Synthetic Aperture Radar Drone Gets Interferometric Imaging | Hackaday

Synthetic Aperture Radar Drone Gets Interferometric Imaging | Hackaday

Backyard Coherence: A €1,000 Drone SAR Achieves Repeat-Pass Interferometry

A hobbyist C-band FMCW radar on a 7-inch FPV quadcopter now produces sub-meter-accurate elevation maps. The enabling step is a residual-motion estimator built on generalized phase-gradient autofocus, not exotic hardware.

---

BLUF

Finnish engineer Henrik Forstén has shown repeat-pass interferometric SAR (InSAR) from a sub-kilogram FPV quadcopter carrying a self-built ~6 GHz polarimetric FMCW radar. He puts total hardware cost for radar and drone at roughly 1,000 EUR, a small fraction of comparable commercial InSAR systems. The hardware is not for sale, but the processing code is published under the MIT license in his torchbp repository. hforsten

The results are these. Two passes with a 1 m vertical baseline produced mean scene coherence rising from 0.21 without motion correction to about 0.47 with autofocus, a new residual-motion-estimation (RME) algorithm, and a reference DEM. The resulting DEM agreed with national lidar to 0.4 m RMS and 0.11 m median absolute error. Three things made this possible: a €144 RTK/PPK GNSS receiver, fixes to ArduPilot's raw-GNSS logging, and a backprojection-domain RME method that Forstén believes is novel.

The work sits alongside a growing peer-reviewed literature on drone InSAR at L-, Ku-, and K-band. It is notable mainly for cost, for being fully open, and for what it implies about the spread of dual-use SAR capability. No court filings or government releases specific to this project were found.

---

From Snow-Field Demo to Interferometer

Forstén's first airborne system appeared in February 2025. He mounted the radar on the cheapest no-name 7-inch FPV kit he could buy, set a radar budget under 500 EUR, and accepted lossy FR4 PCB material for both electronics and antennas as a result. The carrier sits near 6 GHz, which he chose as the highest band with plentiful inexpensive consumer RF parts. Cheap power amplifiers there reach about 30 dBm. Transmit and receive polarization switches let the radar collect all four of HH, HV, VH, and VV. Homemade polarimetric synthetic aperture radar drone +2

The digital back end uses a Zynq-7020 SoC. The antennas are dual-polarized, slot-fed stacked patches surrounded by a hand-cut sheet-metal horn, with a simulated peak gain of 10 dBi. Without its battery the airframe weighs 752 g. With the smaller 1,300 mAh pack the whole system comes to 948 g. The first demonstration flew a straight 500 m track at 110 m altitude and 5 m/s, transmitting VV only with a 400 µs sweep, 500 MHz of bandwidth, and 1 kHz PRF. Homemade polarimetric synthetic aperture radar drone +2

That first system relied on consumer GNSS with meter-class error, so image quality depended on autofocus. In October 2025 Forstén described a new pipeline. It combines generalized phase gradient autofocus (GPGA) with a 3-D trajectory-deviation estimator adapted from Ding et al. The image is split into sub-images, and the phase error in each is converted to a range error. Per-pulse 3-D position corrections then come from a linearized, overdetermined least-squares solution weighted by signal-to-clutter ratio. The same post added antenna-pattern normalization and a posteriori polarimetric calibration following Ainsworth, Ferro-Famil, and Lee. Synthetic aperture radar autofocus and calibration +2

Hardware Changes for Interferometry

Precision positioning. Forstén found an F9P-based RTK module on AliExpress for 144 EUR, bundled with a helical antenna. The listing gave no connector pinout, so he traced the pins himself and filed down part of the PCB to clear the antenna. The base station, also F9P-based and sold with a tripod and antenna, cost 180 EUR. He reports about 2 cm accuracy against roughly 1 m for the non-RTK receiver. hforstenhforsten

Because image formation happens offline, he also uses post-processed kinematic (PPK) solutions. He found that ArduPilot's raw-GNSS logging dropped nearly every measurement, because a fixed buffer held only 32 satellite observations and he routinely saw about 80. It also failed to record the signal-band ID needed for carrier-phase solutions. He submitted a fix upstream. In the pull request, an ArduPilot maintainer said the change should be merged while suggesting the configuration parameter be redesigned later. hforstengithub

A second problem came from the radio link. RTCM correction traffic exceeded the capacity of the ExpressLRS MAVLink link, so he wrote his own bandwidth-limited RTCM server. hforsten

Sweep linearity. The chirp comes from an LMX2491 PLL. Vendor PLL simulators model fixed-frequency operation only, so Forstén built, with LLM assistance, a simulator of a sweeping PLL with a MASH delta-sigma modulator and used it to tune the loop filter. The optimizer favored a very narrow loop, about 60 kHz bandwidth with 55° phase margin. That suppressed integer-boundary spurs and a constant-frequency spur, but at the cost of longer relock time between sweeps. hforstenhforsten

