EUVIMEDEUVIMED entdecken
Crossref

Continual Learning With Memorized Dynamic Visual Description for Remote Sensing Scene Classification

Rui Peng, Wenzhi Zhao, Liqiang Zhang, Kun Jia, Qiao Wang, Fengcheng Ji

IEEE Transactions on Geoscience and Remote Sensing · 2026 · 64 · 1-19

DOI: 10.1109/tgrs.2026.3665622 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Geoscience and Remote Sensing, 2026. Verfasst von Rui Peng und 5 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 53 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Beijing Natural Science Foundation, National Natural Science Foundation of China Major Program.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The continual updating of data renders the design of model architectures time-consuming and energy-intensive when tackling diverse tasks, and it quite commonly confronts weak generalization performance, even leveraging fine-tuning or transfer learning. Furthermore, the model tends to be well adapted for the current task while suffering performance degradation on previous tasks, namely, the catastrophic forgetting problem. Hence, there is an urgent demand to maintain models to learn new knowledge and alleviate knowledge forgetting, specifically balancing the plasticity and the stability of the model during the continual learning (CL) process. To address the aforementioned challenges, we propose a dynamic visual description and playback mechanism that leverages deep feature clustering and memory across diverse tasks. The former extracts stable, invariant features by clustering typical visual characteristics of scenes, accounting for spatial patterns, seasonal variations, and lighting conditions, thereby enhancing the model’s comprehension of continuously evolving remote sensing scenes. The latter selectively memorizes features from specific samples to preserve historical model knowledge. This approach significantly strengthens the model’s capabilities, deepening its understanding of scene imagery while mitigating the forgetting of historical knowledge. We generate several sequence tasks on AID, RSI, NWPU, SEASONET, and different scaled merged datasets to simulate the CL process across different resolutions, seasonal variations, and illumination changes. Compared to the classical CL methods, we demonstrate the superior performance of our model in dealing with continually updated remote sensing data.

Förderung

  • National Natural Science Foundation of China
  • Beijing Natural Science Foundation
  • National Natural Science Foundation of China Major Program

Zitieren

APA

Peng, R., Zhao, W., Zhang, L., Jia, K., Wang, Q., & Ji, F. (2026). Continual Learning With Memorized Dynamic Visual Description for Remote Sensing Scene Classification. IEEE Transactions on Geoscience and Remote Sensing, 64, 1-19. https://doi.org/10.1109/tgrs.2026.3665622

Vancouver

Peng R, Zhao W, Zhang L, Jia K, Wang Q, Ji F. Continual Learning With Memorized Dynamic Visual Description for Remote Sensing Scene Classification. IEEE Transactions on Geoscience and Remote Sensing. 2026;64:1-19. doi:10.1109/tgrs.2026.3665622

BibTeX
@article{Peng2026,
  author = {Peng, Rui and Zhao, Wenzhi and Zhang, Liqiang and Jia, Kun and Wang, Qiao and Ji, Fengcheng},
  title = {Continual Learning With Memorized Dynamic Visual Description for Remote Sensing Scene Classification},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  year = {2026},
  volume = {64},
  pages = {1--19},
  doi = {10.1109/tgrs.2026.3665622}
}
Kostenlos starten →