Research

Compressive sensing rests on a simple observation with far-reaching consequences: a signal with few significant components can be recovered from far fewer measurements than the Nyquist rate demands — provided the measurements are designed correctly. Our work in this area established what that means for real sensing hardware, and what happens when a physical scene refuses to line up with the discrete models the theory assumes.

Subsurface imaging with ground penetrating radar

A stepped-frequency GPR normally acquires a complete frequency sweep at every scan position, which makes surveys slow and the data volume large. Because buried targets occupy only a small part of the imaged volume, the scene is sparse — and we showed that a small, randomly chosen subset of the frequency steps is enough to reconstruct it. Acquisition time and data volume drop together, without the resolution loss that uniform undersampling would cause. The same formulation extends to imaging from multiple scan positions and to direct reconstruction of the subsurface without a separate migration step.

Off-the-grid targets and basis mismatch

Sparse recovery methods represent a scene on a discrete grid of candidate positions, delays, or angles. Real targets rarely sit exactly on that grid, and the mismatch is not benign: a single off-grid target spreads energy across several neighbouring cells, degrading both resolution and amplitude estimates. We addressed this by letting the reconstruction adjust the grid itself — estimating perturbations of the dictionary parameters jointly with the sparse coefficients.

Perturbed Orthogonal Matching Pursuit formalized this idea and gave conditions under which off-grid components are recovered accurately. We then carried it across sensing modalities: GPR imaging of off-grid buried targets, robust reconstruction of sparse radar scenes, and stretch processing for wideband radar.

The same difficulty appears in wireless communications. In millimetre-wave systems the channel is sparse in the angular domain, but the true angles of departure and arrival do not fall on the quantized beam grid used for training. We developed off-grid aware channel and spatial covariance estimation for mmWave networks, estimating angular perturbations alongside the sparse channel coefficients — which improves beam alignment and covariance-based precoding compared with methods that assume the grid is exact.

Compressive SAR imaging

Synthetic aperture radar delivers high resolution, but reconstruction is an inverse problem that degrades in two ways at once: when the phase history is undersampled, and when the platform trajectory is not known precisely. We developed sparsity-based SAR reconstruction that addresses both together. An expectation-maximization based matching pursuit formulation reconstructs images from limited phase-history data, and its off-grid extension estimates scatterer positions continuously instead of snapping them onto a discretized image grid.

Building on that, we treated autofocus and reconstruction as one problem: phase errors caused by platform position uncertainty and the position errors caused by off-grid scattering are estimated jointly, so a sparse spotlight SAR scene is focused and reconstructed in a single optimization rather than in two mismatched steps. The result is sharper imagery from fewer measurements, without the artefacts that appear when autofocus is applied to an already-degraded sparse reconstruction.

Sparsity in array processing

Casting bearing estimation as a spatial sparsity problem allows direction-of-arrival estimation from compressive measurements, resolving closely spaced sources with fewer samples than conventional subspace methods require — an early demonstration that the sparsity prior is as powerful in the angular domain as in range or image domains.

Compressive acquisition also trades measurement count against computation, and that trade is not always favourable. We analysed the energy efficiency of compressive sensing in wireless sensor networks and characterized the regimes — network size, sparsity level, radio and processing cost — in which compressive acquisition genuinely reduces total energy consumption rather than merely relocating it.

This foundation carries directly into Area 2 — Learning to Sense, where the measurement operator is no longer designed by hand but learned jointly with the reconstruction and the downstream task.

Selected publications

1. A. C. Gurbuz, J. H. McClellan, and W. R. Scott, “A Compressive Sensing Data Acquisition and Imaging Method for Stepped-Frequency GPRs,” IEEE Transactions on Signal Processing, vol. 57, no. 7, pp. 2640–2650, July 2009.
2. A. C. Gurbuz, J. H. McClellan, and W. R. Scott, “Compressive Sensing for Subsurface Imaging using Ground Penetrating Radar,” Signal Processing, vol. 89, no. 10, pp. 1959–1972, October 2009. [2013 EURASIP Best Paper Award for the Signal Processing Journal]
3. A. C. Gurbuz, V. Cevher, and J. H. McClellan, “Bearing Estimation via Spatial Sparsity Using Compressive Sensing,” IEEE Transactions on Aerospace and Electronic Systems, vol. 48, no. 2, pp. 1358–1369, April 2012.
4. O. Teke, A. C. Gurbuz, and O. Arikan, “Perturbed Orthogonal Matching Pursuit,” IEEE Transactions on Signal Processing, vol. 61, no. 24, pp. 6220–6231, December 2013.
5. O. Teke, A. C. Gurbuz, and O. Arikan, “A Robust Compressive Sensing Based Technique for Reconstruction of Sparse Radar Scenes,” Digital Signal Processing, vol. 27, pp. 23–32, 2014.
6. A. C. Gurbuz, O. Teke, and O. Arikan, “Sparse Ground Penetrating Radar Imaging Method for off-the-grid Target Problem,” Journal of Electronic Imaging, vol. 22, no. 2, 2013.
7. I. Ilhan, A. C. Gurbuz, and O. Arikan, “Compressive Sensing based Robust Off-the-Grid Stretch Processing,” IET Radar, Sonar & Navigation, vol. 11, pp. 1730–1735, 2017.
8. S. Camlica, A. C. Gurbuz, and O. Arikan, “Autofocused Spotlight SAR Image Reconstruction of Off-Grid Sparse Scenes,” IEEE Transactions on Aerospace and Electronic Systems, vol. 53, no. 4, pp. 1880–1892, August 2017.
9. S. Ugur, O. Arikan, and A. C. Gurbuz, “SAR Image Reconstruction by Expectation Maximization Based Matching Pursuit,” Digital Signal Processing, vol. 37, pp. 75–84, 2015.
10. S. Camlica, A. C. Gurbuz, and O. Arikan, “SAR image reconstruction with joint off-grid target and phase error corrections,” IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Milan, 2015, pp. 4502–4505.
11. S. Camlica, H. E. Guven, A. C. Gurbuz, and O. Arikan, “Analysis of sparsity based joint SAR image reconstruction and autofocus techniques,” 3rd International Workshop on Compressed Sensing Theory and its Applications to Radar, Sonar and Remote Sensing (CoSeRa), Pisa, 2015, pp. 99–103.
12. S. Ugur, O. Arikan, and A. C. Gurbuz, “Off-grid sparse SAR image reconstruction by EMMP algorithm,” IEEE Radar Conference (RadarCon), Ottawa, 2013.
13. C. K. Anjinappa, A. C. Gurbuz, Y. Yapici, and I. Guvenc, “Off-Grid Aware Channel and Covariance Estimation in mmWave Networks,” IEEE Transactions on Communications, vol. 68, no. 6, pp. 3908–3921, June 2020.
14. C. K. Anjinappa, A. C. Gurbuz, Y. Yapici, and I. Guvenc, “Off-Grid Aware Spatial Covariance Estimation in mmWave Communications,” Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA, 2019.
15. C. Karakus, A. C. Gurbuz, and B. Tavli, “Analysis of Energy Efficiency of Compressive Sensing in Wireless Sensor Networks,” IEEE Sensors Journal, vol. 13, no. 5, pp. 1999–2008, May 2013.

