SAGRNet: A novel object-based graph convolutional neural network for diverse vegetation cover classification in remotely-sensed imagery

• We propose a novel SAGRNet model for efficiently classifying diverse vegetation cover in remote sensing images. • This classification works through fusion of sampling aggregation and multi-attention mechanisms. • The SAGRNet model surpasses other mainstream pixel-based and object-based deep learning algorithms. • We offer a analysis of the SAGRNet model’s sensitivity across different environments, parameter , and model architectures. • This model has the advantages of high accuracy, adaptability in different vegetation landscapes, and high efficiency. Growing global population, changing climate, and shrinking land resources demand for quicker, efficient, and more accurate methods of mapping and monitoring vegetation cover in remote sensing datasets. Many deep learning-based methods have been widely applied for semantic segmentation tasks in remote sensing images of vegetated environments. However, most existing models are pixel-based, which introduces challenges such as high time consumption, cumbersome implementation, and limited scalability. This paper presents the SAGRNet model, a Graph Convolutional Neural Network (GCN) that incorporates sampling aggregation and self-attention mechanisms, while leveraging the ResNet residual network structure. A key innovation of SAGRNet is its ability to fuse features extracted through diverse algorithms, enabling comprehensive representation and enhanced classification performance. The SAGRNet model demonstrates superior performance over leading pixel-based neural networks, such as U-Net++ and DeepLabV3, in terms of both time efficiency and accuracy in vegetation image classification tasks. We achieved an overall mapping accuracy of ∼90 % using SAGRNet, compared to ∼87% and ∼85% from U-Net++ and DeepLabV3, respectively. Additionally, it offers more convenience in data processing. Furthermore, the model significantly outperforms cutting-edge graph-based convolutional networks, including Graph U-Net (achieved overall accuracy ∼65%) and TGNN (achieved overall accuracy ∼75%), showcasing exceptional generalization capability and classification accuracy. This paper provides a comprehensive analysis of the various processing aspects of this object-based GCN for vegetation mapping and emphasizes its significant potential for practical use. The model’s versatility can also be expanded to other image processing domains, offering unprecedented possibilities of information extraction from satellite imagery. The code for practical application experiment is available at https://github.com/baoling123/GCN-remote-sensing-classification.git .