Frequency-Modulated Continuous-Wave Radar Perspectives on Unmanned Aerial Vehicle Detection and Classification: A Primer for Researchers with Comprehensive Machine Learning Review and Emphasis on Full-Wave Electromagnetic Computer-Aided Design Tools

Abstract

Unmanned Aerial Vehicles (UAVs) represent a rapidly increasing technology with profound implications for various domains, including surveillance, security, and commercial applications. Among the number of detection and classification methodologies, radar technology stands as a cornerstone due to its versatility and reliability. This paper presents a comprehensive primer written specifically for researchers starting on investigations into UAV detection and classification, with a distinct emphasis on the integration of full-wave electromagnetic computer-aided design (EM CAD) tools. Commencing with an elucidation of radar’s pivotal role within the UAV detection paradigm, this primer systematically navigates through fundamental Frequency-Modulated Continuous-Wave (FMCW) radar principles, elucidating their intricate interplay with UAV characteristics and signatures. Methodologies pertaining to signal processing, detection, and tracking are examined, with particular emphasis placed on the pivotal role of full-wave EM CAD tools in system design and optimization. Through an exposition of relevant case studies and applications, this paper underscores successful implementations of radar-based UAV detection and classification systems while elucidating encountered challenges and insights obtained. Anticipating future trajectories, the paper contemplates emerging trends and potential research directions, accentuating the indispensable nature of full-wave EM CAD tools in propelling radar techniques forward. In essence, this primer serves as an indispensable roadmap, empowering researchers to navigate the complex terrain of radar-based UAV detection and classification, thereby fostering advancements in aerial surveillance and security systems.

1. Introduction

Due to the widespread proliferation of Unmanned Aerial Vehicles (UAVs) worldwide, they have increasingly been utilized in various illicit activities, including but not limited to drug and weapons smuggling across borders or into prisons, interference with aircraft operations, invasion of privacy, and potential involvement in terrorist acts [1,2,3,4,5,6,7,8,9,10,11]. Consequently, the imperative to accurately identify and classify UAVs from other airborne targets is primary for ensuring safety and security measures. Therefore, it is essential to detect and classify UAVs from a distance.

Three primary methods are commonly employed for the detection of UAVs: acoustic, optical, and Radio Frequency (RF) [12,13,14,15,16,17,18]. Acoustic methods, recognized for their ease of installation and relatively modest cost, offer the advantage of not mandating a Line of Sight (LOS) for UAV detection. Moreover, these methods can be effectively coupled with Machine Learning (ML) algorithms for classification purposes. However, acoustic techniques are limited in their capacity for long-range detection and are notably vulnerable to interference from ambient environmental noise. In contrast, optical methods, particularly when integrated with ML algorithms, provide a robust solution capable of capturing visual images of targets, tracking autonomous UAVs, and are generally considered cost-effective, with expenses ranging from low to medium. Nonetheless, optical systems are susceptible to weather conditions, necessitate LOS for effective UAV detection, and may face challenges in reliably detecting small UAVs emitting insufficient thermal heat signatures utilizing thermal cameras [12,13,14,15,16,17,18].

RF detection methods encompass the utilization of radio sensors and radar-based systems. In radio sensor detection, a receiver intercepts RF communications between UAVs and their controllers. This approach is regarded as cost-effective, falling within the low- to medium-cost range, and not only facilitates the detection of UAVs and their controllers, but also exhibits a low false alarm rate. Furthermore, it boasts the capability of detecting UAVs over extended distances. Nevertheless, radio sensors falter in detecting autonomous UAVs due to the absence of communication with controllers and may encounter difficulties in simultaneously detecting numerous UAVs while mandating LOS for optimal operation [12,13,14,15,16,17,18].

Radar-based detection systems, distinct from optical methods, remain impervious to adverse weather conditions and operate proficiently both day and night. These systems can track autonomous flights, concurrently detect and track multiple UAVs, and when integrated with ML algorithms, are capable of classifying various targets and discerning between different types of UAVs. However, radar-based detection systems are considered a high-cost solution compared to other detection methodologies, and necessitate LOS for UAV detection [12,13,14,15,16,17,18].