Classical sensing pipelines are built in isolation: a sensor is designed, then a reconstruction algorithm is written for it, then an analysis step is bolted on top. Each block is optimized for its own criterion, and information discarded early can never be recovered later. This area replaces that chain with a single learnable pipeline — the measurement operator, the reconstruction, and the decision are trained together, against the task that actually matters.

Joint learning of measurement and reconstruction

The starting question is direct: for a given class of signals, what is the best set of measurements to take? Rather than choosing a random or fixed sensing matrix, we learn the measurement operator jointly with a reconstruction network, so the two adapt to each other and to the signal class. Learned operators consistently outperform random projections at the same measurement budget, and the gap widens as the budget shrinks.

Because real hardware cannot implement arbitrary matrices, a central part of this work is learning under physical constraints — binary or quantized entries, limited dynamic range, restricted patterns — so that what the network learns is something a sensor can actually do. We also showed how a single trained network can serve a whole range of measurement rates, avoiding the usual need to train and store one model per sensing budget.

Learning in Compressed Domain

If the end goal is a decision rather than an image, reconstruction is an expensive detour. We learn measurements optimized for the task itself — classification, detection, parameter estimation — and perform inference directly on the compressed data. The measurement operator is shaped by what discriminates the classes, not by what reproduces every pixel, so accuracy at very low measurement rates is far better than reconstructing first and classifying afterwards.

Choosing what to measure: bands, antennas and array elements

The same principle applies when the sensor cannot be redesigned but its configuration can be chosen. In hyperspectral imaging, acquiring hundreds of contiguous bands is expensive in acquisition time, storage and bandwidth — yet only a few carry the information a classifier needs. We formulate band selection as a learning problem, training the selection jointly with the classifier so that the chosen bands are the ones that actually separate the classes, and reaching full-cube accuracy with a small fraction of the bands.

In array processing, the analogous question is which antenna elements to keep. We learn sparse element selections that preserve direction-of-arrival accuracy while reducing the number of active RF chains — a direct saving in hardware cost and power. Related work estimates the number of sources present directly with a network, replacing information-theoretic model-order criteria that degrade badly at low sample support or low signal-to-noise ratio.

Computational imaging: learning the sensor itself

In a colour camera, the filter array on the sensor and the demosaicing algorithm that follows it are usually designed separately, and the classical Bayer pattern predates the reconstruction methods now used with it. We optimize both together: the binary colour filter pattern is learned end to end with the demosaicing network, subject to the constraint that the pattern must be physically realizable. The jointly designed pair reconstructs more faithfully than a hand-designed mosaic paired with the same network — a concrete demonstration that the optical front end belongs inside the optimization.

Physics-aware learning for sensing hardware

Learning also accelerates the design of the hardware itself. Computing the excitation coefficients that steer a reconfigurable antenna array normally requires repeated electromagnetic simulation, which is far too slow for real-time or cognitive operation. By embedding the underlying physics into the network structure, we predict beam coefficients for planar and reconfigurable arrays directly, retaining accuracy while cutting computation by orders of magnitude — making adaptive beam control practical inside a sensing loop.

Selected publications

1. R. Mdrafi and A. C. Gurbuz, “Joint Learning of Measurement Matrix and Signal Reconstruction via Deep Learning,” IEEE Transactions on Computational Imaging, vol. 6, pp. 818–829, 2020.
2. C. O. Ayna, B. K. Gunturk, and A. C. Gurbuz, “DREAM-CFA: Joint Learning of Binary Color Filter Array and Demosaicing,” Journal of Electronic Imaging, vol. 34, no. 2, 023063, 2025.
3. G. D. King, M. Asaduzzaman Towfiq, A. C. Gurbuz, and B. A. Cetiner, “Beam Coefficient Prediction for Antenna Arrays Using Physics-Aware Convolutional Neural Networks,” IEEE Access, vol. 12, pp. 176908–176919, 2024.
4. C. O. Ayna, R. Mdrafi, Q. Du, and A. C. Gurbuz, “Learning-Based Optimization of Hyperspectral Band Selection for Classification,” Remote Sensing, vol. 15, no. 18, 4460, 2023.
5. C. O. Ayna and A. C. Gurbuz, “Antenna Selection for Direction of Arrival Estimation with Deep Learning,” IEEE International Radar Conference (RADAR), Atlanta, GA, 2025, pp. 1–6.
6. R. Mdrafi and A. C. Gurbuz, “Learning Measurement for Classification in Compressed Domain,” SPIE Machine Learning from Challenging Data, vol. 13460, 134600D, 2025.
7. C. O. Ayna and A. C. Gurbuz, “A Single Network for Reconstruction with Dynamic Measurement Rates,” 58th Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, 2024, pp. 1328–1332.
8. J. Rogers, J. E. Ball, and A. C. Gurbuz, “Estimating the Number of Sources via Deep Learning,” IEEE Radar Conference (RadarConf), Boston, MA, 2019, pp. 1–5.
9. R. H. Mdrafi and A. C. Gurbuz, “Data Driven Measurement Matrix Learning for Sparse Reconstruction,” IEEE Data Science Workshop (DSW), Minneapolis, MN, 2019, pp. 253–257.
10. R. H. Mdrafi and A. C. Gurbuz, “Learning to Sense and Reconstruct A Class of Signals,” IEEE Radar Conference (RadarConf), Boston, MA, 2019, pp. 1–5.

