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

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

SAMBA, a novel Scatter-Guided Masked Bidirectional Mamba foundation model, is proposed for Synthetic Aperture Radar Automatic Target Recognition (SAR ATR). This model addresses the high computational complexity of Transformer-based architectures and the inadequacy of generic masking strategies for SAR imagery. SAMBA integrates a linear-complexity Mamba encoder with a mid-sequence class token, a three-level hierarchical Scattering-Guided Masked Autoencoder (SG-MAE) masking strategy, and a lightweight SpatialMix feature interaction module. Utilizing a two-stage cross-domain pre-training pipeline, SAMBA demonstrates superior performance across all pre-training configurations. It features only 27M parameters, significantly fewer than ConvNeXt-V2 (89M) and HiViT (66M), while achieving state-of-the-art results on seven downstream SAR classification and detection datasets, including MSTAR, SSDD, and SARDet-100K. The SG-MAE strategy alone boosted 5-shot classification accuracy from 71.5% to 80.6%.

Key takeaway

For Machine Learning Engineers developing SAR Automatic Target Recognition systems, you should consider Mamba-based architectures like SAMBA. Its linear complexity and scattering-guided masking offer superior performance and efficiency over traditional Transformer models, especially for high-resolution SAR imagery. Implement a two-stage cross-domain pre-training approach to maximize transfer capability and achieve state-of-the-art results on classification and detection tasks.

Key insights

SAR ATR foundation models benefit from Mamba architectures and scattering-guided masking for efficiency and accuracy.

Principles

• SAR target recognition requires specialized masking strategies.
• Linear-complexity models outperform quadratic Transformers for SAR.
• Two-stage cross-domain pre-training is optimal for SAR.

Method

SAMBA employs an overlapped patch embedding, a mid-positioned CLS token in a bidirectional Mamba encoder, and a three-level hierarchical SG-MAE masking strategy, complemented by a SpatialMix decoder.

In practice

• Use Mamba-based backbones for SAR ATR.
• Implement scattering-guided masking for SAR data.
• Adopt two-stage pre-training for SAR domain adaptation.

Topics

• SAR Automatic Target Recognition
• Mamba Foundation Models
• Self-Supervised Learning
• Masked Autoencoders
• State Space Models
• Remote Sensing Imagery

Code references

• mynswkk/SAMBA

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Physics-Driven Semantic Scattering Structure Understanding of Aircraft Target in SAR Images — Physics-driven semantic scattering structures improve SAR aircraft interpretation by linking electromagnetic responses to physical components.
• ConvVitMamba: Efficient Multiscale Convolution, Transformer, and Mamba-Based Sequence modelling for Hyperspectral Image Classification — ConvVitMamba efficiently classifies hyperspectral images by combining multiscale CNNs, Vision Transformers, and Mamba-inspired sequence mixing.
• MamBOA: State-Space Architecture for Video Recognition — MamBOA recasts selective state-space recurrence as a native motion synthesizer for robust video action recognition.
• CamoNAS: Neural Architecture Search for Enhanced Camouflaged Object Detection — CamoNAS uses frequency-aware Neural Architecture Search to automatically design top-performing models for camouflaged object detection.
• SSMamba: A Self-Supervised Hybrid State Space Model for Pathological Image Classification — SSMamba improves pathological image classification by addressing domain shift and enhancing local-global feature learning.
• On Sequence-to-Sequence Models for Automated Log Parsing — Mamba and Transformer models offer superior log parsing accuracy and sample efficiency, with Mamba providing cost advantages.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.