Remote Sensing Retrieval-Augmented Generation: Bridging Remote Sensing Imagery and Comprehensive Knowledge with a Multi-Modal Dataset and Retrieval-Augmented Generation Model

Recent progress in VLMs has demonstrated impressive capabilities across a variety of tasks in the natural image domain. Motivated by these advancements, the remote sensing community has begun to adopt VLMs for remote sensing vision-language tasks, including scene understanding, image captioning, and visual question answering. However, existing remote sensing VLMs typically rely on closed-set scene understanding and focus on generic scene descriptions, yet lack the ability to incorporate external knowledge. This limitation hinders their capacity for semantic reasoning over complex or context-dependent queries that involve domain-specific or world knowledge. To address these challenges, we first introduced a multimodal Remote Sensing World Knowledge (RSWK) dataset, which comprises high-resolution satellite imagery and detailed textual descriptions for 14,141 well-known landmarks from 175 countries, integrating both remote sensing domain knowledge and broader world knowledge. Building upon this dataset, we proposed a novel Remote Sensing Retrieval-Augmented Generation (RS-RAG) framework, which consists of two key components. The Multi-Modal Knowledge Vector Database Construction module encodes remote sensing imagery and associated textual knowledge into a unified vector space. The Knowledge Retrieval and Response Generation module retrieves and re-ranks relevant knowledge based on image and/or text queries, and incorporates the retrieved content into a knowledge-augmented prompt to guide the VLM in producing contextually grounded responses. We validated the effectiveness of our approach on three representative vision-language tasks, including image captioning, image classification, and visual question answering, where RS-RAG significantly outperformed state-of-the-art baselines.

💡 Research Summary

The paper addresses a critical gap in remote‑sensing vision‑language models (VLMs): the inability to incorporate external, domain‑specific, and general world knowledge when answering complex queries. To remedy this, the authors introduce two major contributions. First, they construct the Remote Sensing World Knowledge (RSWK) dataset, a large‑scale multimodal benchmark that pairs 14,820 high‑resolution satellite images of globally recognized landmarks with richly annotated textual descriptions. Each entry contains (i) remote‑sensing metadata such as surface reflectance, spectral indices (NDVI, EVI), atmospheric conditions, and acquisition parameters, and (ii) world‑knowledge narratives covering historical background, cultural significance, construction periods, and major events. The dataset spans 16 semantic categories across 184 countries, providing a breadth of information far beyond earlier remote‑sensing caption datasets (e.g., UCM‑Captions, RSICD) which only offer simple scene descriptions.

Second, the authors propose the Remote Sensing Retrieval‑Augmented Generation (RS‑RAG) framework. RS‑RAG consists of (1) a multimodal knowledge vector database construction module and (2) a knowledge retrieval and response generation module. In the first module, a CLIP‑style dual encoder is trained to map images (ViT‑B/16) and texts (BERT‑large) into a shared embedding space using contrastive loss and label smoothing, thereby aligning visual and textual modalities. All RSWK items are encoded and indexed with FAISS for efficient nearest‑neighbor search. In the second module, a query—either an image, a textual question, or a combination— is embedded, and the top‑K candidates are retrieved. Candidates are re‑ranked using a fused similarity score that combines image‑text and text‑text cosine similarities. The top‑N retrieved knowledge snippets are then inserted into a “knowledge‑augmented prompt” (e.g., “