.png){kind=icon}
Return
Parameter-efficient dynamically evolved networks for multi-modal class-incremental remote sensing image classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2026.114768.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
•
We investigate a practical and challenging M2CI-RSIC task.
•
Modality-isolated, task-indexed LoRA alleviates feature-level representation drift.
•
Task-conditioned similarity calibration mitigates classifier-level logit mismatch.
•
Perturbed prototype replay reduces the instability of old class features.
•
Extensive experiments on three benchmarks demonstrate the superiority of PDEN.
Abstract
Class-incremental remote sensing image classification (CI-RSIC) has primarily been studied within a single-modality context, leaving the potential benefits of complementary information across modalities underexplored. In this paper, we investigate a more practical and challenging task, termed multi-modal class-incremental remote sensing image classification (M2CI-RSIC), where models effectively exploit multi-modal information while preserving previously acquired knowledge and adapting to newly emerging classes. Directly extending existing CI-RSIC methods to the above setting often leads to degraded old class representations and biased classifiers during incremental updates. To address these issues, we propose a parameter-efficient dynamically evolved network (PDEN) that jointly performs multi-modal feature adaptation and task-conditioned similarity calibration. Specifically, modality-isolated and task-indexed LoRA modules are introduced into the multi-modal feature branches to constrain parameter updates within low-rank subspaces, thereby strengthening old class representations and maintaining discrimination between old and new classes. Built upon the adapted features, a task-conditioned similarity calibration module further calibrates task-specific logits to alleviate classifier bias and stabilize decision boundaries across incremental tasks. Extensive experiments on three multi-modal remote sensing benchmarks demonstrate that PDEN consistently outperforms state-of-the-art CI-RSIC methods.
Keywords:
Remote sensing image classification
Class-incremental learning
Multi-modal learning
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
View more organizations
Cited Papers
No cited papers available