Integrated Deep Learning Segmentation and Graph Convolutional Networks with Relaxation Labeling for Object Relationship Extraction in Satellite Imagery

Analyzing spatial and semantic relationships in satellite imagery is essential for applications ranging from urban planning to disaster management. This paper presents a novel approach that combines semantic segmentation and graph convolutional networks (GCNs) to model these relationships in high-resolution satellite images. We employ state-of-the-art segmentation models, including U-Net, ResUNet, and SegNet, to accurately delineate object boundaries for detailed analysis. After segmentation, we extract object coordinates to build a graph representation, where each object is represented as node with features based on the coordinates. Our method leverages a Custom Graph Convolutional Network(GCN) to evaluate relationships between nodes. The GCN uses an iterative message passing process, where the first iteration classifies spatial and semantic relationships based on the spatial distances between the nodes, while the second iteration refines these relationships by applying a relaxation labeling technique to eliminate incorrect edges. The proposed framework effectively integrates precise segmentation with sophisticated graph-based analysis, offering significant improvements in understanding spatial and semantic interactions within satellite images. This methodology is demonstrated through its applicability in urban planning, environmental monitoring, and disaster response, where visualizing relationships as graphs enhances decision-making. Comprehensive evaluations against existing methods confirm the framework’s capability in accurately capturing and interpreting complex spatial relationships.