StreamAZ: Line-by-Line SAR Focusing with State Space Models
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "StreamAZ: Line-by-Line SAR Focusing with State Space Models".
Dev: Here's the gist of the paper. Conventional focusing methods for Synthetic Aperture Radar (SAR) employ block processing efficiently but remain latency-heavy processes that prevent the realisation of a closed-loop cognitive SAR vision system.
Rosa: We present the first Online SAR Processor (OSP), an online image-formation framework that treats SAR sensing as a stream and produces focused SAR image output line by line during acquisition.
Dev: OSP uses a tiny state-space surrogate model trained with teacher-student distillation and multi-stage losses.
Rosa: Relative to a linewise digital-signal-processing baseline, OSP delivers approximately 70× lower latency and 130× lower memory use.
Title and authors: Rosa: So we've talked about the basics of StreamAZ: Line-by-Line SAR Focusing with State Space Models. Now let's look at what it means in terms of the specific title and authors, and what that implies for the research direction.
Dev: The authors are Fieldhouse, Del Prete, and Daga, along with Rensly and Meoni from ESA and National Tsing Hua University. They're clearly bringing together expertise from radar signal processing, space agency applications, and advanced control systems.
Rosa: Exactly. It tells us right away that this isn't just a theoretical math exercise; it’s rooted in the practical challenges of real satellite data processing. The focus on state-space models points toward a dynamic approach to solving this problem as the data flows in.
Taro: From my side, I'm thinking about the implication of focusing on linear synthetic aperture assumption. That assumption is quite restrictive, so it makes me wonder how robust this system really is when the world doesn't behave perfectly linearly.
Rosa: That's a fair question, Taro. The authors themselves flag that they assume a linear synthetic aperture and stripmap acquisition with minimal range cell migration as their operating regime for this work.
Dev: They also made it clear that failure cases are most likely when assumptions are violated by large residual motion errors or stronger than expected range migration.
Taro: That's a crucial piece of information for anyone trying to implement this in the real world. It means we can't just assume it works perfectly everywhere; we have to be ready for those motion errors.
Rosa: And they also noted that because OSP is causal and doesn't store the full aperture, you should expect some loss of fidelity relative to the best offline methods when you are in challenging scenes where those assumptions break down.
Dev: So, it’s not a perfect replacement for everything; it’s a component in a larger chain, which is what they suggest.
Taro: Right. It sounds like we're looking at a tool that excels under specific conditions but needs to be integrated carefully into the overall pipeline.
Rosa: So, the big picture here is moving SAR focusing away from heavy block processing and toward a stream-based inference problem.
The paper's summary: Dev: We’ve covered the setup now. To summarize what StreamAZ actually proposes, it re-formulates SAR image formation as an online inference problem under the linear synthetic aperture assumption.
Rosa: That's right. The main contribution is introducing the Online SAR Processor, OSP, which is a state-space surrogate for line by line azimuth focusing with a fixed size recurrent state and about two hundred trainable parameters.
Taro: So it’s not trying to solve the whole imaging problem at once; it's focusing on one range line at a time, updating the hidden state as you process each subsequent azimuth sample.
Dev: Precisely. It processes input data, gets an output for that specific range line, and then updates its internal state based on the next sample in sequence.
Rosa: And they define an azimuth focusing training loss called LAF which combines complex data fidelity terms, log amplitude terms, and structural regularization components to model the SAR structure along the azimuth dimension.
Taro: I'm interested in that loss function again. How does it specifically help the model learn what a good focused image looks like?
Dev: The LAF includes terms like Lcomplex for the primary data fidelity anchor and an Lfocus width term to penalize mainlobe broadening, which is key for getting that sharp focus.
Rosa: And they use teacher-student distillation to train this student model, which helps it preserve the dominant morphology of both held out scenes while achieving higher amplitude correlation and complex coherence than if it were only trained on hard targets.
Taro: So they’re not just training it to hit a peak; they're training it to learn the shape of the focus itself.
Dev: That's right. They are modeling SAR structure along that azimuth dimension using this compact model while maintaining a very small parameter count, around two hundred parameters.
