Related Experiment Videos
BiCM-Prompt: Bidirectional Cross-Modal Prompt Tuning for Class-Incremental Learning on Multisource Remote Sensing
Summary
This study introduces a new prompt-based framework for class-incremental learning in multi-source remote sensing image classification. The method enhances stability and plasticity for continuous learning across diverse data sources.
Area of Science:
- Computer Science
- Remote Sensing
- Artificial Intelligence
Background:
- Class-incremental learning (CIL) is crucial for expanding category recognition in remote sensing without forgetting past data.
- Existing CIL methods often fail in multi-source scenarios by not leveraging complementary information across different data origins.
- Remote sensing applications require adaptive classification systems capable of long-term learning.
Purpose of the Study:
- To develop a novel prompt-based framework for class-incremental learning specifically designed for multi-source remote sensing image classification.
- To address the limitations of single-source CIL methods in exploiting multi-modal data.
- To enhance both model stability and plasticity for effective incremental learning in complex remote sensing environments.
Main Methods:
- Proposed a Bidirectional Cross-Modal Prompt Tuning (BiCM-PT) module for dynamic prompt selection and preservation of cross-modal relationships.
- Introduced a Prompt-Guided Knowledge Aggregator (PGKA) to guide feature aggregation and extract discriminative multi-modal representations.
- Utilized prompt-based learning to improve stability by freezing modality relation projectors and plasticity for new-class learning.
Main Results:
- Demonstrated effective and stable class-incremental learning in multi-source remote sensing environments.
- Achieved a balance between stability (preserving old knowledge) and plasticity (learning new knowledge).
- Validated the approach's effectiveness on three real-world remote sensing benchmarks.
Conclusions:
- The proposed prompt-based framework significantly improves class-incremental learning performance in multi-source remote sensing.
- The BiCM-PT and PGKA modules effectively handle the challenges of catastrophic forgetting and multi-modal data integration.
- The method offers a promising solution for long-term adaptive classification in remote sensing applications.
