A survey of multimodal remote sensing foundation models: Evolution, datasets, applications, and future directions

Recent advances in multimodal deep learning have accelerated the emergence of multimodal remote sensing foundation models (Mul-RSFMs), which are increasingly viewed as a promising paradigm for general-purpose geospatial intelligence. Unlike general multimodal foundation models developed for natural-image domains, Mul-RSFMs must address more complex challenges arising from distinct imaging mechanisms across sensors, large spatial–spectral–temporal variations, scarce or inconsistently annotated supervision, and substantial cross-region and cross-sensor distribution shifts. These challenges make remote sensing not merely a new application domain for foundation models, but a distinct setting that demands modality-aware representation learning, scalable pretraining, and robust adaptation across tasks and scenarios. To clarify the current landscape of Mul-RSFMs and provide a roadmap for their continued development, this survey presents a structured review that connects theoretical foundations, technical evolution, datasets and benchmarks, and downstream applications within a unified framework. From this perspective, the development of Mul-RSFMs can be understood as a progression from single-modal self-supervised pretraining to multimodal alignment and collaborative modeling, and more recently toward unified generative pretraining, while downstream applications have gradually expanded from low-level perception to mid-level interpretation and high-level reasoning. On this basis, three overarching challenge dimensions—representation, computation, and cognition—are further summarized, together with future directions toward more robust, efficient, and trustworthy remote sensing intelligence.