⚡ electrical engineering

Radar Detection in the CBRS Band: Techniques, Challenges, and Future Directions

This paper surveys radar detection techniques within the 3.5 GHz CBRS band, comparing traditional methods with emerging machine learning approaches to address regulatory requirements, performance challenges, and future directions for ensuring coexistence between commercial networks and critical government radar systems.

Original authors: Madan Baduwal, Priyanka Paudel

Published 2026-08-04

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Madan Baduwal, Priyanka Paudel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the air around us is filled with invisible rivers of information, carrying everything from your favorite songs to emergency alerts. For a long time, we treated these "rivers" like private roads, assigning specific lanes to specific groups. But recently, scientists and regulators realized we could be smarter: what if we let different groups share the same road, as long as they knew when to pull over? This is the world of spectrum sharing. In this crowded highway, there are two main types of drivers: the commercial trucks (like your 5G phones and Wi-Fi) and the heavy, critical emergency vehicles (like military radar systems). The rule is simple: the emergency vehicles always have the right of way. If a radar ship is scanning the ocean, the commercial traffic must instantly clear the lane to avoid a crash. The challenge? The commercial trucks can't see the radar ships; they are invisible. So, we need a super-quick, super-smart traffic cop that can spot these invisible ships from miles away and shout, "Move!" before a collision happens. This is the story of the Citizens Broadband Radio Service (CBRS), a special zone where this high-stakes sharing happens, and the paper you are about to read is the ultimate guidebook on how to build that traffic cop.

This paper, titled "Radar Detection in the CBRS Band: Techniques, Challenges, and Future Directions," acts as a massive review of all the different ways we try to spot those invisible radar ships. The authors, Madan Baduwal and Priyanka Paudel, don't just list old tricks; they compare the "old school" methods with the shiny new "AI" methods to see which one actually works best in the messy, noisy real world. They look at how we currently listen for radar signals, how we test these listening systems, and what happens when things go wrong. The big takeaway? While the old methods are like reliable, simple flashlights that work great in a quiet room, the new AI methods are like super-powered night-vision goggles that can see through fog and noise, but they are hungry for data and power. The paper suggests that the future isn't about picking one or the other, but combining them into a hybrid team that is both fast and incredibly smart.

The Invisible Traffic Cop

Let's dive into the heart of the problem. The CBRS band is a stretch of radio waves (around 3.5 GHz) that the government decided to share. On one side, you have the Incumbents—the U.S. Navy and their radar systems. These are the "big bosses" of the spectrum. They are like a lighthouse keeper who must be able to see ships in a storm, no matter what. If a commercial signal gets in the way, it could blind the radar, which is a huge safety risk. On the other side, you have Commercial Users (like private LTE and 5G networks) who want to use that same space to send data.

To make this work, the government created a system called the Environmental Sensing Capability (ESC). Think of the ESC as a network of super-sensitive ears placed along the coastlines. These ears are constantly listening for the specific "whistle" of a radar signal. When an ear hears a radar, it immediately tells a central manager called the Spectrum Access System (SAS). The SAS then acts like a traffic controller, telling all the nearby commercial devices to "stop talking" or "change lanes" for a moment. The rules are strict: the system must hear the radar 99% of the time, even if the signal is incredibly weak (as low as -89 dBm/MHz), and it must do it within 60 seconds.

The Detective's Toolkit: Old vs. New

The paper reviews the different "detective tools" scientists have built to hear these radar whistles.

The Old School Detectives (Classical Methods)
First, there are the traditional signal processing methods. These are like detectives who rely on a specific rulebook.

Energy Detection: This is the simplest method. It's like listening for any loud noise. If the volume goes up, it assumes a radar is there. It's fast and easy, but it's easily fooled. If a nearby Wi-Fi router or a 5G tower gets loud, this method might scream "Radar!" when there isn't one. It's too sensitive to background noise.
Matched Filtering: This detective has a photo of the radar's "face" (its exact signal pattern). It compares what it hears to that photo. If it's a perfect match, it's a radar! This is very accurate, but it only works if the radar looks exactly like the photo. If the radar changes its pattern even a little, this detective gets confused.
Cyclostationary Detection: This one looks for repeating patterns, like a heartbeat. Radar signals often pulse in a regular rhythm. This method ignores random noise and only listens for that specific rhythm. It's great at ignoring noise, but it's slow and requires a lot of brainpower to calculate.

The New School Detectives (Machine Learning & Deep Learning)
Then, the paper introduces the new generation: Machine Learning (ML) and Deep Learning (DL). These aren't detectives with a rulebook; they are like students who learn by studying thousands of examples.

Instead of being told "radars look like this," these systems are fed massive amounts of data (like pictures of sound waves called spectrograms). They learn to recognize the "shape" of a radar signal on their own.
The paper highlights that these AI models, especially Deep Learning ones like CNNs (which are great at spotting patterns in images), can achieve accuracy rates over 99%. They are much better at handling the "fog" of interference from other wireless signals. They can spot a radar even when it's mixed up with 4G or 5G noise, something the old methods struggle with.
However, these AI detectives have a downside: they need a lot of training data (thousands of examples) and they need powerful computers to run. They can also be "black boxes," meaning we don't always know how they decided something was a radar.

The Hybrid Approach: The Best of Both Worlds
The paper suggests that the future might not be choosing between the old and new, but mixing them. Imagine a security guard who first uses a simple motion sensor (the old method) to see if something is moving. If the sensor trips, they then pull up a video feed and use AI (the new method) to check if it's a person or just a cat. This Hybrid Approach uses the speed of the old methods to filter out the noise and the smarts of the AI to make the final call. This could give us the speed we need with the accuracy we want.

The Real-World Hurdles

The authors are careful to point out that while these methods look great in a lab, the real world is messy.

False Alarms: If the system thinks it hears a radar when there isn't one, it shuts down the commercial networks for no reason. This is annoying and costly. The paper notes that balancing the need to catch every real radar with the need to avoid false alarms is a huge challenge.
Speed: The rule says the system has 60 seconds to react. But in the real world, waiting a minute is too long. The paper mentions that newer systems are trying to react in milliseconds, which requires very fast hardware.
The "Weak Signal" Problem: Radar signals can be incredibly faint by the time they reach the sensor. The system must be able to hear a whisper in a hurricane. The paper shows that while AI is getting better at this, it's still hard to guarantee 99% accuracy when the signal is extremely weak.
Data Scarcity: To train the AI, you need lots of real radar data. But collecting real radar signals is hard and expensive. The paper discusses how researchers are using "fake" (simulated) data to help train these models, but there's a risk that the AI might learn the wrong things if the simulation isn't perfect.

What the Paper Says About the Future

The paper concludes that there is no single "magic bullet." The old methods are reliable but rigid, while the new AI methods are flexible but hungry for resources. The most promising path forward is a hybrid system that combines the best of both.

The authors also point out some exciting (and tricky) frontiers. They suggest that future systems might need to be "adaptive," meaning they can learn new radar patterns on the fly without needing to be retrained from scratch. They also mention the need for cooperative sensing, where multiple sensors talk to each other to confirm a radar sighting, making the system much harder to fool. Finally, they warn about security: if AI is used to detect radars, could someone trick the AI with a fake signal? This is a new kind of problem that researchers are just starting to tackle.

In short, this paper is a roadmap for building the ultimate traffic cop for the wireless world. It tells us that while we have made huge strides with AI, the job isn't done yet. We need systems that are not only smart but also fast, cheap, and tough enough to handle the chaos of the real world, ensuring that our phones stay connected while the Navy's radar stays safe.

Technical Summary: Radar Detection in the CBRS Band

Problem Statement

The Citizens Broadband Radio Service (CBRS) band (originally 3550–3700 MHz, later expanded to include 3450–3550 MHz, covering 3.45–3.7 GHz) operates as a shared spectrum environment in the United States, allowing commercial networks (LTE/5G) to coexist with incumbent government systems, primarily U.S. Navy radar. The Federal Communications Commission (FCC) mandates that commercial users must vacate or adjust their transmissions immediately upon the detection of radar signals to prevent harmful interference. This requirement is enforced through the Environmental Sensing Capability (ESC), a network of sensors that must detect radar signals with a probability of detection ( $P_d$ ) of at least 99% within 60 seconds, even for extremely weak signals (down to −89 dBm/MHz).

The core challenge lies in the variability of radar signals (pulse width, repetition frequency, and strength) and the presence of significant interference from commercial wireless systems (LTE, Wi-Fi, 5G) and adjacent bands. Traditional detection methods often struggle in these complex, low-SNR environments, while modern data-driven approaches face challenges regarding computational latency, the need for large labeled datasets, and generalization to unseen signal conditions.

Methodology

This paper presents a comprehensive survey of radar detection techniques within the CBRS band. The methodology follows a structured literature review process guided by the PRISMA framework (Fig. 2).

Search Strategy: The authors queried major digital libraries (IEEE Xplore, ACM Digital Library) and regulatory documents (FCC 47 CFR Part 96, WInnForum specifications) for studies published between 2018 and 2026.
Selection Criteria: From an initial pool of 150 records, 25 studies were selected after screening for relevance to radar detection in the CBRS band, technical depth, and methodological clarity.
Taxonomy: The survey categorizes detection approaches into four distinct groups:
1. Classical Signal Processing: Methods such as Energy Detection, Matched Filtering, Cyclostationary Feature Detection, and Constant False Alarm Rate (CFAR).
2. Feature-Based Machine Learning: Approaches using handcrafted features (e.g., pulse width, PRF) with classifiers like Support Vector Machines (SVMs).
3. Deep Learning (DL) Methods: Models, particularly Convolutional Neural Networks (CNNs) and YOLO-based architectures, that learn patterns directly from spectrograms or raw IQ data.
4. Hybrid Approaches: Systems that integrate classical detectors (for low-latency screening) with deep learning models (for refined classification).

The paper also reviews available datasets (synthetic NIST datasets vs. real-world measurements) and simulation platforms used to validate these methods.

Key Contributions

The survey offers a unified technical overview of the CBRS radar detection landscape with the following specific contributions:

Unified Taxonomy: It integrates classical and learning-based methods into a cohesive system-level pipeline, clarifying the transition from model-driven to data-driven strategies.
Comparative Analysis: It provides a detailed comparison of detection techniques under specific CBRS constraints, including low-SNR operation (down to −89 dBm/MHz) and strict latency requirements (60-second evacuation window).
Resource Review: It compiles a comprehensive list of datasets (e.g., NIST synthetic data, TN 1954 real-world measurements) and simulation platforms (MATLAB, GNU Radio) used in ESC validation.
Challenge Identification: It highlights critical deployment hurdles, including false alarm management, interference from modern wireless systems, hardware limitations, and the need for real-time operation on edge devices.
Future Directions: It outlines open research problems, such as low-SNR detection, adaptive/transferable models, adversarial robustness, and the integration of sensing into 5G/6G architectures (e.g., O-RAN).

Results and Findings

The survey synthesizes findings from the selected literature to draw several technical conclusions:

Performance Trade-offs: Classical methods (e.g., Matched Filtering) offer theoretical optimality and low latency when radar waveforms are known but degrade significantly in heterogeneous environments with unknown signals or strong interference. Conversely, deep learning models (e.g., CNNs, DeepRadar, ESC+) demonstrate superior robustness, achieving detection accuracies exceeding 99% in complex, interference-heavy scenarios (e.g., at 17–20 dB SINR).
Hybrid Effectiveness: Hybrid approaches, which combine the speed of classical detectors (like CFAR) with the pattern recognition of deep learning, emerge as a promising solution. These systems aim to reduce computational costs while maintaining high accuracy and minimizing false alarms.
Data Dependency: While deep learning models outperform traditional methods in real-world conditions, they rely heavily on large, labeled datasets. The survey notes that synthetic data (e.g., NIST) is widely used for training, but real-world over-the-air datasets remain limited.
Latency and Hardware: Although the FCC allows up to 60 seconds for detection, modern implementations (e.g., SenseORAN) aim for sub-second latency (e.g., ~866 ms) using hardware acceleration (FPGAs) and lightweight model architectures (e.g., quantization, YOLO adaptations).

Significance and Claims

The paper positions itself as a critical resource for understanding the evolution of radar detection in shared spectrum environments. It claims that while traditional methods remain simple and reliable in controlled settings, the increasing complexity of the RF environment necessitates a shift toward learning-based and hybrid approaches.

The authors assert that the future of CBRS radar detection lies not in a single "perfect" method, but in hybrid strategies that leverage the reliability of classical signal processing and the adaptability of machine learning. The significance of this work is framed around enabling the safe and efficient sharing of the 3.5 GHz band, ensuring that critical government radar systems are protected without unnecessarily disrupting commercial 5G and private LTE networks. The paper concludes that achieving accurate, fast, and robust performance in real-world deployments requires addressing open challenges in low-SNR detection, model generalization, and adversarial security.