.png){kind=icon}
Return
Difference Approximation Network With Semantic Alignment for Remote Sensing Change Captioning
.png){kind=icon}
PRE
AI
DOI：10.1109/tgrs.2026.3739247.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Remote sensing change captioning (RSCC) plays an important role in advancing remote sensing spatiotemporal understanding with vision-language generation. However, due to the temporal evolution of geospatial patterns, bitemporal images are susceptible to the influence of lighting conditions and seasonal effects. Existing methods fail to effectively utilize the cross-sample difference information and perform explicit multimodal alignment of key changes, which often results in inaccurate descriptions of changes in key objects and their relationships. To address these issues, the article proposes a novel difference approximation network with semantic alignment (DANSA) for RSCC. Specifically, a cross-temporal feature transform (CFT) is introduced to interact with feature representations extracted from the contrastive language-image pretraining (CLIP) model. Then, to learn a reliable difference representation, a difference memory approximation (DMA) is designed to enhance the difference representation of the current scene by activating difference information from training samples. Furthermore, considering that pretrained language embedding models can serve as a powerful supervisory signal, a change semantic alignment (CSA) is constructed to achieve fine-grained alignment between visual and textual features. In this way, DANSA generates comprehensive scene change captions that accurately depict objects and relationships across time. Extensive experiments and analyses conducted on three benchmark datasets demonstrate that DANSA achieves competitive performance with fewer training parameters. The code will be available at https://github.com/zcsisiyao/DANSA
Keywords:
Change captioning
contrastive language-image pretraining
cross-modal alignment
remote sensing
Journal
IF:
8.6
Papers:
2.2W
Citations:
10.7W
Organization
No organization information available
Cited Papers
No cited papers available