Data throughput. The 12-bit, 50 MSPS ADC produces 75 MB/s. The Zynq's SD interface tops out at 25 MB/s in theory and about 20 MB/s in practice. His fix exploits a property of FMCW data: after dechirp, near-range returns are strong and low in frequency, so successive samples change slowly. He delta-encodes adjacent samples, zigzag-maps the result, and uses a three-code nibble-aligned entropy scheme. This gives about 2.3:1 lossless compression, roughly 5.7 bits per sample on the mission data. FLAC compresses better but is impractical to run in real time on the FPGA. The compressor runs in programmable logic at about three samples per clock. With 2× decimation enabled, the stream fits comfortably within SD bandwidth. Interferometric Synthetic Aperture Radar Drone +2

Ground control. Forstén replaced Mission Planner with his own ground station, mostly written with AI coding help. It auto-generates linear, circular, and interferometric SAR missions. All flight-critical logic stays in ArduPilot on the flight controller. hforsten

The InSAR Processing Chain

The geometry is conventional repeat-pass, flown with a multicopter. Both passes follow the same track in the same direction, separated by 1 m in altitude. Flying the same direction avoids differential Doppler phase and lets common timing errors between the GNSS tag and the sweeps cancel. hforsten

The processing chain runs in five steps:

1. Each pass is autofocused independently with GPGA.
2. Each pass is backprojected onto a flat ground plane or a reference DEM.
3. The two images are resampled to a common grid.
4. RME corrects the differential trajectory error between passes.
5. The interferogram is formed, filtered, unwrapped, and converted to height.

Because backprojection onto a ground plane already accounts for the path-length difference between passes, perfectly flat ground produces zero interferometric phase. Classical slant-range processing would instead show a flat-earth phase ramp. hforsten

The RME contribution. The best-known airborne RME technique is multisquint, which has well-established backprojection variants. Examples include Cao et al. in IEEE TGRS (2018) and the BP-MSQ algorithm of Xie et al., which derives an explicit expression for residual motion error in a backprojected image. nih

Forstén took a different route. He extended his GPGA formulation so that it works directly on the per-pulse backprojection terms rather than only on finished images, and says that to his knowledge the method is new. The central problem is that real topography contaminates a naive global estimate. Each sweep illuminates the scene differently, so pixels with different topographic phase partly cancel and leak terrain into the motion estimate. He fixes this by summing over small blocks where topographic phase is roughly constant. He then estimates each block's phase and removes it before solving for the per-sweep motion phase common to all blocks. hforstenhforsten

Pixels are weighted by squared power coherence, which ignores topographic phase and suppresses decorrelated vegetation and poorly lit regions. Blocks are weighted by how consistent their phase history is across sweeps. Cross-track and vertical errors come from a least-squares fit across range bands that are evenly spaced in the sine of elevation angle. Along-track error is solved separately by differencing blocks ahead of and behind the platform. This needs a beam covering both sides, so it would fail with a narrow squinted beam. Unlike autofocus, the method needs no point targets and no iteration, and it actually works better over flat ground. Interferometric Synthetic Aperture Radar Drone +3

Measured Performance

The main scene used 17,000 sweeps imaged onto a 4,156 × 16,473-pixel grid. On an RTX 3090 Ti, autofocus took 22 seconds and fast factorized backprojection took 0.9 seconds per polarization. Ground truth was the National Land Survey of Finland 2 m lidar elevation model. hforsten

The coherence progression shows what each correction stage buys:

• No autofocus or RME: mean coherence 0.210 hforsten
• Autofocus only: 0.297, still very poor at near range hforsten
• Autofocus plus RME: 0.453 including decorrelated forest and poorly lit edges, with near-range ground above 0.9 hforsten
• Autofocus, RME, and reference DEM: 0.468 hforsten

The solved RME was mostly vertical, which is both the axis where RTK GNSS is weakest and the axis left out of autofocus. For radar engineers, the practical point is this: the amplitude images before and after RME look almost the same, yet the interferograms differ dramatically. In other words, a well-focused SAR image says little about whether it will support interferometry. hforstenhforsten

After unwrapping with SNAPHU, the SAR DEM differed from lidar by 0.4 m RMS and 0.11 m median absolute error, counting areas where the scene had genuinely changed. A constant baseline error of (9.2, 24.4, 2.3) mm was estimated by fitting to the reference DEM and removed. Uncorrected, it would have produced several meters of height error at the image edges with a 1 m baseline. hforstenhforsten

A second mission, flown minutes later on a track rotated 90°, reached 0.480 mean coherence, with a smaller baseline error. hforsten

The limitations are standard physics, and Forstén states them plainly:

• Forest decorrelates between passes, shadowed areas return nothing, and smooth roads give little backscatter at these look angles. hforsten
• At a bridge, the deck and the ground beneath share the same range. That breaks the single-surface assumption the interferometry relies on. hforsten
• The 1 m baseline works at close range, but farther ranges would benefit from a longer baseline to increase height sensitivity. hforsten
• Without special permits, flight altitude is capped at 120 m, which forces very shallow grazing angles and long shadows at range. hforsten

Context: The Drone InSAR Research Landscape

Forstén's result follows several years of institutional work across multiple bands.

