Remote sensing scene classification with relation-aware dynamic graph neural networks

Remote sensing scene classification (RSSC) is challenging due to the complexity and diversity of scenes. Existing methods need help to capture long-range and structural relationships among image regions, limiting their performance. This paper proposes a novel graph-based model that learns relation-aware dynamic graph representations for remote sensing scene classification tasks. The proposed model consists of three main components: Multi-Scale Feature Extraction (MSFE), Relation-Aware Graph Processing (RAGP), and Scene Classification with Weighted Pooling (SCWP). MSFE uses a multi-scale feature extraction strategy to generate low-level feature nodes from remote sensing images. RAGP applies several cascaded graph processing blocks to dynamically learn the relations between nodes in high-level semantic spaces using relation-aware graph convolutional and node feature update operators. SCWP performs weighted pooling on the learned node features from RAGP to obtain global representations of remote sensing images and makes scene decisions using a fully feed-forward network-based classifier. We evaluate our model on three benchmark datasets and compare it with state-of-the-art RSSC methods. Our experimental results show that our model outperforms existing methods on all three datasets, demonstrating the effectiveness of a graph-based model with the proposed techniques for RSSC tasks.