D3GNN: Double dual dynamic graph neural network for multisource remote sensing data classification
Convolutional Neural Network (CNN) has garnered attention due to its outstanding performance in multisource remote sensing (RS) image classification. However, classical CNN-based methods primarily concentrate on information within a fixed-size neighborhood and a standard square region, neglecting lo...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-05-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225001438 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Convolutional Neural Network (CNN) has garnered attention due to its outstanding performance in multisource remote sensing (RS) image classification. However, classical CNN-based methods primarily concentrate on information within a fixed-size neighborhood and a standard square region, neglecting long-range and global information. As non-Euclidean data, the topological structure enables flexible construction of relationships between objects, which can be served as an effective carrier of global information. Graph Neural Network (GNN), capable of extracting features from the topological structure, is considered as a solution for capturing global information. Nevertheless, GNN faces challenges as the manually defined static graph structure might not accurately capture the complexity of the data. We propose a double dual dynamic graph neural network (D3GNN) with dynamic topological structure refinement for multisource RS data classification. D3GNN generates multiple topological structures to achieve a comprehensive perception of scene features by utilizing local spatial information and distinctive data from various sources. Given the characteristics of heterogeneous-structure data, D3GNN implements targeted topological structure remodeling and refinement to overcome the limitations imposed by static graph, thereby enabling the network to generate feature embeddings with enhanced discriminative power. The experimental results show that D3GNN achieves superior performance compared to other current methods. |
|---|---|
| ISSN: | 1569-8432 |