State Space Models Meet Remote Sensing: A Survey

Overview of SSMs in Remote Sensing

State Space Models (SSMs) have recently gained significant traction in the remote sensing community as a powerful alternative to traditional Transformer-based architectures. By leveraging linear computational complexity, SSMs address the bottleneck of quadratic scaling inherent in self-attention mechanisms, making them particularly well-suited for high-resolution, multi-modal, and temporal remote sensing data. This survey synthesizes the rapid development of these models from early 2024 through late 2025.

Evolution and Architectural Innovation

The transition of SSMs into the vision domain—often termed Vision-Mamba—has been marked by the adaptation of sequential modeling techniques to 2D spatial data. Remote sensing researchers have pushed this further by developing specialized scanning strategies (such as 2D selective scanning) and hybrid architectures that combine SSM blocks with CNNs, Transformers, and Graph Neural Networks. These innovations are designed to capture both global context and local spectral-spatial correlations, which are critical for tasks like hyperspectral image classification and change detection.

Task-Specific Applications

SSMs are currently being applied across a broad spectrum of remote sensing tasks. In classification, research focuses on multi-directional scanning to improve spatial sensitivity. In dense prediction tasks like semantic segmentation and object detection, SSMs are utilized to manage large spatial extents efficiently. Additionally, for temporal tasks such as change detection, SSMs excel at perceiving feature differences across time periods. The survey highlights that while many methods exist, the field is actively exploring how to best integrate these models into foundation-level frameworks for broader remote sensing utility.

Future Directions and Challenges

Despite the rapid progress, the field faces ongoing challenges, including the need for more robust multi-modal fusion techniques and the development of unified remote sensing foundation models based on SSMs. Future research is expected to focus on optimizing these architectures for hardware efficiency and improving their performance on noisy or heterogeneous data, ensuring that the theoretical advantages of linear scaling translate into practical, scalable solutions for real-world Earth observation.