Comparison Graph Neural Networks for Remote Sensing Scene Classification

Remote sensing scene classification has been a research hotspot in recent years, and convolution neural networks have been widely used in this field. However, remote sensing scene images have large-scale variance with regard to a specific category, making it still difficult to individually train and obtain an excellent classifier by adopting features of single sample extracted by a CNN model to make prediction. To tackle above issue, a novel framework named Comparison Graph Neural Networks (CGNN) is proposed for remote sensing scene classification. The framework constructs comparison graph based on sample features extracted by CNN model. Then CGNN is employed on the graph for sample comparison and aggregates node features according to node connection weights learned by metric-learning neural networks from node similarities. Experiments are conducted on the benchmark dataset and the proposed framework obtains competitive performance compared with powerful baselines.