Table 1. Comparison of UAV detection methods.

Method	Acoustic	Optical	Radio Sensors	Radars
LOS Independence	✓	×	×	×
Long Range Detection	×	×	✓	✓
Weather Conditions	×	×	✓	✓
Detect Autonomous UAVs	✓	✓	×	✓
Classify UAVs and Birds	×	✓	×	✓
Compatibility with ML	✓	✓	✓	✓
Detect Controller Location	×	×	✓	×

summarizes the key distinctions among the four UAV detection methods. Notably, radar systems exhibit numerous advantages over alternative techniques.

However, a primary limitation in this application lies in the generation of radar datasets containing UAV information for training ML algorithms [12]. This undertaking proves to be both resource-intensive and time-consuming, constrained by various factors including radar parameters, available UAV models, and environmental backgrounds. Moreover, the scarcity of accessible radar UAV datasets further exacerbates this challenge, with existing datasets often limited in scope to specific UAV types and radar configurations employed in prior studies [19,20,21,22,23,24]. To overcome these constraints, numerical simulations offer a dependable alternative approach.

Full-wave EM CAD tools have been used in prior studies that attempted to mathematically simulate Radar Cross Section (RCS) of UAVs and the spectrograms of rotor blades on UAVs, or have employed CAD models of propellers to generate micro-Doppler signatures through the conversion of 3D CAD models into point clouds [25,26,27,28,29,30,31,32]. Notably, certain trials, exemplified by [30], entail simulations that may span up to 160 h. Recently, a new method has emerged to expedite these simulations, enabling significantly more complex simulations to be completed within minutes or even seconds when utilizing Graphics Processing Units (GPUs) [33,34,35,36,37,38,39,40,41]. Moreover, this approach replicates UAV motion utilizing full-wave EM CAD tools, encompassing more than just rotor blade rotation. The incorporation of time-based full-wave analysis, an unprecedented feature integrated into this method, further accentuates this uniqueness. Furthermore, this methodology not only facilitates the generation of micro-Doppler signatures of UAVs, but also yields a diverse array of outcomes, including but not limited to range profiles, waterfalls, range-Doppler maps, Inverse Synthetic Aperture Radar (ISAR) images, and range-angle maps when employing Multiple-Input Multiple-Output (MIMO) radars.

This paper presents a primer review aimed at researchers, delineating methodologies for detecting and classifying UAVs utilizing radar systems. The review encompasses an introductory overview of radar fundamentals, elucidating radar detection methods, simulation setups, and the generation of datasets essential for training ML models. Additionally, the paper delves into the application of ML models to these datasets. As shown in Figure 1, the process begins with acquiring radar measurements, which are then subjected to signal processing algorithms to extract range-Doppler and micro-Doppler information. Subsequently, utilizing the range-Doppler information and micro-Doppler signatures, the necessary datasets for this application are generated. These datasets play a pivotal role in training machine learning models for accurate detection and classification of UAVs. Through the exposition of relevant case studies, the paper highlights successful implementations of radar-based UAV detection and classification systems, while also elucidating encountered challenges and insights garnered from these endeavors. Anticipating future trajectories, the paper contemplates emerging trends and potential avenues for research, emphasizing the indispensable role of full-wave EM CAD tools in advancing radar techniques. Furthermore, a block diagram illustrating the procedural steps involved in radar UAV detection and classification application is provided, offering a comprehensive overview of the methodology discussed.

The paper is organized as follows: Section 2 provides an introduction to the fundamentals of FMCW radar systems. Section 3 elucidates the generation of radar datasets that contain UAV information. Section 4 introduces the application of Machine Learning for UAV classification. Finally, Section 5 addresses challenges, outlines future trends, and presents concluding remarks.

