DC-GCN: A Lightweight Graph Convolutional Network Integrating Local and Global Context for Semantic Segmentation

Graph convolutional networks (GCNs) have demonstrated remarkable performance in hyperspectral image classification and remote sensing scene recognition. However, their application to semantic segmentation remains relatively underexplored. Traditional convolutional neural networks (CNNs), constrained by fixed-grid kernels and inherently limited receptive fields, struggle to capture long-range dependencies critical for fine-grained segmentation. In contrast, GCNs operate on irregular, non-Euclidean domains, enabling explicit modeling of global contextual relationships among adjacent land cover types. Leveraging outputs from geographic object-based image analysis (GEOBIA), raw remote sensing imagery can be transformed into structured graphs, facilitating more effective interpretation of complex scenes. Here, we propose a lightweight dual-channel GCN framework (DCGCN) that jointly models local and global contextual features within a unified inductive learning paradigm. The architecture supports efficient parallel training and generalizes seamlessly to unseen data without retraining. Specifically, local context is extracted from the original graph topology, while global dependencies are captured via a reconstructed graph representation. These complementary features are fused through a gated attention mechanism (GAM), enhancing contextual integration. To address the oversmoothing issue prevalent in deeper GCNs, we incorporate a densely connected residual module that preserves feature discriminability and promotes training stability. Experimental results on three semantic segmentation datasets show that DCGCN-GATn (where n denotes the number of attention heads) delivers superior performance, achieving an optimal balance between accuracy metrics (precision, overall accuracy, and kappa) and efficiency indicators (FPS, FLOPs, and parameter size).