L-band. Frey and Werner of Gamma Remote Sensing reported repeat-pass interferograms and a first tomographic profile at EUSAR 2021. Their compact FMCW L-band SAR flew on Aeroscout's Scout B1-100 VTOL UAV, with temporal baselines up to 43 days. Gamma now fields the system on a Freefly Alta X quadcopter, specifying 1.2–1.4 GHz operation and 0.75 m range resolution. vde-verlaggamma-rs

Ku-band. Ruiz-Carregal et al. demonstrated repeat-pass interferometry with a dual-channel Ku-band FMCW drone SAR that also has single-pass cross-track capability. They used the single-pass DEM to coregister the repeat-pass stack. Their 2025 IEEE TGRS follow-up addresses the fact that GNSS errors are comparable to wavelengths at X-band and above. They introduced a complex-domain enhancement of multisquint RME that is more robust to large errors and decorrelation. The same group has since published drone-based multi-temporal DInSAR for large 3-D displacement retrieval with daily revisits in IEEE JSTARS (2026). Remote Sensing (Oct 2024) +2

K-band. In IEEE Transactions on Radar Systems (March 2026), Li, Bekar, Martorella, and Antoniou of the University of Birmingham quantified motion-error tolerances for high-frequency UAV InSAR. They showed that autofocus can restore height-estimation performance and validated this with a 24 GHz UAV demonstrator, which they describe as the first UAV InSAR demonstration in that band. birmingham

The Birmingham group's earlier low-cost drone SAR work was, by Forstén's account, part of what motivated his own effort. He noted that the published system was quoted at £15,000, beyond his personal budget. hforsten

Ultra-wideband. Multi-baseline UWB UAV interferometry has been demonstrated experimentally by Mustieles-Perez et al. at EUSAR 2024.

A shared theme runs through all of this work. Whether the method is Gamma's DEM-aided time-domain backprojection, the Ku-band team's multisquint refinements, the Birmingham autofocus analysis, or Forstén's block-wise GPGA estimator, the limiting factor is residual trajectory error at the millimeter level, not radar hardware. Forstén's RME has not been peer reviewed or benchmarked against multisquint on common data. That comparison would be a natural next step.

Dual-Use and Supply-Chain Notes

Two points deserve attention from AES readers.

First, component access. Mouser accepted Forstén's order for an RF component and then declined to ship it because the supplier bars sales to individuals, apparently to keep parts out of defense applications. He substituted an obsolete pin-compatible part rated only to 4 GHz. hforsten

Second, export classification. SAR capability is a named parameter in the multilateral dual-use lists. Wassenaar-derived entry 6A008.d covers radar able to operate in SAR, ISAR, or side-looking airborne modes. In U.S. regulations the corresponding text is ECCN 6A008 in 15 CFR Part 774. Separately, publicly available software is treated differently under the General Software and Technology Notes. Whether and how these controls apply to a non-commercial, unsold hobby build with MIT-licensed processing code is a question for export counsel. This article draws no conclusion on it. thetradehub

The broader policy point stands regardless. Autofocus, RME, and GPU backprojection are now open source. Sub-€200 carrier-phase GNSS is widely available. Together these lower the barrier to coherent change detection and terrain mapping from small UAS. That is a boon for geoscience, infrastructure monitoring, and disaster response, and a factor for counter-UAS and operational-security planners.

---

Sources

Sources [1]–[3], [5]–[16], and [18]–[20] were accessed or confirmed during research for this article. Source [4] is the article supplied by the requester. Source [17] is cited in [1] and was not independently opened.