2. Fundamentals of FMCW Radar Systems

Radar systems operate across diverse frequency bands and function by emitting electromagnetic waves and subsequently analyzing the reflections thereof. Notably, Frequency-Modulated Continuous-Wave (FMCW) radars are highly sought after owing to their relatively simplistic architecture, customizable resolution, and capacity to furnish both distance and velocity data [42,43]. FMCW radar systems utilize a variety of waveforms, with the chirp waveform emerging as particularly prevalent in commercial radar applications. This waveform, depicted in Figure 2, is commonly denoted as the sawtooth Linear Frequency-Modulated Continuous Wave (LFMCW) radar waveform. The transmitted FMCW signal can be delineated as follows:

$$X_t\left(t\right)=A_{t}cos(2\pi f_{min}t+\pi k{t}^2),0\le t\le T_{sweep}$$

(1)

where $X_t\left(t\right)$, $A_t$, $f_{min}$, $f_{max}$, k, t, and $T_{sweep}$ represent the transmitted signal, the amplitude of the transmitted signal, the minimum frequency utilized, the maximum frequency utilized, the chirp rate, the time vector, and the sweep time per one chirp, respectively. The transmitted bandwidth, $BW$, is determined by the difference between the maximum and minimum frequencies, as expressed by Equation (2).

$$BW=f_{max}-f_{min}$$

(2)

Meanwhile, the chirp rate k is defined as the ratio of the bandwidth $BW$ to the sweep time $T_{sweep}$, as depicted in Equation (3).

$$k={\displaystyle \frac{BW}{T_{sweep}}}$$

(3)

The received signal $X_r\left(t\right)$, delayed by $\tau$, is described by Equation (4), where $A_r$ represents the received signal amplitude, and

$$\begin{array}{c}\hfill X_r\left(t\right)=A_{r}cos(2\pi f_{min}(t-\tau )+\pi k{(t-\tau )}^2),\\ \hfill 0\le t\le T_{sweep}\end{array}$$

(4)

where $A_r$ is the amplitude of the received signal, while $\tau$ denotes the time delay of the signal, which correlates with the target range according to the relationship:

$$\tau ={\displaystyle \frac{2R}{c}}$$

(5)

where $c=3\times {10}^8$ m/s denotes the speed of light. After incorporating the target radial velocity $v_r$, Equation (5) transforms into:

$$\tau ={\displaystyle \frac{2(R-v_{r}t)}{c+v_r}}$$

(6)

Given that $v_r\ll c$, Equation (6) simplifies to:

$$\tau ={\displaystyle \frac{2(R-v_{r}t)}{c}}$$

(7)

When mixing the transmitted and received signals through a mixer and filtering the mixer output with a Low Pass Filter (LPF), the resulting signal, termed the Intermediate Frequency (IF) signal, is described as:

$$X_{IF}\left(t\right)={\displaystyle \frac{1}{2}}{A}_t{A}_{r}cos(2\pi f_{min}\tau +2\pi kt\tau -\pi k{\tau }^2)$$

(8)

Here, $X_{IF}\left(t\right)$ denotes the IF signal. Given the insignificance of the delay, $2\pi f_{min}\tau \gg \pi k{\tau }^2$, the IF signal can be reformulated as:

$$X_{IF}\left(t\right)={\displaystyle \frac{1}{2}}{A}_t{A}_{r}cos2\pi f_{min}\tau +2\pi kt\tau$$

(9)

Consequently, the IF frequency of the IF signal $X_{IF}\left(t\right)$ is expressed as:

$$f_{IF}=k\tau t+f_{min}\tau =k{\displaystyle \frac{2R}{c}}-{\displaystyle \frac{2v_r}{c}}(2kt+f_{min})$$

(10)

Here, $k\frac{2R}{c}$ represents the beat frequency utilized for target range measurement, and the Doppler frequency is denoted as $f_D=\frac{2v_r}{c}{f}_c=\frac{2v_r}{\lambda }$, where $f_c$ is the center frequency, and $\lambda$ is the wavelength. Thus, the IF frequency is:

$$f_{IF}=k{\displaystyle \frac{2R}{c}}-{\displaystyle \frac{2v_r}{\lambda }}$$

(11)