Rosa: It really shows how this state-space surrogate is used as a backbone for efficiently performing the focusing operation in an online manner.
The paper's improvements: Taro: Now that we know what it does, let's talk about the actual improvements they claim. What makes this approach better than just running some standard linewise digital signal processing?
Dev: The primary improvement is the massive reduction in latency and memory usage. They report seventy times lower latency and one hundred thirty times lower memory use compared to a linewise digital signal processing baseline.
Rosa: That's a huge jump, especially when you look at the operational constraints of real-time systems. It moves us closer to those closed-loop cognitive SAR vision systems we were aiming for.
Taro: And from an engineering standpoint, the paper talks about how they reduced the complexity of processing. They state that their computational complexity is O(N r log(N r)) for full scene total processing when using fixed memory footprint state space model layers.
Dev: That’s better than the O(N r log N r) complexity that you get from a batched RDA implementation, which confirms the efficiency gain from this streaming approach.
Rosa: They also achieved a very efficient scaling of O(N r log two N r) when processing full scans because the model uses fixed memory footprint state space model layers that do not require buffering data in the azimuth direction <ref:2605.10340#pg1>.
Taro: So, by avoiding that full aperture buffer, you’re essentially removing a major source of computational overhead. That's a big win for real time processing.
Dev: And this efficiency is tied directly to how they implemented the state-space surrogate as a compact causal dynamical system using those state space models.
Rosa: The way they manage the memory footprint by not buffering data in the azimuth direction is what leads to that superior memory footprint compared to both batched and linewise RDA implementations.
Conclusion: Dev: So, wrapping up this discussion on StreamAZ: Line-by-Line SAR Focusing with State Space Models. The paper shows a way to treat image formation as an online inference problem under the linear synthetic aperture assumption.
Rosa: In essence, they present the Online SAR Processor, OSP, which is a state space surrogate for line by line azimuth focusing that achieves seventy times lower latency and one hundred thirty times lower memory use than traditional linewise digital signal processing.
Taro: For me, the implication is that we can use this to support immediate decision making through CFAR vessel detection and flood mapping in a fast, stream-based fashion.
Dev: And that speed allows for rapid localization of a vessel in a restricted maritime area when you need it most, because they demonstrate rapid localization capabilities.
Rosa: We're leaving the discussion on this paper by acknowledging its limitations: it’s strictly tied to linear synthetic aperture and stripmap acquisition with minimal range cell migration.
Taro: And they explicitly state that OSP is causal, so we can't exploit arbitrarily large future context like a fully offline batch imager, meaning some loss of fidelity relative to the best offline methods should be expected in challenging scenes where those assumptions break down.
Dev: So, the big picture is that StreamAZ gives us a fast way to get focused outputs with good enough quality for immediate situational awareness without needing to replace all our classical estimation machinery.
Rosa: We've gone through the details of StreamAZ: Line-by-Line SAR Focusing with State Space Models, and that’s where we are for this paper. Thanks for tuning in.
College of Semiconductor Research National Tsing Hua University European Space Agency
eess.IV, cs.CE, cs.ET
Submitted: 2026-05-11
Updated: 2026-11-06
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Conventional focusing methods for Synthetic Aperture Radar (SAR) employ block processing efficiently but remain latency-heavy processes that prevent the realisation of a closed-loop cognitive SAR…
Key concepts
- Online SAR Processor (OSP)
- OSP treats SAR sensing as a continuous stream, producing focused image output line by line during data acquisition. It uses a tiny state-space surrogate model to perform azimuth focusing without storing the entire aperture, drastically reducing latency and memory compared to conventional block processing.
- State-Space Surrogate Model
- This is a compact mathematical model used within OSP to approximate the complex process of line-by-line azimuth focusing. It uses a fixed-size recurrent state with only about 200 trainable parameters, allowing it to efficiently capture the necessary focusing dynamics for each incoming data row.
