EFF-GAT: an expanded feature fusion graph attention network for side-scan sonar image classification

Side-scan sonar is a crucial remote sensing technique for marine exploration. However, low resolution, limited information, and noise interference in sonar images hinder fine-grained classification. Existing deep learning-based methods for side-scan sonar image classification typically employ convolutional neural networks to extract local features using various convolutional kernels. These methods predominantly focus on local features while neglecting global feature modelling, limiting their ability to capture global spatial relationships in sonar images. To address these issues, this paper proposes a side-scan sonar image classification method based on an expanded feature fusion graph attention network, named EFF-GAT. First, multi-dimensional features such as pixel, texture, shape, and spatial position are extracted to construct an expanded feature set, with correlation analysis used to reduce redundancy and enhance the discriminative power of the features. Second, a nonlinear mapping and weighted fusion strategy integrates the expanded features, thereby enhancing their expressive ability. Subsequently, the KNN algorithm is used to model both local and global features, enabling feature aggregation and propagation. Finally, a graph attention mechanism is introduced to dynamically adjust the weights between nodes and their neighbours, optimizing the representation of node features. Experimental results show that the proposed method outper- forms existing state-of-the-art CNN- and GNN-based classification methods on real side-scan sonar datasets.