Alternatively, it can be expressed as:

$$f_{IF}=f_b-f_D$$

(12)

Most UAVs exhibit micro-Doppler signatures as a common characteristic, which have been extensively utilized for radar object classification. Consequently, range-Doppler maps and micro-Doppler signatures play a crucial role in the classification process of UAVs, as evidenced by several studies [44,45,46,47,48,49,50,51]. Figure 2 illustrates the procedural steps involved in generating range-Doppler maps in FMCW radars. To generate the range-Doppler maps, the following process is applied: initially, a Fast Fourier Transform (FFT) is performed on the fast-time axis, denoted as $t_f$, for each chirp, yielding range information for each chirp. Subsequently, another FFT is applied to the slow-time axis, denoted as $t_s$, spanning the entire frame. This operation captures variations in range across the chirps, resulting in both Doppler information and the final range-Doppler maps [21]. This process can be represented as follows:

$$\begin{array}{cc}\hfill RP(R,t_s)& \iff FFT\left\{X_{IF}(t_f,t_s)\right\}\hfill \\ \hfill RDM(R,f_D)& \iff FFT\left\{RP(R,t_s)\right\}\hfill \end{array}$$

(13)

Here, $X_{IF}$ denotes the received radar signal in the IF domain, $RP(R,t_s)$ signifies the range profile across the entire frame, and $RDM(R,f_D)$ represents the range-Doppler map, which serves as a two-dimensional representation of radar echoes in the range (R) and Doppler frequency ($f_D$) domains.

The micro-Doppler signatures of UAVs, or spectrograms, are generated by applying a Short-Time Fourier Transform (STFT) to the received radar signal to analyze its frequency content over time. The received signal is divided into segments or windows, and the STFT is applied to represent the signal’s frequency content within each window. The length of each window is selected to balance the trade-off between time and frequency resolutions, as depicted in Equation (14) [47,48,51,52,53,54,55,56]:

$$X(\tau ,\omega )={\int }_{-\infty }^{\infty }{X}_{IF}\left(t\right)\Omega (t-\tau )e^{-j\omega t}dt$$

(14)

Here, $\Omega \left(\tau \right)$ represents the window function, $X(\tau ,\omega )$, and denotes the Fourier transform of the received signal’s segment according to the window length. Subsequently, the spectrogram representation is obtained by taking the magnitude squared of the STFT in a logarithmic scale, as shown in Equation (15) [47,48,51,52,53,54,55,56]:

$$S(\tau ,\omega )=10{log}_{10}{\left|X(\tau ,\omega )\right|}^2$$

(15)

Here, $S(\tau ,\omega )$ signifies the spectrogram of the received signal in a logarithmic scale. This representation proves beneficial for visualizing signals with a wide range of power spectral density in the logarithmic scale, as it compresses the dynamic range of the spectrogram.

3. Data Collection

Training ML models requires radar datasets containing UAV information to improve their accuracy in UAV classification. Furthermore, these datasets play a crucial role in the development of new ML models and signal processing techniques within the field of radar-based UAV detection and classification. However, the process of generating radar measurement datasets for training ML algorithms for UAV classification is both time-consuming and costly. The accuracy of these datasets is inherently limited by factors such as radar system parameters, UAV composition, detection range, and environmental conditions [19]. However, simulation techniques offer a viable solution to overcome these constraints by enabling the generation of large datasets encompassing diverse types and sizes of UAVs [33,34,35,36,37,38,39,40,41]. Controlled simulations allow for the investigation and analysis of the effects of various radar parameters, UAV characteristics, and environmental factors. Despite the existence of limited radar UAV datasets, their utility is contingent upon specific considerations such as UAV types, radar system parameters, detection range, and environmental conditions [19,20,21,22,23].