- Ground-Truth-Aligned Azimuth-Focusing Loss (LAF)
- This loss function guides the model training by combining several components: complex data fidelity, log-amplitude terms, and structural regularization. It specifically penalizes mainlobe broadening and ensures the model learns to accurately reconstruct SAR structures across the azimuth dimension.
- Teacher–Student Distillation
- This training technique involves a high-capacity 'teacher' model guiding a smaller 'student' model. The student is trained to mimic the teacher’s complex fidelity and structural morphology, ensuring it preserves important features from held-out scenes while achieving better amplitude correlation.
Terminology
Summary
Conventional focusing methods for Synthetic Aperture Radar (SAR) employ block processing efficiently but remain latency-heavy processes that prevent the realisation of a closed-loop cognitive SAR vision system. p. 0 int
Kea-Tiong Tang
Department of Electrical Engineering
National Tsing Hua University
kttang@ee. tw
Abstract
Conventional focusing methods for Synthetic Aperture Radar (SAR) employ block
processing efficiently but remain latency-heavy processes that prevent the realisation of a closed-loop cognitive SAR vision system. tw
Abstract
Conventional focusing methods for Synthetic Aperture Radar (SAR) employ block
processing efficiently but remain latency-heavy processes that prevent the realisation of a closed-loop cognitive SAR vision system. We present the first Online SAR
Processor (OSP), an online image-formation framework that treats SAR sensing
as a stream and produces focused SAR image output line by line during acquisition. We present the first Online SAR
Processor (OSP), an online image-formation framework that treats SAR sensing
as a stream and produces focused SAR image output line by line during acquisition. OSP uses a tiny state-space surrogate model trained with teacher-student
distillation and multi-stage losses. We evaluate the method on 300GB of SAR
data from Maya4, a Sentinel-1-derived dataset containing raw, range-compressed,
range-cell-migration-corrected, and azimuth-compressed products. Relative to a
linewise digital-signal-processing baseline, OSP delivers approximately 70× lower
latency and 130× lower memory use; on a single AMD CPU core it processes
one row in 16 ms with a memory footprint of 6 MB whilst maintaining a focusing
quality high enough to support downstream decisions, which we illustrate with
vessel detection and flood-mapping tasks.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The core contribution is the Online SAR Processor (OSP), which re-formulates SAR image formation as an online inference problem under the linear synthetic aperture assumption. This framework introduces a state-space surrogate for lineby-line azimuth focusing with fixed-size recurrent state, approximately 200 trainable parameters, and an explicit deployment mode that processes each incoming row without storing the full aperture. p. 1 In order to address these limitations and take a step towards a CSAR system, this work makes the
following contributions:
• We re-formulate SAR image formation as online inference problem under the linear synthetic
aperture assumption. • We introduce the very first Online SAR Processor (OSP), a state-space surrogate for lineby-line azimuth focusing with fixed-size recurrent state, ∼200 trainable parameters, and
an explicit deployment mode that processes each incoming row without storing the full
aperture. We evaluate the method on 300GB of SAR
data from Maya4, a Sentinel-1-derived dataset containing raw, range-compressed,
range-cell-migration-corrected, and azimuth-compressed products. Relative to a
linewise digital-signal-processing baseline, OSP delivers approximately 70× lower
latency and 130× lower memory use; on a single AMD CPU core it processes
one row in 16 ms with a memory footprint of 6 MB whilst maintaining a focusing
quality high enough to support downstream decisions, which we illustrate with
vessel detection and flood-mapping tasks.Where the paper says it
Where the paper says it
How the training signal is built
The azimuth-focusing loss is defined as LAF, which combines several terms including complex data-fidelity, log-amplitude terms, and structural regularization components. The loss function LAF includes terms such as Lcomplex for the primary data-fidelity anchor and Lfocus-width to penalize mainlobe broadening. p. 4 We summarize each of the subterms of LAF in Table 1, and further elaborate in
Appendix A. The complex term Lcomplex is the primary data-fidelity anchor: because it acts directly in
complex space, it penalizes residual phase and amplitude mismatch before magnitude-only surrogates