Related projects

NSF CAREER — Learning to Sense: Joint Learning of Task Oriented Cognitive Sensing with Data Driven Reconstruction and Inference (NSF ECCS, Award #2047771, 2021–2027)

Deep networks applied to radar and RF data are usually treated as black boxes: generic convolutional layers are trained on spectrogram images, and whatever they learn is difficult to relate back to the physics of the signal. This area takes the opposite approach — architectures are built from signal processing primitives, so their parameters carry physical meaning, they need far fewer weights, and their behaviour can be examined rather than trusted.

Parameterized learnable filters

A conventional convolutional layer learns dozens of unconstrained weights per filter, and the resulting frequency response has no interpretable structure. We replace those layers with parameterized filters — sinc, Gaussian and related kernels — where learning adjusts only a centre frequency and a bandwidth. The network therefore learns which parts of the spectrum matter for the task, and that answer can be read directly off the trained filters.

CV-SincNet extends this idea to complex-valued radar returns, learning band-pass filters directly from raw in-phase and quadrature data instead of from pre-computed spectrograms. Preserving phase and skipping the time-frequency transform improves human motion recognition accuracy while cutting computation, since the expensive front-end transform disappears from the pipeline. PLFNets generalize the family to complex-valued parameterized learnable filters for RF classification more broadly, reaching comparable or better accuracy than deep CNN baselines with orders of magnitude fewer parameters — which matters for embedded and real-time RF systems.

The same principle carries into waveform analysis: learnable Gaussian filter front ends perform automatic modulation and waveform recognition on raw RF data, with the learned filter banks aligning with the occupied bands of the signals they are asked to separate.

High-resolution micro-Doppler reconstruction

Micro-Doppler signatures — the time-varying Doppler shifts produced by limbs, rotors or vibrating structures — are the raw material for radar-based activity recognition. Extracting them means choosing a window length, and that choice is a compromise: long windows resolve frequency but blur fast transitions, short windows track transitions but smear the spectrum. Neither setting shows the full picture, and classification accuracy suffers accordingly.

HRSpecNET treats this as a reconstruction problem rather than a parameter-tuning problem. A deep model learns to produce a high-resolution micro-Doppler signature directly from raw complex IQ observations, recovering detail in both time and frequency simultaneously. The reconstructed signatures are not merely sharper to look at: downstream human activity recognition accuracy improves, and the same approach applies to high-resolution frequency estimation for sparse radar range profiles.

Physics- and kinematics-driven training data

Radar data sets for human motion are small, expensive to collect, and awkward to share. Rather than accept that limit, we generate training data from motion models. Kinect skeletal recordings drive an electromagnetic model that synthesizes realistic micro-Doppler signatures, so a large labelled corpus can be produced from a modest number of real recordings.

Purely synthetic data eventually diverges from reality, so we also constrain generative models with domain knowledge: adversarial synthetic data generation for sign language recognition is conditioned on kinematics and signing fluency, so that generated samples remain physically and linguistically plausible instead of merely fooling a discriminator. Earlier work in the same spirit exploited prior knowledge of human motion to improve micro-Doppler classification when training data are scarce.

RF sensing for American Sign Language recognition

Automatic recognition of American Sign Language is usually attempted with cameras, which raises privacy concerns and fails in poor lighting. RF sensors avoid both problems: they are non-contact, work in the dark, and capture the kinematics of hands and arms without recording an image of the person. In a long-running collaboration spanning engineering, linguistics and human-computer interaction, we developed RF-based ASL recognition from the sensing hardware upward.

Different RF sensors see different things — an FMCW radar resolves range and Doppler, a CW radar captures fine micro-Doppler, an ultra-wideband impulse radar gives fine range resolution — so we fuse multiple frequencies and waveforms rather than relying on one. This fusion is what makes word-level recognition of fluent signing feasible: signing in natural conversation is continuous and co-articulated, quite unlike the isolated, deliberately produced signs that most datasets contain.

A recurring finding is that linguistic structure matters as much as signal processing. Models trained only on isolated signs degrade sharply on fluent signing, and synthetic data that ignores signing fluency transfers poorly. Building those constraints into data generation and model design is what closes the gap.

More detail about radar based ASL recognition please check out our project website at here.

From gestures to traffic and automotive applications

The same time-frequency machinery transfers directly to the road: classifying traffic signalling motions — the hand signals of a police officer or road worker — from automotive FMCW radar, so that an autonomous vehicle can interpret human direction from the sensor it already carries, at night and in weather that defeats cameras. The same multi-modal fusion methods — combining radar, LiDAR and camera data — underpin our work on off-road mobility assessment, subterranean sensing and threat detection, and disaster damage assessment for defence and infrastructure sponsors.

Selected publications

1. S. Biswas, C. O. Ayna, S. Z. Gurbuz, and A. C. Gurbuz, “CV-SincNet: Learning Complex Sinc Filters From Raw Radar Data for Computationally Efficient Human Motion Recognition,” IEEE Transactions on Radar Systems, vol. 1, 2023.
2. S. Biswas, C. O. Ayna, and A. C. Gurbuz, “PLFNets: Interpretable Complex-Valued Parameterized Learnable Filters for Computationally Efficient RF Classification,” IEEE Transactions on Radar Systems, vol. 2, 2024.
3. B. Hicks, S. Biswas, and A. C. Gurbuz, “Learnable Gaussian Filter-Based Automatic RF Waveform Modulation Recognition,” IEEE Radar Conference, 2025.
4. S. Biswas, A. Manavi Alam, and A. C. Gurbuz, “HRSpecNET: A Deep Learning-Based High-Resolution Radar Micro-Doppler Signature Reconstruction for Improved HAR Classification,” IEEE Transactions on Radar Systems, vol. 2, 2024.
5. M. M. Rahman, E. A. Malaia, A. C. Gurbuz, D. J. Griffin, C. Crawford, and S. Z. Gurbuz, “Effect of Kinematics and Fluency in Adversarial Synthetic Data Generation for ASL Recognition With RF Sensors,” IEEE Transactions on Aerospace and Electronic Systems, vol. 58, no. 4, pp. 2732–2745, August 2022.
6. S. Z. Gurbuz, M. M. Rahman, E. Kurtoglu, A. C. Gurbuz, E. A. Malaia, D. J. Griffin, and C. Crawford, “Multi-Frequency RF Sensor Fusion for Word-Level Fluent ASL Recognition,” IEEE Sensors Journal, vol. 22, no. 12, pp. 11373–11381, June 2022.
7. S. Z. Gurbuz, A. C. Gurbuz, E. A. Malaia, D. J. Griffin, C. Crawford, M. M. Rahman, E. Kurtoglu, R. Aksu, T. Macks, and R. Mdrafi, “American Sign Language Recognition Using RF Sensing,” IEEE Sensors Journal, vol. 21, no. 3, pp. 3763–3775, February 2021.
8. C. Karabacak, S. Z. Gurbuz, A. C. Gurbuz, M. B. Guldogan, G. Hendeby, and F. Gustafsson, “Knowledge Exploitation for Human Micro-Doppler Classification,” IEEE Geoscience and Remote Sensing Letters, vol. 12, no. 10, pp. 2125–2129, October 2015.
9. B. Erol, C. Karabacak, S. Z. Gurbuz, and A. C. Gurbuz, “Simulation of Human Micro-Doppler Signatures with Kinect Sensor,” IEEE Radar Conference, Cincinnati, OH, 2014, pp. 863–868.
10. S. Biswas, B. Bartlett, J. E. Ball, and A. C. Gurbuz, “Classification of Traffic Signaling Motion in Automotive Applications Using FMCW Radar,” IEEE Radar Conference (RadarConf23), San Antonio, TX, 2023, pp. 1–6.

Related projects

• RF-SHIELD: AI-Driven RF Sensemaking — Laboratory for Analytic Sciences (LAS), NC State · PI · 2026
• CAREER: Learning to Sense — NSF ECCS, Award #2047771 · PI · 2021–2027
• CPS Small: RF Sensing for Sign Language Driven Smart Environments — NSF CPS · PI · 2019–2022
• Interpretable Complex Sinc-Nets for RF Waveform Detection and Classification with SDR Implementation — AFRL · PI · 2023–2025

Radar and array signal processing is where the lab builds, measures and tests. The work spans the full chain — designing the experiment, collecting real data, and developing the estimation, imaging and detection algorithms that turn returns into answers — with both model-based methods and, increasingly, learned ones.

Ground penetrating radar, landmine detection and subsurface imaging

Detecting buried objects is hard for a reason: the ground reflection is far stronger than anything beneath it, soil properties vary from site to site, and clutter looks a great deal like a target. Our work began with multistatic GPR — using several transmitter and receiver positions rather than a single co-located pair, so that a buried object is observed from many angles at once. Experiments with a purpose-built multistatic system showed how much more can be extracted from the same survey when the geometry is exploited properly.

Subsurface images contain more than isolated targets. Pipes, cables, walls and voids appear as linear and planar structures, and separating them from point-like objects is what makes an image interpretable. We developed an iterative dimension-reduction approach that detects these structures directly in the 3-D image volume, progressively peeling off lower-dimensional components.

Compressive sensing later reduced the acquisition burden itself, letting a stepped-frequency GPR reconstruct the subsurface from a small random subset of frequency measurements — covered in more detail under Area 1. Applying it in the field required one more step: the air–ground interface reflection is so dominant that it consumes the sparse budget before any target is found, so we developed ground reflection removal designed to operate inside the compressive sensing framework rather than as a separate pre-processing stage.

SAR imaging

Synthetic aperture radar turns platform motion into resolution, but the reconstruction degrades when the phase history is undersampled and when the trajectory is imperfectly known. We developed expectation-maximization based matching pursuit for SAR reconstruction from limited data, and extended it so that scatterer positions are estimated continuously rather than forced onto an image grid. Autofocus is then handled jointly with reconstruction: platform-induced phase errors and off-grid scattering are resolved in a single optimization, producing focused imagery from fewer measurements. The off-grid machinery behind this is described under Area 1.

Direction of arrival estimation

Estimating where signals come from is a classical array problem, and one where structure pays. Treating the angular scene as sparse turns bearing estimation into a compressive recovery problem, resolving closely spaced sources from fewer snapshots than conventional subspace methods need.

A second line concerns multifunctional reconfigurable antennas, whose radiation pattern can be switched electronically among many modes. Which modes an array should use is then a design question with a precise answer: we derived Cramér–Rao lower bounds for DOA estimation as a function of the selected mode combination, and used them to choose modes that minimize the achievable estimation error — turning a hardware capability into a measurable accuracy gain.

AI/ML for radar

Learning enters radar most usefully where classical estimators are brittle. Sample covariance matrices, for instance, are poorly conditioned when snapshots are scarce — exactly the regime of fast-moving or agile scenarios. We learn covariance estimation and covariance reconstruction directly from data, and feed the result into DOA estimation, recovering accuracy where sample-starved classical estimators break down. Related work learns which antenna elements to keep, so that estimation accuracy is maintained with fewer active RF chains.

On the hardware side, physics-aware networks predict the excitation coefficients needed to form a desired beam, replacing repeated electromagnetic simulation with a fast forward pass and making adaptive beam control feasible inside a real-time loop.

These pieces come together in cognitive radar: a system that adapts what it transmits and how it listens based on what it has already learned about the scene. With multifunctional reconfigurable antennas providing the degrees of freedom, target detection and tracking performance improves as the radar reallocates its resources dwell by dwell rather than following a fixed schedule.

