Hierarchical Deep Features Progressive Aggregation for Remote Sensing Images Scene Classification

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 17; pp. 9442 - 9450
• Main Authors: Zhao, Yang, Liang, Jiaqi, Huang, Sisi, Huang, Pingping
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Aggregation; Agricultural production; Classification; Complexity; Convolutional neural networks; Cooperative learning; Digital agriculture; Distribution functions; Ecological monitoring; Feature extraction; Graphical models; Image classification; Interference; Modules; Multilevel local (MLL) and multilevel global (MLG) feature modules; progressive aggregation (PA); Redundancy; Remote sensing; remote sensing images scene classification; Representations; Scene classification; Semantics; Spatial discrimination; Spatial resolution
• DOI: 10.1109/JSTARS.2024.3391332
• ISSN: 1939-1404

Remote sensing image scene classification is essential, and it can promote the rational planning of land and ecological monitoring in the practical application of agricultural production. High spatial resolution (HSR) remote sensing images are widely used in smart agriculture because of their wide coverage and HSR. The HSR remote sensing images have a more detailed description of the local scene. However, the complexity of scene details intensifies the intraclass diversity and interclass similarity of scenes, and the interference to scene classification is more significant. To distinguish scene categories effectively in complex background, this article proposes a scene classification method of remote sensing images based on progressive aggregation (PA) with local and global cooperative learning. Specifically, multilevel local and global feature modules are employed at different levels to describe the influence of local objects and their global distribution on scene category determination. Then, the PA module is introduced to explore the collaboration of the same level features and reduce the interference of shallow redundancy. The residual structure can establish the correlation between multilevel representations, thereby improving the representation of the aggregate features. To verify the performance of the proposed method, we implemented cross-domain experiments on four internationally available remote sensing image classification datasets: NWPU-RESISC45, WHU-RS19, RSSCN7, and AID. The experimental results show that the proposed method is effective and robust in remote sensing scene classification.