Previous studies have employed full-wave EM CAD tools to mathematically simulate the RCS of UAVs and the spectrograms of rotor blades, or have utilized CAD models of propellers to generate micro-Doppler signatures by converting 3D CAD models into point clouds [25,26,27,28,29,30,31,32]. Notably, some of these simulations, such as [30], have required extensive time, with trials lasting up to 160 h. Recently, a new approach has emerged to expedite these simulations, allowing for significantly more complex simulations to be completed within minutes or even seconds when employing GPUs [33,34,35,36,37,38,39,40,41]. Furthermore, this approach goes beyond merely replicating rotor blade rotation, as it utilizes full-wave EM CAD tools to simulate UAV motion comprehensively. The integration of time-based full-wave analysis, a feature unique to this method, further underscores its distinctiveness.

Full-wave EM CAD tools effectively simulate the scattering and RCS of UAVs. Initially, a 3D model of the UAV and its environment is created within the design environment, specifying shape, materials, and dimensions for the UAV, as well as surrounding structures affecting radar scattering. A mesh is then applied to ensure a precise representation of structures and electromagnetic interactions. Material properties are assigned to the UAV and surrounding objects, utilizing the library of materials or defining custom materials with specific electromagnetic characteristics. Radar system parameters such as frequency, polarization, and incident angle are defined to analyze UAV scattering behavior. Simulation parameters including frequency range, solver settings, and convergence criteria are configured. Post-simulation, results are analyzed and visualized to show radar signatures of the UAV. These outcomes aid in UAV detection and classification based on scattering characteristics. The accuracy and realism of simulations depend on factors such as model quality, material selection, and accurate assignment of physical properties. As illustrated in Figure 3a–c [40], the DJI S900 UAV and its corresponding model along with three-time stamps, separated by 1 millisecond (ms), depicting UAV blade rotation, were utilized for simulation, with the 3D model sourced from the GrabCAD library [57,58]. In Figure 3, Ansys HFSS [59], a full-wave EM CAD tool, is showcased.

4. Machine Learning for UAV Classification

Traditional radar signal processing methods have demonstrated certain limitations in target classification tasks. Although effective for basic target detection by relying on fundamental signal characteristics such as amplitude, time-of-arrival, and Doppler frequency shift, these methods often struggle to distinguish between different types of targets, especially in complex environments with clutter or multiple targets. They may also lack the sophistication needed to discern intricate features or patterns that could aid in target classification and can be susceptible to noise and interference, further complicating the classification process. Moreover, traditional radar signal processing techniques may not fully leverage advanced machine learning algorithms or pattern recognition techniques, which have shown promise in improving target classification accuracy but often require large datasets for training and may pose computational challenges for real-time applications. Integrating machine learning (ML) with radar detection algorithms significantly enhances the accuracy of UAV detection and classification.

4.1. Classification of UAVs and Birds

In [60], a low-cost radar-based system for UAV detection and classification is presented, distinguishing UAVs from common urban objects. It utilizes a two-step process, initially extracting micro-Doppler features and providing preliminary classifications. Subsequently, it combines data from successive time segments to make the final determination. In [61], the challenge of classifying multiple UAVs and birds based on micro-Doppler signatures is addressed. It introduces three protocols and demonstrates the most effective one for distinguishing UAVs from birds in real observation scenarios using radar and a Convolutional Neural Network (CNN). In [46], an X-band radar system is used for UAV detection. Experimental results with measured radar signals show a good performance in detection accuracy and computational efficiency. In [44], micro-Doppler signatures of a small UAV are used. Experimental results show the potential of micro-Doppler features for UAV detection and classification. In [45], an iterative adaptive technique for enhancing Doppler processing in ground-based surveillance radar systems used for UAV detection is proposed. The method overcomes limitations in dwell time and improves Doppler resolution, resulting in better discrimination of micro-Doppler signatures and accurate UAV classification.

In [47], micro-Doppler signatures for distinguishing UAVs and birds using radar data are explored. An accuracy of $96\%$ in UAV and bird classification and $85\%$ accuracy in classifying individual UAVs and birds into five categories are achieved using a Support Vector Machine (SVM). In [62], a system to automatically classify UAVs from other objects, especially birds is introduced. It proposes a complex-log-Fourier transform to better handle micro-Doppler signatures, addressing the phase spectrum. Additionally, it presents a dimension-reduction technique called subspace reliability analysis, reducing the error rate from $6.68\%$ to $3.27\%$ compared to the Cadence Velocity Diagram (CVD). In [52], classification features from micro-Doppler signatures are extracted and employs eigenpairs from correlation matrices for classification, showing an accuracy of $95\%$ to classify different UAVs and birds.

In [48], micro-Doppler signatures are extracted from spectrograms and cepstrograms to distinguish between birds and small UAVs. In [50,51], an analysis of micro-Doppler features from small UAVs using a 94 GHz radar in CW and FMCW modes is discussed in the first work. Wavelet decomposition extracts high-frequency components corresponding to small UAV movements, and time-frequency analysis is employed for feature visualization and interpretation. In the second work, radar micro-Doppler properties of UAVs and birds are explored. Experimental measurements were conducted using K-band and W-band radars with various UAV models and bird species. The results demonstrate that both radar frequencies can retrieve reliable and distinct micro-Doppler signatures, with the W-band radar having a higher Signal-to-Noise Ratio (SNR). In [22], using a 77 GHz radar, a CNN classifier’s ability to discriminate UAVs from non-UAV targets, such as birds, is evaluated when they are not represented in the training data. It is shown that the mean accuracy for out-of-distribution UAVs is $78\%$. By introducing a synthetic UAV class created through a mathematical model, the accuracy improves to $86\%$. When trained on all UAV types, the mean accuracy across all classes reaches $90\%$.

In [26,27,28,63], in the first work, a method is proposed to extract a symmetry feature from micro-Doppler signatures of UAV targets, maintaining explainable AI for safety applications. In the second work, the use of micro-Doppler signatures for differentiating UAVs from other objects, like birds, is highlighted. It also explores how the signatures vary with different UAV component shapes. It involves the design and performance of a 94 GHz radar for lab-based micro-Doppler measurements. In the third work, the classification of UAV types using micro-Doppler signatures is addressed. It presents a simulation approach that models UAV components as point clouds derived and compares simulated results with experimental data. In the fourth work, models to simulate radar returns from multi-rotor UAVs using real motor speed data are created, generating synthetic spectrograms that capture the micro-Doppler features.

4.2. Micro-Doppler Signatures for UAV Classification

In [49], a method uses the micro-Doppler effect, and pattern recognition is presented to identify UAVs in real-time with accuracy. In [64], a precise UAV detection and localization system using mm-Wave radar, combining spatial heatmaps, micro-Doppler analysis, and Gaussian process regression to improve accuracy over direct spectral analysis is proposed. In [65], micro-Doppler results using FMCW radar and ML are used to locate and classify UAVs, achieving an accuracy of $85.1\%$ in classifying four UAV models. In [66], a CNN algorithm is applied to the Continuous Wavelet Transform (CWT) and Discrete Wavelet Transform (DWT) scalogram of reflected radar signals from UAVs. In [67], radar micro-Doppler is used to detect and classify small UAVs based on their movements. Radar measurements of UAVs and birds are analyzed using Time Velocity Diagram (TVD), and a boosting classifier is employed for target detection and classification offering advantages over the SVM classifier. In [56], the challenging task of distinguishing between rotary-wing UAVs and flying birds using K-band FMCW radar is addressed. A data augmentation method is used to enhance the collected dataset. A CNN model is used to improve feature extraction and reduce computational complexity, achieving high classification accuracy. In [25], CNN models are utilized to classify UAVs based on radar signals, using STFT spectrograms. In [68], a method for tracking and recognizing low-altitude UAVs and birds using traditional surveillance radar data are introduced. It establishes motion models, filters and smooths target trajectories, calculates model occurrence probabilities, and identifies target types.