Radar work described elsewhere on this page

Several radar topics sit more naturally under other areas, and are covered there:

• Micro-Doppler analysis, human activity and gesture recognition, learnable filter architectures — Area 3
• Passive radar, signals of opportunity and GNSS reflectometry — Area 7
• Integrated sensing and communications, spectrum sharing and radar–communication coexistence — Area 5
• Automatic waveform and modulation recognition, RF anomaly detection — Area 6
• Compressive sensing, off-grid estimation and sparse reconstruction foundations — Area 1

Selected publications

1. T. Counts, A. C. Gurbuz, W. R. Scott Jr., J. H. McClellan, and K. Kim, “Multistatic Ground-Penetrating Radar Experiments,” IEEE Transactions on Geoscience and Remote Sensing, vol. 45, no. 8, pp. 2544–2553, August 2007.
2. A. C. Gurbuz, J. H. McClellan, and W. R. Scott, “Detection of Linear and Planar Structures in 3D Subsurface Images by Iterative Dimension Reduction,” Digital Signal Processing, vol. 20, no. 2, pp. 391–400, March 2010.
3. A. C. Gurbuz, J. H. McClellan, and W. R. Scott, “A Compressive Sensing Data Acquisition and Imaging Method for Stepped-Frequency GPRs,” IEEE Transactions on Signal Processing, vol. 57, no. 7, pp. 2640–2650, July 2009.
4. A. C. Gurbuz, J. H. McClellan, and W. R. Scott, “Compressive Sensing for Subsurface Imaging using Ground Penetrating Radar,” Signal Processing, vol. 89, no. 10, pp. 1959–1972, October 2009. [2013 EURASIP Best Paper Award]
5. M. A. C. Tuncer and A. C. Gurbuz, “Ground Reflection Removal in Compressive Sensing Ground Penetrating Radars,” IEEE Geoscience and Remote Sensing Letters, vol. 9, no. 1, pp. 23–27, January 2012.
6. S. Ugur, O. Arikan, and A. C. Gurbuz, “SAR Image Reconstruction by Expectation Maximization Based Matching Pursuit,” Digital Signal Processing, vol. 37, pp. 75–84, February 2015.
7. S. Camlica, A. C. Gurbuz, and O. Arikan, “Autofocused Spotlight SAR Image Reconstruction of Off-Grid Sparse Scenes,” IEEE Transactions on Aerospace and Electronic Systems, vol. 53, no. 4, pp. 1880–1892, August 2017.
8. A. C. Gurbuz, V. Cevher, and J. H. McClellan, “Bearing Estimation via Spatial Sparsity Using Compressive Sensing,” IEEE Transactions on Aerospace and Electronic Systems, vol. 48, no. 2, pp. 1358–1369, April 2012.
9. A. C. Gurbuz and B. A. Cetiner, “CRLB Based Mode Selection and Enhanced DOA Estimation for Multifunctional Reconfigurable Arrays,” Physical Communication, vol. 38, 100894, 2020.
10. A. C. Gurbuz, R. Mdrafi, and B. A. Cetiner, “Cognitive Radar Target Detection and Tracking With Multifunctional Reconfigurable Antennas,” IEEE Aerospace and Electronic Systems Magazine, vol. 35, no. 6, pp. 64–76, June 2020.
11. A. Manavi Alam, C. O. Ayna, S. Biswas, J. T. Rogers, J. E. Ball, and A. C. Gurbuz, “Deep Learning-Based Direction-of-Arrival Estimation with Covariance Reconstruction,” IEEE Radar Conference, 2024.
12. J. T. Rogers, J. E. Ball, and A. C. Gurbuz, “Data-Driven Covariance Estimation,” IEEE International Symposium on Phased Array Systems & Technology, 2022.
13. C. O. Ayna and A. C. Gurbuz, “Antenna Selection for Direction of Arrival Estimation with Deep Learning,” IEEE International Radar Conference, Atlanta, GA, 2025.
14. G. D. King, M. Asaduzzaman Towfiq, A. C. Gurbuz, and B. A. Cetiner, “Beam Coefficient Prediction for Antenna Arrays Using Physics-Aware Convolutional Neural Networks,” IEEE Access, vol. 12, pp. 176908–176919, 2024.

Related projects

• CAREER: Learning to Sense — NSF ECCS, Award #2047771 · PI · 2021–2027
• Interpretable Complex Sinc-Nets for RF Waveform Detection and Classification with SDR Implementation — AFRL · PI · 2023–2025
• Detecting Biological and Chemical Threats in Complex Subterranean Environments — DoD / ERDC · PI · 2021–2024
• UAS-Based Site Characterization — DoD · PI · 2018–2020

Spectrum is finite, and demand for it is not. Communication systems, radars and scientific instruments are being pushed into the same bands, and the traditional answer — give each service its own exclusive allocation — no longer scales. This area works on the two halves of the resulting problem: letting active and passive users share spectrum without ruining each other’s measurements, and designing systems that sense and communicate with the same waveform.

Active–passive coexistence and RFI mitigation

Passive microwave radiometers measure faint natural emission to retrieve soil moisture, sea-surface salinity and atmospheric state. They cannot transmit, cannot repeat a measurement, and cannot move away from an interferer. When a nearby active system radiates into their band, the observation is corrupted — and because the science signal itself is weak and noise-like, conventional threshold detectors either miss subtle interference or discard large amounts of usable data.

We treat detection and mitigation as a learning problem in the time-frequency domain, where different interference types — narrowband carriers, pulsed emissions, chirps, wideband bursts — have distinct signatures. Deep models localize interference at fine resolution and excise only the affected cells, so far more of the observation survives than with blanket flagging. Applied to NASA SMAP radiometer data, learned detection recovers measurements that conventional detectors discard, and the same time-frequency approach extends to high-resolution suppression in passive systems generally.

Instrument calibration is the companion problem: radiometer calibration drifts, and the reference information available on orbit is limited. We developed learning-based calibration that works from reduced reference information together with two-dimensional spectral features, tracking instrument behaviour over time and improving the quality of the brightness temperatures that everything downstream depends on.

Because coexistence research needs data that does not yet exist, we also built a physical testbed in which wireless communication systems and passive sensing operate together under controlled, repeatable conditions, and released the resulting labelled dataset openly so that other groups can benchmark detection and mitigation methods against realistic measurements rather than simulation alone.

