Answers are grounded in this paper and cite their sources.
Research on remote sensing multimodal large language models (RS-MLLMs) has advanced rapidly in recent years, but existing studies remain scattered across different model architectures and application tasks, highlighting the need for a systematic synthesis to establish a clear overall understanding of the field. In addition, most existing surveys on RS-MLLMs narrowly focus on early vision-language models (VLMs) or merely provide catalogs of architectures and datasets. This paper presents a comprehensive review of RS-MLLMs from both system-level and task-oriented perspectives on a tutorial-oriented basis. First, the necessity of RS-MLLMs is discussed by examining the domain gap and internal heterogeneity of RS data, the limitations of visual-only paradigms, and the transitional role of VLMs. Subsequently, the system-level evolution of RS-MLLMs is comprehensively summarized, including architectural components, training and fine-tuning strategies, evaluation, and RS-Agent paradigms. Meanwhile, downstream RS-MLLM applications are analyzed with a unified capability-oriented taxonomy, covering language generation, visual question answering, grounding and localization, cross-modal retrieval, and temporal change understanding. Finally, the current challenges of RS-MLLMs are discussed, and future research directions, including parameter-efficient adaptation, spatio-temporal reasoning, and knowledge-driven architectures, toward more reliable, efficient, and geospatially aware multimodal intelligence, are envisioned.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: