Rasad (رصد): AI Anti-Drone Monitoring for Qatar's Critical Sites

Team QH-180 "gradient descent" · QAI × Hugging Face Hackathon 2026

Try it: Rasad app · Model: Galal123/team-180-rasad · Dataset: Galal123/rasd-tracks-synthetic

How warfare changed

For decades, air defense assumed the threat was expensive: aircraft and missiles worth millions, flown in small numbers. Cheap one-way attack drones broke that assumption.

The best-known example is the Iranian HESA Shahed 136, a delta-wing loitering munition driven by a small piston engine. Public estimates put it at roughly $20,000–50,000 per unit, with a range of up to 2,000–2,500 km and a warhead of about 30–50 kg (heavier variants are reported). It is not built to be clever. It is built to be numerous: launched in waves to saturate defenses. The idea has spread fast; the U.S. has reportedly fielded its own low-cost copy, LUCAS (Low-Cost Uncrewed Combat Attack System).

That creates two problems for anyone protecting critical sites:

1. Cost asymmetry. Shooting down a $30,000 drone with an interceptor that costs many times more is a losing trade, and saturation attacks are designed to exhaust expensive defenses.
2. Seeing them in time. A small, slow, low-flying drone looks a lot like a bird on radar. Cameras mislabel at range, at night or in dust. A hostile drone has no reason to broadcast Remote ID or keep a radio link. Meanwhile the sky is full of legitimate traffic: airliners, helicopters, and authorised inspection drones.

Qatar's most important sites are exactly the kind of places this matters: Hamad International Airport, the LNG facilities at Ras Laffan, and the desalination plants that supply the country's water. Operators there need to know, early and with an honest level of confidence, what is in the air.

Our approach

Rasad is a monitoring system, built and tested entirely in simulation:

1. Fuse every sensor. Simulated radar, RF detectors, visible and thermal cameras, Remote ID and ADS-B feed a multi-sensor tracker. Each track is described in plain English: speed, altitude, turn rate, radar return, RF link, cooperative ID, camera labels, zone and distance to critical areas.
2. Classify with a fine-tuned model. Our fine-tuned Laya reads that description and answers drone, bird, aircraft or unknown with a calibrated probability. Calibration matters: the confidence feeds the threat score and the alert, so "90%" should mean 90%.
3. People decide. A track is flagged as a suspected drone only if the model is at least 80% sure it is a drone and it broadcasts no Remote ID (so authorised maintenance drones are never flagged). A voice alert says "Three suspected drones incoming. Please confirm the defense launch." One click from the operator launches simulated low-cost stinger interceptors at all of them, highest threat first. Nothing launches on its own.

How we built it

A simulator of three Qatari sites. Flyers (quadcopters, fixed-wing and delta-wing drones, gull flocks, airliners, a helicopter shuttle), sensors with realistic flaws (clutter, missed detections, camera mislabels, night, haze, sandstorm, radar outages, bad Remote ID messages) and a real tracker. We ran it hundreds of times:

A synthetic dataset, built without cheating. From 180 generated scenarios we produced 160,148 track descriptions. The text is built only from tracker and sensor output; the simulator's ground truth is used only for the label. Our tests rewrite the ground truth and check that the text never changes. Splits are by scenario, never by row, so the test set is made of runs the model never saw.

Fine-tuning on Hugging Face Jobs. We fine-tuned convaiinnovations/laya (ModernBERT-large) with the laya library on balanced samples, running four configurations in parallel on A100s and picking the winner on the validation split. Each run took about 10 minutes; the whole sweep cost about $2.

Results (held-out simulated test set, 23,358 tracks)

On the hard cases:

The confidence is well calibrated: when Rasad says 95%, it is right about 95% of the time. The weak spot is tracks that sit a few hundred metres off the object they follow, and drones and gulls are still occasionally swapped (264 drones called birds, 367 birds called drones, out of about 21,000).

The demo

The Rasad app runs the whole pipeline live on a GPU: simulator → tracker → fine-tuned Laya → alerts → dashboard. Press Start scenario and three unregistered drones approach Ras Laffan from different directions while an inspection drone, a helicopter and gulls share the sky. Every visitor gets their own private simulation. Voice alerts are generated with the open Kokoro-82M model.

Limits

• Everything here is simulated. Flight behaviour, sensor physics and errors are simplified models we chose; the numbers say nothing about real sensors or real airspace.
• The model only classifies. Interceptors in the demo are simulated, and they launch only when a person confirms.
• Real airspace has more than four kinds of object, and real attackers adapt.

Related work

Rasad builds on several lines of research. Here is where it sits.

1. Telling drones from birds with radar. Narayanan, Tsang & Bharadwaj, "Classification and Discrimination of Birds and Small Drones Using Radar Micro-Doppler Spectrogram Images," Signals 4(2), 2023 (doi:10.3390/signals4020018). Using a 10 GHz radar and the spinning-rotor "micro-Doppler" signature, they separate drones from birds with about 96% accuracy. This confirms drone vs. bird is the core difficulty. Rasad works one level up: instead of raw radar signals, it reasons over the fused track (motion, radar return, RF, ID and camera evidence together).

2. The Drone-vs-Bird Detection Challenge. Coluccia et al., Drone-vs-Bird Detection Challenge, first edition at IEEE AVSS 2017, run yearly since (latest results in IEEE Open Journal of Signal Processing, 2024). A long-running benchmark where systems must raise an alarm for drones but not for birds. Its results show the hardest cases are very distant drones and moving cameras. Rasad's design target is the same rule: flag drones, never alarm on gulls.

3. Anti-drone systems need many sensors. "Securing the Skies: A Comprehensive Survey on Anti-UAV Methods, Benchmarking, and Future Directions," arXiv:2504.11967, 2025. This survey reviews detection with RGB, infrared, audio, radar and RF, and argues for multi-sensor fusion. It flags real-time performance and stealthy targets as open gaps. Rasad follows that advice: radar, RF, cameras, Remote ID and ADS-B are fused into one track before classification.

4. Confidence you can trust. Guo, Pleiss, Sun & Weinberger, "On Calibration of Modern Neural Networks," ICML 2017 (PMLR, arXiv:1706.04599). They show modern networks are often overconfident, introduce the widely used expected calibration error (ECE) measure, and fix overconfidence with temperature scaling. Because Rasad's confidence decides whether an operator is asked to act, we report ECE (0.005) and a reliability diagram, and Laya's fine-tuning fits temperatures on held-out data.

5. The encoder under the hood. Warner, Chaffin, Clavié et al., "Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference" (ModernBERT), arXiv:2412.13663, ACL 2025. ModernBERT is the fast encoder that Laya (our base model) is built on. Being a classifier rather than a text generator is what makes it quick enough to re-check every track every few seconds.

6. Training on synthetic data. Wisniewski et al., "Drone Detection using Deep Neural Networks Trained on Pure Synthetic Data," arXiv:2411.09077, 2024. They train a drone detector only on simulated images and show it carries over to real footage. That supports our approach of building the dataset entirely in simulation. Testing Rasad on real sensor recordings is the next step.

Before publishing:

• I only read the search results, not the full papers. Open each link once to confirm the details you'll be quoting.
• Paper 3 (the survey): the search didn't show its author names, so the entry gives only the title and arXiv number. Add the authors from the arXiv page if you want them.
• Paper 1 (radar): the ~96% figure is the authors' own result on their own radar data. It isn't comparable to our 95.6%; don't put the two side by side.
• A possible seventh paper: a NATO conference paper on classifying drones vs. birds from simulated radar tracks is very close to what we do, but I couldn't confirm its authors or year, so I left it out.

Sources:

• Narayanan et al. 2023, Signals (DOAJ) (https://doaj.org/article/42cf100a43a14a4aad32de604a4c2a80)
• Drone-vs-Bird challenge papers (Univ. Salento IRIS) (https://iris.unisalento.it/handle/11587/445027)
• 2019 edition PDF (Zenodo) (https://zenodo.org/records/4557743/files/08909876.pdf)
• Anti-UAV survey, arXiv 2504.11967 (https://arxiv.org/abs/2504.11967)
• Guo et al. 2017, PMLR (https://proceedings.mlr.press/v70/guo17a)
• Guo et al. 2017, arXiv (https://arxiv.org/abs/1706.04599)
• ModernBERT, arXiv 2412.13663 (https://arxiv.org/abs/2412.13663)
• ModernBERT, ACL Anthology (https://aclanthology.org/2025.acl-long.127)
• Synthetic-data drone detection, arXiv 2411.09077 (https://arxiv.org/pdf/2411.09077)
• NATO STO-MP-MSG-SET-183 (simulated tracks, drones vs birds) (https://www.sto.nato.int/publications/STO%20Meeting%20Proceedings/STO-MP-MSG-SET-183/MP-MSG-SET-183-15.pdf)

Future work

• Protecting Qatar's critical infrastructure. Extend Rasad to airspace monitoring around Hamad International Airport, Hamad Port, Ras Laffan Industrial City and major energy and water facilities, with evaluations focused on correct classification, few false alarms and support for human operators.
• A model built for Qatar's conditions. Train and validate on more realistic scenarios: dust, haze, high temperatures, poor visibility and night operations. Add more object classes than drone, bird, aircraft and unknown.
• Stronger multi-sensor fusion. Better fusion of radar, optical and thermal cameras, RF signals and aircraft identification, especially when sensors fail or disagree.
• Real-world testing in Qatar. Move beyond simulation with real sensor recordings and controlled field trials, subject to approval from the Qatar Civil Aviation Authority and other relevant authorities.
• Tracking and human verification. Track many objects at once more reliably, and add AI-assisted camera checks so operators can review uncertain detections before deciding.
• A better operator console. A faster, interactive 3D dashboard showing Qatar-based scenarios, sensor coverage, aircraft movements, alerts and the evidence behind each one, with Arabic voice alerts.
• Running on site. Run the classifier on local hardware at each site, so monitoring keeps working without an internet connection.
• Scaling across Qatar. Study coordinated monitoring across multiple sites, with secure communication, reliable data processing and one shared picture of the airspace.
• Collaboration with Barzan Holdings. Propose Rasad as a Qatar-based research and development initiative, and explore technical mentorship, prototype improvement, industry expertise and controlled evaluation for future critical-infrastructure monitoring.
• Qatar National Vision 2030. Contribute to a knowledge-based economy by advancing locally developed AI, building national research capability and strengthening infrastructure resilience.

Long-term vision

We want to grow Rasad into a Qatar-focused, AI-powered airspace monitoring platform: one that helps protect national infrastructure, improves situational awareness, and helps human operators spot unauthorised aerial activity early, with people making every decision.

By working with organisations such as Barzan Holdings, running controlled evaluations in Qatar and aligning our research with Qatar National Vision 2030, we aim to contribute to Qatar's technological innovation, research capability and long-term resilience.

Our ambition is to move from a hackathon prototype to a locally developed, validated technology that can serve Qatar's future airspace monitoring needs.