Fluxamba: Topology-Aware Anisotropic State Space Models for Geological Lineament Segmentation in Multi-Source Remote Sensing

TL;DR

Fluxamba tackles the topology mismatch between rigid-state space model traversals and curvilinear geological lineaments by introducing a differentiable, topology-aware flux mechanism. At its core, the Structural Flux Block (SFB) combines the Anisotropic Structural Gate (ASG) with a Prior-Modulated Flow (PMF) to align information propagation with intrinsic geometry, while Hierarchical Spatial Regulator (HSR) and High-Fidelity Focus Unit (HFFU) enhance multi-scale alignment and noise suppression. The Boundary-Modulated Fusion head ensures sharp, boundary-preserving segmentation, yielding state-of-the-art results across LROC-Lineament, LineaMapper, and GeoCrack with a lightweight model (3.39M parameters) and real-time speed (>24 FPS). The approach achieves a new Pareto frontier for onboard planetary exploration, enabling high-fidelity, efficient mapping of geological lineaments under low-SNR conditions and setting a foundation for 3D extensions and self-supervised data utilization.

Abstract

The precise segmentation of geological linear features, spanning from planetary lineaments to terrestrial fractures, demands capturing long-range dependencies across complex anisotropic topologies. Although State Space Models (SSMs) offer near-linear computational complexity, their dependence on rigid, axis-aligned scanning trajectories induces a fundamental topological mismatch with curvilinear targets, resulting in fragmented context and feature erosion. To bridge this gap, we propose Fluxamba, a lightweight architecture that introduces a topology-aware feature rectification framework. Central to our design is the Structural Flux Block (SFB), which orchestrates an anisotropic information flux by integrating an Anisotropic Structural Gate (ASG) with a Prior-Modulated Flow (PMF). This mechanism decouples feature orientation from spatial location, dynamically gating context aggregation along the target's intrinsic geometry rather than rigid paths. Furthermore, to mitigate serialization-induced noise in low-contrast environments, we incorporate a Hierarchical Spatial Regulator (HSR) for multi-scale semantic alignment and a High-Fidelity Focus Unit (HFFU) to explicitly maximize the signal-to-noise ratio of faint features. Extensive experiments on diverse geological benchmarks (LROC-Lineament, LineaMapper, and GeoCrack) demonstrate that Fluxamba establishes a new state-of-the-art. Notably, on the challenging LROC-Lineament dataset, it achieves an F1-score of 89.22% and mIoU of 89.87%. Achieving a real-time inference speed of over 24 FPS with only 3.4M parameters and 6.3G FLOPs, Fluxamba reduces computational costs by up to two orders of magnitude compared to heavy-weight baselines, thereby establishing a new Pareto frontier between segmentation fidelity and onboard deployment feasibility.

Figures (11)

• Figure 1: Conceptual comparison of feature aggregation mechanisms across different architectures. (a) The input image representation with spatial resolution $H \times W$. (b) Convolution (CNNs): Aggregates features within a local receptive field, maintaining a linear computational complexity of $\mathcal{O}(HW)$. (c) Self-Attention (Transformers): Models global dependencies by densely calculating pairwise interactions between all pixels, incurring a quadratic complexity of $\mathcal{O}(H^2W^2)$. (d) State Space Models (SSMs): Recursively integrates context along specific scanning trajectories. While achieving near-linear complexity $\mathcal{O}(HW)$ similar to CNNs, it enables long-range dependency modeling with distance-dependent weight decay.
• Figure 2: Performance evaluation of the proposed Fluxamba. (a) Comparison with state-of-the-art (SOTA) methods in terms of parameter count, computational complexity (FLOPs), and segmentation accuracy (mIoU). Fluxamba (marked with a star) achieves a superior efficiency-accuracy trade-off. (b) Scalability analysis of Fluxamba variants (Tiny, Small, Base, Large) with different depth configurations. The radar chart illustrates the comprehensive impact of model scaling on computational cost (Size, FLOPs, Params) and segmentation metrics (F1, mIoU, OIS, ODS).
• Figure 3: Overall architecture of the proposed Fluxamba. The framework adopts a hierarchical encoder-decoder structure. On the left, the encoder utilizes stacked Structural Flux Blocks (SF Blocks) to extract multi-scale anisotropic features, denoted as $F_1$ to $F_4$, at varying resolutions through downsampling. On the right, the decoder incorporates the Boundary-Modulated Fusion (BMF) module, which dynamically aggregates these hierarchical features via DySample and ScaleGate operations using element-wise addition ($\oplus$). To enhance linear structural details, a specialized boundary head composed of a Conv-BN-ReLU sequence and a $1\times1$ convolution modulates the fused features via element-wise multiplication ($\otimes$). Finally, the segmentation head comprising Conv-BN-ReLU, dropout, and a $1\times1$ convolution layer generates the prediction map.
• Figure 4: Detailed architecture of the proposed Structural Flux Block (SFB). Designed as the core computational unit, the SFB orchestrates a topology-aware information flux through four synergistic stages: (1) Anisotropic Structural Gate (ASG): Extracts geometric priors to model long-range dependencies; (2) Prior-Modulated Flow (PMF): Rectifies the rigid FS2D scanning trajectories by dynamically gating the information flow based on ASG priors; (3) Hierarchical Spatial Regulator (HSR): Aligns feature semantics adaptively, utilizing Lightweight Modulation Refinement (LMR) for local detail preservation in shallow layers (Stages 1-2) and Global Transformer Reorganizer (GTR) for semantic coherence in deep layers (Stages 3-4); (4) High-Fidelity Focus Unit (HFFU): Maximizes the signal-to-noise ratio via dual-polarized modulation.
• Figure 5: Illustration of scanning primitives and the sequence generation pipeline. (a) Visualization of candidate scanning trajectories: diverse path patterns (e.g., Parallel, Diagonal) utilized to capture spatial context from varying orientations. (b) Serialization execution flow: The process of transforming 2D patches into direction-specific 1D sequences. These multi-route sequences provide the necessary anisotropic inputs for the subsequent Prior-Modulated Flow (PMF), allowing the model to dynamically gate and re-weight different strategies based on the target structure.