Scalable Remote Sensing Image Change Captioning using In-Context Learning

Abstract

Remote Sensing Image Change Captioning (RSICC) plays a crucial role in detecting changes between bitemporal remote sensing images and generating descriptive captions that explain these alterations. Conventional approaches to developing such a successful system for RSICC tasks require substantial computational power, finely gathered images and captioning. In contrast, this study explores the advantages of In-Context Learning (ICL) combined with Visual Language Models for RSICC tasks to generate precise and accurate captions. The ICL approach involves presenting a few of demonstrations to the model with the queries to elicit expected outputs. Using the ICL, approximately 9x performance improvement is achieved on the LEVIR-CC dataset. Code available at https://github.com/ChangeCapsInRS/ICL-CC.