Vision Mamba in Remote Sensing: A Comprehensive Survey of Techniques, Applications and Outlook

TL;DR

The paper addresses the efficiency and scalability gap between CNNs and ViTs in remote sensing by advancing Vision Mamba, a selective State Space Model (SSM) framework. It surveys foundational SSM concepts, key Mamba variants (notably Mamba-2), and a broad spectrum of micro- and macro-architectural innovations, including scan strategies and multimodal interactions, benchmarking across tasks like classification, segmentation, detection, and restoration. The work highlights concrete methods (Context-SSM, Cross-SSM, RK2 discretization, Hybrid CNN/Mamba/Transformer designs, frequency-domain operations) and provides taxonomy-driven insights while outlining open challenges and future directions, such as non-causal SSM formulations and foundation-model-scale RS applications. The survey also offers an open-source repository to catalyze community-driven progress and emphasizes the practical impact of Mamba-based remote sensing systems in terms of efficiency and global-context modeling at scale.

Abstract

Deep learning has profoundly transformed remote sensing, yet prevailing architectures like Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) remain constrained by critical trade-offs: CNNs suffer from limited receptive fields, while ViTs grapple with quadratic computational complexity, hindering their scalability for high-resolution remote sensing data. State Space Models (SSMs), particularly the recently proposed Mamba architecture, have emerged as a paradigm-shifting solution, combining linear computational scaling with global context modeling. This survey presents a comprehensive review of Mamba-based methodologies in remote sensing, systematically analyzing about 120 Mamba-based remote sensing studies to construct a holistic taxonomy of innovations and applications. Our contributions are structured across five dimensions: (i) foundational principles of vision Mamba architectures, (ii) micro-architectural advancements such as adaptive scan strategies and hybrid SSM formulations, (iii) macro-architectural integrations, including CNN-Transformer-Mamba hybrids and frequency-domain adaptations, (iv) rigorous benchmarking against state-of-the-art methods in multiple application tasks, such as object detection, semantic segmentation, change detection, etc. and (v) critical analysis of unresolved challenges with actionable future directions. By bridging the gap between SSM theory and remote sensing practice, this survey establishes Mamba as a transformative framework for remote sensing analysis. To our knowledge, this paper is the first systematic review of Mamba architectures in remote sensing. Our work provides a structured foundation for advancing research in remote sensing systems through SSM-based methods. We curate an open-source repository (https://github.com/BaoBao0926/Awesome-Mamba-in-Remote-Sensing) to foster community-driven advancements.

Figures (11)

• Figure 1: A diagram that summarizes the pipeline of the survey.
• Figure 2: The architecture of Mamba-1 mamba and Mamba-2 Mambav2. (a) Mamba-1, a simplified attention block that integrates the SSM transformation directly into the primary computational pathway of the MLP block and uses an activation function on the gating pathway. (b) Sequential Mamba-2 Block, where operations are applied in series, and (c) Parallel Mamba-2 Block, where selected operations execute concurrently before fusion.
• Figure 3: Overview of Recent Advancements in SSM Formulations. This figure categorizes improvements into four major aspects: (a) Context-SSM 62-Mamba-MOC35-AFA-Mamba for local feature enhancement, integrating local context through CNNs; (b) Cross-SSM 84-COMO28-FusionMamba29-MSFMamba5-S2CrossMamba47-SegMamba-OS for multi-modal and bi-temporal feature interaction, enabling cross-modal/temporal feature interaction at the parameter level; (c) New discretization methods for SSM 112-RSVMamba, introducing second-order Runge-Kutta (RK2) for improved continuous-time approximation; and (d) Alternative hidden state definitions 68-TTMGNet, utilizing a Minimum Search Tree (MST) to redefine hidden state transitions. These refinements collectively enhance the efficiency and expressiveness of SSMs in various applications.
• Figure 4: Illustration of the scan strategy pipeline, comprising five key components, including feature preprocessing, scan sampling, scan directions, scan pattern, and feature post-processing. These five stages essentially transform a 2D feature map into multiple 1D sequences that conform to Mamba's processing architecture. For clarity, an additional example is provided demonstrating the case without any preprocessing or postprocessing operations.
• Figure 5: The seven feature preprocessing methods for scan strategy. For clarity, two additional preprocessing methods are visually shown in Fig. \ref{['fig:scan_1_pre2']}.