ISAC-based target detection and classification

If a communication network already floods an area with signal, that signal can also be used to see. Integrated sensing and communications turns the cellular waveform into an illuminator: reflections from moving objects are processed alongside normal traffic, so sensing costs no extra spectrum and no dedicated radar.

Our work targets UAV detection and tracking with 5G NR signals. A two-stage, coarse-to-fine framework first searches a wide volume cheaply, then refines position and velocity estimates only where something was found — a structure that keeps computation tractable while retaining accuracy. Small drones are a demanding test case: low radar cross-section, low altitude, and heavy clutter.

Cellular ISAC alone does not see everything, so we fuse it with passive RF sensing of the UAV’s own control and video links. The two modalities fail in different ways — ISAC struggles when the target is poorly illuminated, passive sensing fails when the drone is not transmitting — and combining them yields detection and tracking that is more robust than either alone.

This work runs on software-defined radio testbeds built in the lab, across four complementary sensing modes:

ISAC with 5G NR PDSCH DMRS

Demodulation reference signals embedded in the 5G downlink are re-used as a sensing waveform, so sensing rides along with normal data transmission. A bistatic testbed built from two USRP X310 software-defined radios — one transmit antenna, three receive antennas, and a rubidium clock for frequency and timing reference — estimates range, Doppler and angle of a UAV target from the channel response measured on the reference signals, with a gimbal-mounted receiver for controlled pointing.

RF drone classification

Drones can also be found from the signals they emit themselves. An omnidirectional antenna feeding a software-defined receiver captures controller uplinks and video downlinks, and learned classifiers identify the drone type and protocol from those emissions. A library of controllers and airframes is used to collect labelled data across models and flight modes, so that classification generalizes beyond a single platform.

Passive ISAC with Zadoff–Chu and SSB signals

Cellular networks continuously broadcast synchronization signals — synchronization signal blocks and Zadoff–Chu sequences — whether or not anyone is transmitting data. Treating these as illuminators of opportunity gives a fully passive sensing mode: two NI USRP-2901 radios disciplined by a rubidium clock provide reference and surveillance channels, and cross-correlation processing recovers range-Doppler information about targets without transmitting anything or requiring cooperation from the network.

Drone Remote ID broadcast detection and classification

Remote ID broadcasts carry a drone’s identity, position and operator information over Wi-Fi or Bluetooth. A B210 software-defined radio captures and decodes these broadcasts, giving positive identification for compliant aircraft — and, just as usefully, a cross-check against the RF and ISAC sensing modes above, which is what reveals aircraft that are present but not broadcasting.

Radar detection in shared bands

Coexistence also runs the other way: radars increasingly operate in bands shared with commercial wireless. In the CBRS band, incumbent radar returns must be detected in the presence of strong LTE and 5G traffic, often at very low signal-to-interference-plus-noise ratio. We use convolutional recurrent architectures that exploit both spectral structure and temporal continuity, detecting radar pulses under interference conditions where energy detection fails — the technical basis for spectrum sharing frameworks that protect incumbents without freezing commercial use.

Selected publications

1. A. Manavi Alam, M. M. Farhad, W. Al-Qwider, A. Owfi, M. Koosha, N. Mastronarde, F. Afghah, V. Marojevic, M. Kurum, and A. C. Gurbuz, “A Physical Testbed and Open Dataset for Passive Sensing and Wireless Communication Spectrum Coexistence,” IEEE Access, vol. 12, 2024.
2. A. Manavi Alam, M. Kurum, and A. C. Gurbuz, “RFI-Net: Enhancing Passive Sensing through Deep Learning Based Time-Frequency Domain Radio Frequency Interference Detection and Mitigation,” IEEE Transactions on Geoscience and Remote Sensing, 2026.
3. A. Manavi Alam, M. Kurum, and A. C. Gurbuz, “Radio Frequency Interference Detection for SMAP Radiometer Using Convolutional Neural Networks,” IEEE JSTARS, vol. 15, 2022.
4. A. M. Alam, M. Kurum, M. Ogut, and A. C. Gurbuz, “Microwave Radiometer Calibration Using Deep Learning With Reduced Reference Information and 2-D Spectral Features,” IEEE JSTARS, vol. 17, pp. 748–765, 2024.
5. A. M. Alam, M. Kurum, and A. C. Gurbuz, “Deep Learning-Based High-Resolution Time-Frequency Domain RFI Suppression in Passive Systems,” IGARSS 2025 — IEEE International Geoscience and Remote Sensing Symposium, Brisbane, Australia, 2025, pp. 404–408.
6. W. Khawaja, A. M. Alam, and A. C. Gurbuz, “A Two-Stage 5G NR ISAC Framework for Coarse-to-Fine UAV Detection and Tracking,” IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), pp. 1–8, 2026.
7. C. Dickerson et al., “Fusion of Cellular ISAC and Passive RF Sensing for UAV Detection and Tracking,” 59th Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, 2025, pp. 1633–1638.
8. A. M. Alam, W. Khawaja, and A. C. Gurbuz, “Interference-Resilient Low-SINR Radar Detection in the CBRS Band Using CRNNs.”

Related projects

• SWIFT-SAT INTERACT: Learning-Based Interference Mitigation for Radiometers — NSF SWIFT-SAT, Award #2332661 · PI · 2024–2026
• Enhancing SMAP Radiometer Performance with Deep Learning — NASA SMAP Science Team, 80NSSC25K7061 · Co-Investigator · 2025–2027
• SWIFT LARGE: AI-Enabled Spectrum Coexistence between Active Communications and Passive Radio Services — NSF · Co-PI · 2020–2024

Related work on waveform and modulation recognition, RF anomaly detection and spectrum sensemaking is described under Area 6; passive radar and signals of opportunity under Area 7.

Turning raw wideband RF observations into an understandable picture: detecting emissions, recognizing waveform and modulation types directly from complex-valued data, flagging signals that match nothing known, and doing all of it with architectures whose decisions can be inspected. Methods are evaluated on recorded and over-the-air data with software-defined radios rather than synthetic signals alone.

