Remote sensing image captioning: a survey from encoder-decoder model to large vision-language model

Remote sensing images provide rich evidence about land cover, land use, and environmental conditions. Effective exploitation of this evidence underpins tasks in resource management, environmental surveillance, and urban planning. Remote sensing image captioning (RSIC) involves obtaining textual descriptions from remote sensing images by capturing semantic relationships between objects and attributes. This paper reviews the development of RSIC with emphasis on remote sensing image change captioning (RSICC) and the transition from task-specific encoder-decoder models to large vision–language models (LVLMs). Firstly, representative encoder–decoder approaches and LVLMs are illustrated. Afterwards, a comparative analysis of typical encoder-decoder models and LVLMs is presented, which shows that encoder–decoder models leverage task-specific mechanisms such as difference-aware attention and multi-scale fusion for RSIC, whereas LVLMs exploit pre-trained vision-language backbones with parameter-efficient alignment to support captioning and visual question answering. Finally, conclusions and recommendations are outlined, highlighting improved captioning accuracy and generalization, supported by better remote sensing datasets and LVLM alignment, as key research techniques in future.