StarBOA: Real-Time Mamba State-Space Unrolling for Sparse Radar Micro-Doppler in ISAC Networks

Mustafa Bora Çelik; Ceren Çelik; Orhan Gazi

cs.LGSep 27, 2026

StarBOA: Real-Time Mamba State-Space Unrolling for Sparse Radar Micro-Doppler in ISAC Networks

Authors: Mustafa Bora Çelik, Ceren Çelik, Orhan Gazi

Abstract

In Integrated Sensing and Communications (ISAC), radar sensing must operate under chirp subsampling with up to 90% missing data. An attention-based baseline, limited to a 52ms buffer, collapses toward maximum uniform entropy ( $H=2.584$ bits) as sparsity increases, failing to capture long-range gait-cycle context. We propose StarBOA, which replaces attention with a causal Mamba state-space model that updates incrementally on a per-window basis without re-scanning past reconstructions. By maintaining a persistent state, StarBOA integrates over $100\times$ more temporal history at no additional per-step computational cost. StarBOA outperforms the baseline's published results across all sparsity levels, with SSIM gains increasing from $+0.0379$ at 50% missing data to $+0.2472$ at 90%. Each window is processed in 1.53ms with zero lookahead, demonstrating efficient causal reconstruction under extreme chirp subsampling.

Figures & tables

Spar.	Model	SSIM $\uparrow$	$\bm{s}$	$\bm{l}$	$\bm{c}$	RMSE $\downarrow$
50%	STAR-Attn [ 4 ]	0.8595	0.9245	0.9461	0.9400	0.0849
50%	StarBOA (Ours)	0.8822	0.9394	0.9565	0.9531	0.0746
75%	STAR-Attn [ 4 ]	0.8204	0.8997	0.9289	0.9240	0.0950
75%	StarBOA (Ours)	0.8536	0.9200	0.9427	0.9433	0.0868
90%	STAR-Attn [ 4 ]	0.7331	0.8661	0.8538	0.8504	0.1340
90%	StarBOA (Ours)	0.8047	0.8889	0.9137	0.9210	0.1107

TABLE I: Decoupled Front-End Benchmark on Frozen LIHT Under

50\%

75\%

, and

90\%

Sparsity on the DISC Dataset.

Fig. 2: Visual reconstruction under

90\%

extreme sparsity on the DISC mmWave benchmark (

v\in[-4.5,4.5]\text{ m/s}

t\in[0,4]\text{ s}

). (a) Ground Truth reference mmWave spectrogram. (b) Raw LIHT input under

90\%

pulse dropouts (severe noise and fragmented velocity profile). (c) STAR-Attention [ 4 ] (

\text{SSIM}=0.536

, diffuse moving-average smearing and blurred limb arcs). (d) Proposed StarBOA (

\text{SSIM}=\mathbf{0.7832}

, sharp torso line, continuous limb oscillation arcs, and high-contrast gait trajectory).

Horizon $\bm{H}$	Duration	SSIM $\uparrow$	Struct ( $\bm{s}$ )	RMSE $\downarrow$
$H=2$	$17.3\text{ ms}$	0.6861	0.8238	0.1457
$H=32$	$276.5\text{ ms}$	0.7560	0.8557	0.1263
$H=128$	$1105.9\text{ ms}$	0.7654	0.8601	0.1233

TABLE II: Reconstruction Fidelity vs. Recurrent Memory Horizon (

H

) at

90\%

Chirp Sparsity.

(a) End-to-End Pipeline Reconstruction Across Sparsity Tiers
Sparsity	Architecture	RMSE $\downarrow$	SSIM $\uparrow$	$\bm{\Delta}$ SSIM
50%	STAR [ 4 ] (reported)	0.0545	0.8840	Reference
50%	StarBOA (Ours)	0.0506	0.9219	$\mathbf{+0.0379}$
75%	STAR [ 4 ] (reported)	0.0779	0.7450	Reference
75%	StarBOA (Ours)	0.0772	0.8679	$\mathbf{+0.1229}$
90%	STAR [ 4 ] (reported)	0.1213	0.5360	Reference

TABLE III: Full End-to-End Pipeline Comparison and STAR Attention Distribution Under

90\%

Sparsity.

Explore similar work

SAMBA: A Scatter-Guided Masked Bidirectional Mamba Foundation Model for SAR Target Recognition
Jun 30, 2026Ke Wang, Xiaoyi Pan, Zhaoyu Gu +43D Masked Autoencoders Mamba-Based Models
College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China
Selection, Not Fusion: Radar-Modulated State Space Models for Radar-Camera Depth Estimation
May 12, 2026Zhangcheng Hou, Tomoaki OhtsukiDepth Estimation Multi-Sensor Fusion
School of Science and Technology Keio University Yokohama 223-8522, Japan
MambaCSP: Hybrid-Attention State Space Models for Hardware-Efficient Channel State Prediction
Apr 23, 2026Aladin Djuhera, Haris Gacanin, Holger BocheState Space Models Mamba-Based Models
Technical University Munich, Germany · RWTH Aachen University, Germany

Jun 30, 2026, cs.CVJ/K move · Enter open · S save

SAMBA: A Scatter-Guided Masked Bidirectional Mamba Foundation Model for SAR Target Recognition

Ke Wang, Xiaoyi Pan, Zhaoyu Gu, Xiaofeng Ai+3

College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China

3D Masked Autoencoders Mamba-Based Models Self-Supervised Learning Wireless Foundation Models

Open paper AskPDF arXiv

Synthetic aperture radar automatic target recognition (SAR ATR) is critical for Earth observation and defense, but its practical deployment is constrained by scarce annotated training data. Self-supervised pre-training alleviates this label bottleneck, yet prevailing Transformer architectures incur prohibitive quadratic computational complexity, and conventional universal masking neglects the unique electromagnetic scattering properties intrinsic to SAR imagery. To address these limitations, we propose SAMBA (Scattering-Guided Bidirectional Mamba), an efficient self-supervised pre-training foundation model for SAR target interpretation. Our framework features three core innovations: (i) a linear-complexity Mamba encoder with a mid-sequence class token to mitigate computational bottlenecks; (ii) a three-level hierarchical Scattering-Guided Masked Autoencoder (SG-MAE) masking strategy guided by SAR physical priors, aligning the pretext task with SAR's intrinsic imaging mechanism; (iii) a lightweight SpatialMix feature interaction module to enhance cross-region feature fusion. We also design a two-stage cross-domain pre-training pipeline to optimize the overall pre-training process. Extensive evaluations demonstrate that SAMBA consistently delivers superior performance across all pre-training configurations, with substantially fewer parameters than both CNN and Transformer baselines. Compared with the default masking strategy in standard MAE, the proposed SG-MAE strategy further boosts the model's few-shot transfer capability. Benchmarking on seven downstream datasets covering classification and detection tasks shows SAMBA achieves state-of-the-art (SOTA) performance on most metrics, fully validating its robust generalizability across diverse SAR interpretation tasks. Source code and pre-trained weights are publicly available at https://github.com/mynswkk/SAMBA.

Abstract

Explore similar work

SAMBA: A Scatter-Guided Masked Bidirectional Mamba Foundation Model for SAR Target Recognition

Selection, Not Fusion: Radar-Modulated State Space Models for Radar-Camera Depth Estimation

MambaCSP: Hybrid-Attention State Space Models for Hardware-Efficient Channel State Prediction