Representative work

• Automatic Waveform Recognition from Complex RF Data with Filter-based Deep Learning, NAECON, 2024.
• Automatic classification of radar and communication waveforms through interpretable deep learning, SPIE, 2025.
• PLFNets: Interpretable Complex-Valued Parameterized Learnable Filters for RF Classification, IEEE Transactions on Radar Systems, 2024.

Related project: RF-SHIELD — AI-Driven RF Sensemaking

Navigation and communication satellites already blanket the Earth with signals. Signals of opportunity sensing puts that existing illumination to scientific use: instead of launching a transmitter, a receiver simply listens to reflections of signals that are already in the air. The reward is a sensing capability with no transmit power, modest hardware, and — with constellations such as GNSS and CYGNSS — revisit times that dedicated missions cannot match.

Machine learning for CYGNSS soil moisture retrieval

NASA’s CYGNSS constellation was designed to measure ocean winds, but its reflected GPS signals also carry a strong soil moisture signature over land. Extracting it is difficult: the reflection depends not only on moisture but on vegetation, surface roughness, terrain and soil texture, and the specular points fall along irregular tracks rather than a regular grid. Physically-based inversion struggles with this tangle of confounders.

We approached the problem as a learning task, training retrieval models against dense in-situ networks and satellite products. Starting from neural-network retrievals at high spatial and temporal resolution, the work grew into systematic evaluation — against ISMN stations across the continental United States, against SMAP observations, and eventually at quasi-global scale, where a machine-learning retrieval combining CYGNSS and SMAP produced soil moisture at finer space-time scales than either source alone.

Later work moved from engineered features to the raw observations themselves, learning retrievals directly from delay–Doppler maps, and assessed how well those deep retrievals generalize across climates and land cover worldwide. Because no single product is best everywhere, we also developed statistically optimal fusion — best linear unbiased and minimum-variance estimators that merge multiple CYGNSS soil moisture products into a single, better-conditioned estimate — and quantified the interpolation errors introduced when sparse specular tracks are gridded, a subtle but consequential source of apparent retrieval error.

Forward modelling and performance bounds

Learning alone is not enough when the goal is to design a future mission or to know what is physically recoverable. Alongside the retrieval work we develop the physics: SCoBi Multilayer models signals of opportunity reflectometry from layered dielectric media, capturing how a stratified soil column — rather than a single homogeneous half-space — shapes the reflected signal.

That model in turn enables a sharper question: how much information about root-zone moisture, well below the surface, is actually present in a reflectometry measurement? We derived Cramér–Rao lower bounds for root-zone soil moisture estimation from SoOp reflectometry, establishing the accuracy achievable as a function of frequency, polarization and observation geometry — guidance that shapes instrument design rather than merely evaluating an algorithm after the fact.

UAS-based GNSS-R: low-cost and ubiquitous sensing

Spaceborne reflectometry averages over kilometre-scale footprints. Farms, fields and small watersheds need much finer detail, which points to small unmanned aircraft flying low over the target. We showed that this can be done with strikingly ordinary hardware: an ordinary smartphone integrated into a small UAS can receive reflected GPS signals and sense water in soil, turning a commodity device into a remote sensing instrument.

The idea was then generalized into a ubiquitous GNSS-R methodology in which a spinning smartphone aboard a small UAS estimates surface reflectivity — the rotation providing the antenna pattern diversity that a fixed, low-gain phone antenna cannot supply on its own.

Reflectometry alone cannot separate every effect, so our UAS campaigns fly multiple sensors together. A three-year field campaign integrated UAS-based GNSS-R with LiDAR and multispectral imaging, using structure and canopy information from the optical sensors to disentangle vegetation effects from the moisture signal — an approach validated over multiple growing seasons rather than a single flight day.

Vegetation sensing with GNSS transmissometry

Pointing the same idea upward turns reflectometry into transmissometry: measuring how much a canopy attenuates GNSS signals passing through it yields vegetation optical depth, a direct measure of canopy water content and biomass. We developed a deep learning framework that combines airborne LiDAR structure with mobile GNSS-T measurements to map vegetation optical depth at high resolution over large areas — scaling a point measurement into a landscape-level product.

