent the actual morphological structure of plants. This study presents PlantGraphNet, a hybrid CNN-GNN framework that integrates visual and structural information for heathland plant species classification. PlantGraphNet extracts image-specific keypoints and local descriptors to construct graphs, where nodes correspond to plant regions and edges encode spatial relationships. The resulting graphs are processed using graph convolution and graph attention layers to capture relational context, while a CNN backbone provides complementary appearance features. These two modalities are combined into a unified representation for classification. To ensure scalability, PlantGraphNet employs distributed data-parallel training, enabling efficient gradient synchronisation across devices and near-linear performance scaling. When evaluated on Danish aerial heathland datasets, PlantGraphNet achieves a precision of 98.98%, substantially outperforming CNN-only baselines. In addition to improved accuracy, the explicit graph construction increases interpretability by associating classification outcomes with specific plant structures. These findings indicate that integrating CNN-derived features with purposefully designed graph representations of plant morphology yields a robust and interpretable approach for fine-grained heathland plant species classification.