can mask physically incorrect solutions. The focus-width term Lfocus-width penalizes mainlobe broadening through an
autocorrelation-derived width statistic that serves as a compact proxy for focusing concentration,
and the azimuth-edge term Laz-edge matches second-order differences to regulate local curvature,
suppress ringing, and reject both over-smoothed and oscillatory solutions. 2 Multi-Stage Teacher Optimization
Teacher selection was carried out as a staged optimization procedure over a shared Stage-0 baseline. normalized input azimuth strip u, reference AZ strip z, teacher prediction zˆT = Tϕ(u), and student
prediction zˆS = Sθ(u), the teacher parameters are frozen and the student is trained with
Lstudent = LAF(ˆzS, z) + λKD D(ˆzS,sg[ˆzT ]), (4)
where sg[·] denotes stop-gradient and LAF is the physically motivated azimuth-focusing loss defined
in Eq. With reference to Figure 3, the distillation discrepancy is evaluated in the same complex
domain as the reconstruction loss,
D(a, b) = αc∥a − b∥1 + αℓ ∥log(|a| + ϵ) − log(|b| + ϵ)∥1
+ αρ (1 − ρ(|a|, |b|)) + αϕ ⟨1 − cos (∠a − ∠b)⟩.Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The efficiency of OSP stems from representing the azimuth-focusing operator as a compact causal dynamical system using state-space models. The model uses fixed memory footprint SSM layers as its backbone and does not require buffering data in the azimuth direction, leading to a superior memory footprint compared to both batched and linewise RDA implementations. p. 8 We state all our assumptions
and derivations in Appendix C. Due to the fact our model uses fixed memory footprint SSM layers as its backbone and does not
require buffering data in the azimuth direction, we obtain a far superior memory footprint to both
baselines; our online formulation also enables drastically reduced data-in-to-data-out focusing latency. Lower latency, memory, and
GFLOPs per row are better. Method Complexity Single-line Full scan Memory GFLOPs GFLOPs
latency (ms)∗
latency (ms)∗
footprint per scan per row
Traditional RDA O(NaNr log(NaNr)) N/A†
27,444 16 GB 122 N/A†
Linewise RDA O(NaNrNb log(NbNr)) 857 (+243)
‡
17,140,000 780 MB 45,400 2.Where the paper says it
Where the paper says it
Downstream Applications
The focused output quality supports simple downstream decisions such as CFAR vessel detection and fixed threshold flood mapping. For vessel detection, a linewise CFAR detector is applied to an OSP-focused scene over Porto de Santos, Brazil, demonstrating rapid localization capabilities. p. 7 By providing a
rapid localization of a vessel in a restricted maritime area using OSP and linewise CFAR, it could be
possible for the SAR system to take a second, high-resolution scan via spotlight whilst still flying
overhead the detected vessel of interest for identification purposes and improve enforcement on fishing
restrictions etc. As a proof-of-concept we apply a linewise CFAR detector to an OSP-focused scene
over Porto de Santos, Brazil. As a proof-of-concept we apply a linewise CFAR detector to an OSP-focused scene
over Porto de Santos, Brazil. Rapid flood-mapping is also another time-sensitive downstream task
8Where the paper says it
Where the paper says it
Limitations and operational considerations
The study is intentionally narrow in its operating regime, assuming a linear synthetic aperture, stripmap acquisition, and minimal range cell migration. Failure cases are most likely when assumptions are violated by large residual motion errors or stronger-than-expected range migration. p. 1 2 Related Work
We position our contribution between (i) classical SAR image formation pipelines and (ii) modern
sequence models for long-range operators. 1 SAR Image Formation
SAR image formation maps raw complex phase history to a reflectivity image by compensating range
migration, azimuth phase evolution, motion errors, and system-dependent distortions. We
also present extended discussion including limitations, failure cases, operational considerations and
robustness in Appendix G. 6 Conclusions
We introduce the Online SAR processor, a streaming formulation of SAR image formation in which
azimuth focusing is performed incrementally via a compressed phase-history rather than after full9Where the paper says it
Where the paper says it
Open access to data and code
The authors use open source data from ESA’s Sentinel-1 satellites for model training and experiments. They intend to make their code openly available on github upon acceptance. p. 1 FPGA, GPU, and embedded
implementations of RDA, CSA, SPECAN, backprojection, and related methods use pipelined FFTs,
reduced precision, hardware-aware memory layouts, and parallel scheduling to lower latency and