Selected publications

1. O. Eroglu, M. Kurum, D. Boyd, and A. C. Gurbuz, “High Spatio-Temporal Resolution CYGNSS Soil Moisture Estimates Using Artificial Neural Networks,” Remote Sensing, vol. 11, no. 19, 2019.
2. V. Senyurek, F. Lei, D. Boyd, M. Kurum, A. C. Gurbuz, and R. Moorhead, “Machine Learning-Based CYGNSS Soil Moisture Estimates over ISMN Sites in CONUS,” Remote Sensing, vol. 12, no. 7, 1168, 2020.
3. V. Senyurek, F. Lei, D. Boyd, A. C. Gurbuz, M. Kurum, and R. Moorhead, “Evaluations of a Machine Learning-Based CYGNSS Soil Moisture Estimates against SMAP Observations,” Remote Sensing, vol. 12, no. 21, 3503, 2020.
4. D. Boyd, M. Kurum, O. Eroglu, A. C. Gurbuz, J. L. Garrison, B. R. Nold, M. A. Vega, J. R. Piepmeier, and R. Bindlish, “SCoBi Multilayer: A Signals of Opportunity Reflectometry Model for Multilayer Dielectric Reflections,” Remote Sensing, vol. 12, no. 21, 3480, 2020.
5. D. R. Boyd, A. C. Gurbuz, M. Kurum, J. L. Garrison, B. R. Nold, M. A. Vega, J. R. Piepmeier, and R. Bindlish, “Cramér–Rao Lower Bound for SoOp-R-Based Root-Zone Soil Moisture Remote Sensing,” IEEE JSTARS, vol. 13, pp. 6101–6114, 2020.
6. M. Kurum, M. M. Farhad, and A. C. Gurbuz, “Integration of Smartphones into Small Unmanned Aircraft Systems to Sense Water in Soil by Using Reflected GPS Signals,” IEEE JSTARS, vol. 14, pp. 1048–1059, 2021.
7. V. Senyurek, A. C. Gurbuz, and M. Kurum, “Assessment of Interpolation Errors of CYGNSS Soil Moisture Estimations,” IEEE JSTARS, vol. 14, pp. 9815–9825, 2021.
8. F. Lei, V. Senyurek, M. Kurum, A. C. Gurbuz, D. R. Boyd, R. Moorhead, W. T. Crow, and O. Eroglu, “A Quasi-Global Machine Learning-Based Soil Moisture at High Spatio-Temporal Scales using CYGNSS and SMAP Observations,” Remote Sensing of Environment, vol. 276, 113041, 2022.
9. M. M. Nabi, V. Senyurek, A. C. Gurbuz, and M. Kurum, “Deep Learning-Based Soil Moisture Retrieval in CONUS Using CYGNSS Delay–Doppler Maps,” IEEE JSTARS, vol. 15, pp. 6867–6881, 2022.
10. M. M. Farhad, M. Kurum, and A. C. Gurbuz, “A Ubiquitous GNSS-R Methodology to Estimate Surface Reflectivity Using a Spinning Smartphone Onboard a Small UAS,” IEEE JSTARS, vol. 16, pp. 6568–6578, 2023.
11. M. M. Nabi, V. Senyurek, F. Lei, M. Kurum, and A. C. Gurbuz, “Quasi-Global Assessment of Deep Learning-Based CYGNSS Soil Moisture Retrieval,” IEEE JSTARS, vol. 16, pp. 5629–5644, 2023.
12. M. M. Nabi, V. Senyurek, M. Kurum, and A. C. Gurbuz, “Best Linear Unbiased Estimators for Fusion of Multiple CYGNSS Soil Moisture Products,” IEEE JSTARS, 2024.
13. E. Hodges, C. Chew, E. E. Small, D. Bai, M. Al-Khaldi, J. D. Ouellette, J. T. Johnson, et al., “A Merged CYGNSS Soil Moisture Product Using a Minimum Variance Estimator,” IEEE Transactions on Geoscience and Remote Sensing, 2025.
14. M. M. Farhad et al., “Integrating UAS-Based GNSS-R, LiDAR, and Multispectral Data for Soil Moisture Estimation: Summary of Results From a Three-Year-Long Field Campaign,” IEEE JSTARS, vol. 18, pp. 16896–16915, 2025.
15. A. Ghosh, M. E. Hoque, M. M. Farhad, A. C. Gurbuz, A. Peduzzi, and M. Kurum, “Deep Learning Framework for High-Resolution Large-Scale Vegetation Optical Depth Mapping Using Airborne LiDAR and Mobile GNSS-T Data,” IEEE JSTARS, vol. 19, pp. 12087–12100, 2026.

Related projects

• Deep Learning Based High-Resolution Field-Level Soil Moisture Mapper from UAVs — USDA AI Center of Excellence · Co-PI · 2024–2025
• Advancement of UAS/UAV Application Systems — USDA-ARS · Investigator · 2019–2023
• UAV-Based Autonomous Unsupervised Weed Detection for Corn Fields — Mississippi Corn Promotion Board · Co-PI · 2023–2025

Methane escapes from the seafloor through gas seeps — plumes of bubbles that rise through the water column and carry consequences for climate, ocean chemistry, geohazards and energy resources. Multibeam echosounders map these plumes acoustically over enormous areas, but the resulting water-column data is where the difficulty starts: a single survey produces terabytes of imagery, seeps occupy a vanishing fraction of it, and identification has traditionally required an expert to inspect the data by eye.

We develop machine learning that automates this search, so that survey data can be screened at the rate it is collected rather than months later.

Detecting seeps in water-column sonar imagery

The first task is deciding whether a plume is present in a given water-column image. This is harder than it sounds: bubble plumes are faint, thin and variable, and the imagery is full of look-alikes — fish schools, sidelobe artefacts, noise from the vessel and the sea state. We showed that learned classifiers substantially outperform threshold-based screening on real survey data, and then improved on them with an attention-guided convolutional network that concentrates on the narrow, vertically coherent structures that distinguish a plume from surrounding clutter. Attention also makes the model auditable: the regions it weights can be inspected against what a human analyst would flag.

Using the sequence, not just the frame

Surveys do not produce isolated images — they produce continuous sequences as the vessel moves along a track, and a real seep persists across consecutive frames while artefacts do not. We built sequential processing models that exploit this temporal structure, deciding whether a seep occurs over a span of pings rather than judging each frame independently. Treating detection as a sequence problem cuts false alarms driven by transient noise and recovers weak plumes that no single frame shows convincingly.

From detection to localization and mapping

Knowing that a seep is present is only useful if you know where. Our recent work moves from classification to localization: an object-detection framework locates plumes within the water-column image, and a Hough-transform stage exploits their near-vertical geometry to trace each plume back to its origin on the seafloor. The output is a map of seep locations rather than a list of flagged images — the form scientists actually need for characterizing seep fields and monitoring them over time.

Selected publications

1. S. M. Manjur, V. Senyurek, A. Skarke, and A. C. Gurbuz, “Deep Learning-Based Sequential Processing of Multibeam Echosounder Images for Automated Detection of Seafloor Gas Seep Occurrence,” IEEE JSTARS, vol. 18, pp. 25549–25561, 2025.
2. S. M. Manjur, V. Senyurek, R. Kalski, S. Gupta, A. Skarke, and A. C. Gurbuz, “Automated Detection of Seafloor Gas Seeps in Multibeam Echosounder Data With an Attention-Guided Convolutional Neural Network,” IEEE JSTARS, vol. 18, pp. 5633–5645, 2025.
3. S. M. Manjur, V. Senyurek, R. Kalski, A. Skarke, and A. C. Gurbuz, “Machine Learning-Based Seafloor Gas Seep Detection in Sonar Water Column Images,” SPIE Ocean Sensing and Monitoring XVII, vol. 13482, pp. 83–89, 2025.
4. S. M. Manjur, A. Skarke, and A. C. Gurbuz, “From Detection to Localization: A YOLO and Hough Transform Based Framework for Seafloor Gas Seep Mapping in Multibeam Sonar Data,” IEEE Transactions on Geoscience and Remote Sensing, under review, 2026.

Related projects

• Machine Learning Based Automated Detection of Seafloor Gas Seeps — NOAA Ocean Exploration · Co-PI · 2022–2024