Learning to Change by Critique and Correction: A Synergistic Framework for Remote Sensing Change Detection and...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Change Detection (CD) and Change Captioning (CC) are crucial for monitoring land-cover evolution using remote sensing imagery.
• Current joint CD-CC models often suffer from mutual interference due to limited cross-task feedback, hindering performance.

Purpose of the Study:

• To propose a novel framework, Learning to Change by Critique and Correction (LCCC), for synergistic multi-task modeling of CD and CC.
• To introduce explicit cross-task feedback mechanisms to mitigate interference and enhance performance.

Main Methods:

• The LCCC framework reformulates CD and CC as a closed-loop critique-correction process with bidirectional feedback.
• A Text-Guided Critique Attention (TGCA) mechanism allows CC to supervise CD, while CD provides spatial constraints for CC.
• Reciprocal Suppression and Enhancement (RSE) and Key Complementary Feature Fusion (KCFF) modules are designed for representation purification and feature bridging.

Main Results:

• LCCC significantly outperforms existing methods in both change detection accuracy and change captioning quality.
• Experiments validate the effectiveness of the critique-correction paradigm for synergistic multi-task learning.
• The proposed method demonstrates generality across CD-CC tasks.

Conclusions:

• The LCCC framework establishes a synergistic relationship between CD and CC through mutual critique and correction.
• This approach effectively addresses the limitations of previous joint modeling techniques.
• LCCC offers a promising direction for advancing multi-task learning in remote sensing image analysis.