downlink burden [10–14]. However, this notion of real time is mainly throughput-oriented: processing
is accelerated once the required aperture support is available. 4
4 Experiments
We evaluate OSP using a fixed Maya4 split, deterministic strip sampling, origin-preserving complex
normalization, and a recurrent deployment protocol matching the causal SSM boundary conditions
used during training. Maya4 [25] is a Sentinel-1-derived dataset of 2TB, which we split into
training, validations and test data, that provides intermediate SAR representations from Level 0 raw
measurements to Level 1 focused imagery.Where the paper says it
Where the paper says it
Experimental setting/details
Training was run on PBS gpu4 std CUDA nodes with one GPU, 16 CPU cores, and 32GB RAM per training job. The student training uses single-range-bin azimuth strips of length L = 1000, sampled in row order with patch size=[1000,1] and stride=[1000,1]. p. 1 In order to address these limitations and take a step towards a CSAR system, this work makes the
following contributions:
• We re-formulate SAR image formation as online inference problem under the linear synthetic
aperture assumption. • We introduce the very first Online SAR Processor (OSP), a state-space surrogate for lineby-line azimuth focusing with fixed-size recurrent state, ∼200 trainable parameters, and
an explicit deployment mode that processes each incoming row without storing the full
aperture. 3) approach: first, a higher-capacity
teacher is optimized for reconstruction quality under the azimuth-focusing objective described in
Section 4. 1, and then the compact student is then trained to reproduce both the focused target and
the teacher’s response. The LSTM/GRU screen is listed for
completeness. Stage N Swept parameters Selected setting R ↓ C ↑ Q ↑ ∆
‡
R ∆
‡
C
NA†
16 Sequence-model debug sweep: model ∈
{LSTM, GRU}, hidden dimension ∈
{6, 9}, layers ∈ {1, 2}, learning rate ∈
{10−4
, 3×10−4}.Where the paper says it
Where the paper says it
Where the paper says it
Experiment statistical significance
Some of the experiments don’t necessarily suit using error bars or statistical significance as a way of representing information. The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.
Experiments compute resources
The training was run on PBS gpu4 std CUDA nodes with one GPU, 16 CPU cores, and 32GB RAM per training job. Inference and metric evaluation used one GPU, 16 CPU cores, and up to 64GB RAM. p. 0 We evaluate the method on 300GB of SAR
data from Maya4, a Sentinel-1-derived dataset containing raw, range-compressed,
range-cell-migration-corrected, and azimuth-compressed products. Relative to a
linewise digital-signal-processing baseline, OSP delivers approximately 70× lower
latency and 130× lower memory use; on a single AMD CPU core it processes
one row in 16 ms with a memory footprint of 6 MB whilst maintaining a focusing
quality high enough to support downstream decisions, which we illustrate with
vessel detection and flood-mapping tasks.Where the paper says it
Code of ethics
The authors conform to the NeurIPS code of Ethics. They use open source data from ESA’s Sentinel-1 satellites for our model training and experiments. p. 3 4
4 Experiments
We evaluate OSP using a fixed Maya4 split, deterministic strip sampling, origin-preserving complex
normalization, and a recurrent deployment protocol matching the causal SSM boundary conditions
used during training. Maya4 [25] is a Sentinel-1-derived dataset of 2TB, which we split into
training, validations and test data, that provides intermediate SAR representations from Level 0 raw
measurements to Level 1 focused imagery. Real-time and on-board SAR
processing has largely focused on accelerating classical pipelines. FPGA, GPU, and embedded
implementations of RDA, CSA, SPECAN, backprojection, and related methods use pipelined FFTs,
reduced precision, hardware-aware memory layouts, and parallel scheduling to lower latency and
downlink burden [10–14]. Learning to Focus Synthetic Aperture Radar On-line
with State-Space Models
Sebastian Fieldhouse∗
College of Semiconductor Research
National Tsing Hua University
sebastianfieldhouse. com
Roberto Del Prete∗
Φ-lab, European Space Agency (ESA)
roberto.Where the paper says it
Where the paper says it
Where the paper says it
Broader impacts
The paper discusses potential positive societal impacts of the work performed, especially with regard to the downstream applications of our OSP, for example flood detection and marine area protection. They are limited from having a very in-depth discussion on societal impacts by the 9-page page limit. We are not aware of any specific potential negative societal impacts of our work.