[1] H. Forstén, "Interferometric synthetic aperture radar drone," Henrik's Blog, Sep. 14, 2026. https://hforsten.com/interferometric-synthetic-aperture-radar-drone.html

[2] H. Forstén, "Homemade polarimetric synthetic aperture radar drone," Henrik's Blog, Feb. 11, 2025. https://hforsten.com/homemade-polarimetric-synthetic-aperture-radar-drone.html

[3] H. Forstén, "Synthetic aperture radar autofocus and calibration," Henrik's Blog, Oct. 7, 2025. https://hforsten.com/synthetic-aperture-radar-autofocus-and-calibration.html

[4] "Synthetic aperture radar drone gets interferometric imaging," Hackaday, Sep. 28, 2026. https://hackaday.com/2026/09/28/synthetic-aperture-radar-drone-gets-interferometric-imaging/

[5] "Budget-minded synthetic aperture radar takes to the skies," Hackaday, Feb. 13, 2025. https://hackaday.com/2025/02/13/budget-minded-synthetic-aperture-radar-takes-to-the-skies/

[6] Ttl (H. Forstén), "AP_GPS: GPS_RAW_DATA fixes for PPK processing," ArduPilot pull request #32395, GitHub. https://github.com/ArduPilot/ardupilot/pull/32395

[7] H. Forstén, torchbp: differentiable GPU SAR image formation and autofocus (MIT license), GitHub. https://github.com/Ttl/torchbp

[8] Y. Li, A. Bekar, M. Martorella, and M. Antoniou, "Unmanned aerial vehicle (UAV)-based, K-band interferometric synthetic aperture radar (SAR)," IEEE Trans. Radar Syst., vol. 4, pp. 535–548, Mar. 2026, doi: 10.1109/TRS.2026.3661483. https://research.birmingham.ac.uk/en/publications/unmanned-aerial-vehicle-uav-based-k-band-interferometric-syntheti/

[9] G. Ruiz-Carregal et al., "Ku-band SAR-Drone system and methodology for repeat-pass interferometry," Remote Sens., vol. 16, no. 21, Art. no. 4069, Oct. 2024, doi: 10.3390/rs16214069. https://www.mdpi.com/2072-4292/16/21/4069

[10] G. Ruiz-Carregal et al., "Accurate residual motion error estimation for high-frequency drone-borne SAR interferometry," IEEE Trans. Geosci. Remote Sens., vol. 63, pp. 1–20, 2025. https://ieeexplore.ieee.org/document/11104077/

[11] G. Ruiz-Carregal et al., "Drone-based MT-DInSAR for high-magnitude 3-D displacement retrieval with daily revisits," IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 19, pp. 850–874, 2026. Bibliographic record: https://dblp.dagstuhl.de/pid/386/1752.html

[12] O. Frey and C. L. Werner, "UAV-borne repeat-pass SAR interferometry and SAR tomography with a compact L-band SAR system," in Proc. EUSAR 2021, 2021, pp. 1–4. https://vde-verlag.de/proceedings-en/455457040.html

[13] O. Frey, C. L. Werner, and R. Coscione, "Car-borne and UAV-borne mobile mapping of surface displacements with a compact repeat-pass interferometric SAR system at L-band," in Proc. IEEE IGARSS, 2019, pp. 274–277. https://www.research-collection.ethz.ch/entities/publication/4fdfccd7-499c-4e71-af51-cee08f08be21

[14] Gamma Remote Sensing, "GAMMA SAR systems information," v1.8, Oct. 14, 2025. https://www.gamma-rs.ch/files/instruments/SAR/GAMMA_SAR_Systems_information.pdf

[15] A. Bekar, M. Antoniou, and C. J. Baker, "Low-cost, high-resolution, drone-borne SAR imaging," IEEE Trans. Geosci. Remote Sens. https://pure-oai.bham.ac.uk/ws/portalfiles/portal/136457382/Final_Version_TGRS.pdf

[16] P. Xie, M. Zhang, L. Zhang, and G. Wang, "Residual motion error correction with backprojection multisquint algorithm for airborne synthetic aperture radar interferometry," Sensors, vol. 19, no. 10, Art. no. 2342, May 2019, doi: 10.3390/s19102342. https://pmc.ncbi.nlm.nih.gov/articles/PMC6567132

[17] Z. Ding et al., "An autofocus approach for UAV-based ultrawideband ultrawidebeam SAR data with frequency-dependent and 2-D space-variant motion errors," IEEE Trans. Geosci. Remote Sens. https://ieeexplore.ieee.org/document/9380507

[18] N. Cao et al., "Estimation of residual motion errors in airborne SAR interferometry based on time-domain backprojection and multisquint techniques," IEEE Trans. Geosci. Remote Sens., vol. 56, no. 4, pp. 2397–2407, Apr. 2018. https://smu.edu/lyle/departments/cee/faculty/-/media/EF97FA1312EA4C7597D128EE31F15C51.ashx

[19] V. Mustieles-Perez et al., "Experimental demonstration of UAV-based ultra-wideband multi-baseline SAR interferometry," in Proc. EUSAR 2024, pp. 1156–1161. Cited in the reference list of: https://arxiv.org/pdf/2507.20792

[20] EU Dual-Use Annex I / Wassenaar Arrangement, entry 6A008 (radar systems), secondary summary: https://www.thetradehub.eu/de/customs/export-control/annex-i/6a008. Authoritative U.S. text: 15 CFR Part 774, Supp. 1, ECCN 6A008.