Safeguards
We are not aware of any risk of misuse of our work. Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model. p. 0 Relative to a
linewise digital-signal-processing baseline, OSP delivers approximately 70× lower
latency and 130× lower memory use; on a single AMD CPU core it processes
one row in 16 ms with a memory footprint of 6 MB whilst maintaining a focusing
quality high enough to support downstream decisions, which we illustrate with
vessel detection and flood-mapping tasks. Keywords: synthetic aperture radar; sequence models; image formation.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Online SAR Processor (OSP) can perform image formation
line by line during acquisition
with approximately 70 times lower latency and 130 times lower memory use compared to traditional linewise digital-signal-processing baselines. This allows for aclosed-loop cognitive SAR vision system.
-
The OSP can process one row in 16 ms on a single AMD CPU core with a memory footprint of 6 MB while maintaining focusing quality high enough to support downstream decisions like vessel detection and flood mapping.
-
The system can be deployed for rapid decision support by applying CFAR-based vessel detection and fixed threshold flood mapping to OSP-focused outputs in a line-by-line fashion, enabling
rapid localization of a vessel in a restricted maritime area.
-
The model uses a tiny state-space surrogate with only 208 trainable parameters, which is shared across all range bins while maintaining range-bin specific hidden states, allowing it to
focus an entire line while keeping the number of trainable weights extremely small.
-
The system can utilize a Ground-Truth-Aligned Azimuth-Focusing Loss (LAF) that incorporates terms like the complex data fidelity anchor, log amplitude term, and focus width term to model SAR structure along the azimuth dimension.
-
Teacher–Student Distillation enables the student model to preserve
dominant morphology of both held-out scenes
while achieving higher amplitude correlation and complex coherence than if trained only on hard targets. -
The Online Processor can be implemented in a
linewise RDA
mode, which maintains a rolling buffer of 972 range lines, allowing for focusing of a single range line at a time with an estimated total complexity of O(NaNr Nb log(NbNr)). -
The system can achieve computational efficiency scaling as O(NaNr log2 Nr) when processing full scans because the model uses fixed memory footprint SSM layers that do not require buffering data in the azimuth direction.
Abstract
Conventional focusing methods for Synthetic Aperture Radar (SAR) employ block-wise processing efficiently but remain latency-heavy processes that prevent the realisation of a closed-loop cognitive SAR system. We present an online SAR image-formation framework, StreamAZ, that treats sensing as a stream and produces focused images in a line by line fashion. Our method uses a tiny state-space surrogate model trained with teacher-student distillation and multi-stage losses. We evaluate the method on real SAR data from Maya4, a Sentinel-1-derived dataset containing raw, range-compressed, range-cell-migration-corrected, and azimuth-compressed products. Relative to a linewise digital-signal-processing baseline, our framework delivers approximately 70 times lower latency and 130 times lower memory use; on a single AMD CPU core it processes one row in 16 ms with a memory footprint of 6 MB whilst maintaining a focusing quality high enough to support downstream decisions, which we illustrate with vessel detection and flood-mapping tasks.
Sources
Related papers
- HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- FDIR: Fidelity-Anchored Decompressed Image Restoration for Human and Machine Perception
- A Disease-Centric Vision-Language Foundation Model for Precision Oncology in Kidney Cancer
- CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography
- Task-Aware Bitrate Thresholds for Codec-Robust Ultrasound Segmentation