In [54,69,70], various methods for mini-UAV classification based on radar data are explored. The first approach employs Empirical Mode Decomposition (EMD) to extract oscillating waveforms from radar echo signals, which are then used to derive features capturing blade flashes for classification using a nonlinear SVM. In the second work, FMCW radar returns from various UAVs and non-UAV objects in terms of micro-Doppler signatures are analyzed. Based on this analysis, it uses a UAV classification system consisting of five key components: burst selection, rule-based scan pruning, EMD-based micro-Doppler analysis and feature extraction, error counting minimization for class label estimation, and scan-to-scan filtering. The third investigation delves into the challenges posed by the weak micro-Doppler signals from small UAV rotor blades, using EMD to extract the rotation components effectively, which has been validated as a viable method for UAV identification. In [71], six entropies extracted from Intrinsic Mode Functions (IMFs) for radar-based mini-UAV classification are investigated. Three entropies are selected due to their efficiency and accuracy, and are used in combination with different sets of IMFs to improve classification accuracy. In [72], a model for micro-Doppler effects to analyze UAV signatures is introduced. It focuses on the number of UAV motors and their rotation speed. The model is validated by comparing simulated radar images with data from a 77 GHz FMCW MIMO radar. Results demonstrate the model’s accuracy in classifying UAVs based on the number of motors, with motor rotation speed not significantly affecting the classification.

In [53], UAVs are recognized using micro-Doppler features generated by rotor rotation. A FFT and Two-Dimension Principal Component Analysis (2DPCA) are applied to extract features efficiently without high-dimensional data conversion. The method achieves a $98.44\%$ recognition rate for small UAVs, reducing computational complexity and feature extraction time. In [73], a method for detecting and recognizing rotor UAVs based on their micro-motion characteristics is presented. It compensates for translational motion, estimates micro-Doppler parameters, and enhances target detection and identification. In [74], FMCW radar signal models are examined in the presence of UAVs, and an approach for high-frequency micro-Doppler signature analysis is introduced. In [75], the limitations of micro-Doppler methods for detecting small UAVs with weak radar returns from plastic propellers are addressed. A rotary UAV detector based on the number of Helicopter Rotation Modulation (HERM) lines is introduced and two parametric methods are evaluated, Minimum Description Length (MDL) and Akaike Information Criterion (AIC), for HERM line estimation. In [76], a data augmentation method for generating simulated radar micro-Doppler signatures to classify UAVs is presented. It demonstrates that using this method to train a Deep Neural Network (DNN) improves accuracy. The study focuses on classifying the number of UAV motors using a 77 GHz radar system, achieving a classification accuracy of $78.68\%$.

In [23,24], a dataset of micro-Doppler signatures from various small aerial targets is developed in the first work and transfer learning-based DCNN is used, achieving high classification accuracy. The second work introduces a lightweight DCNN model, for accurate detection and classification of small UAVs based on their micro-Doppler signatures, achieving a classification accuracy of $97.1\%$ to $97.3\%$. In [77], the radar signatures of various consumer UAVs through laboratory measurements are examined. The UAVs are rotated on a turntable, and their back-scattered radar data are collected at two different frequency bands. The data are then processed to create ISAR images.

In [78,79,80], UAVs are classified based on micro-Doppler signatures created by rotor blade rotation in the first work. It suggests using both K-band and X-band radar data to enhance accuracy. Spectrograms are formed, PCA extracts features, and SVM is used for classification. Dual-band radar fusion outperforms single-band radar in classification accuracy. In the second work, direct-path and multi-path micro-Doppler radar signatures are combined to improve micro-UAV classification accuracy in urban environments, achieving a $5\%$ increase in accuracy. The third work introduces a recognition method for multiple micro-UAVs based on their micro-Doppler signatures using dictionary learning, with a $93\%$ recognition performance in indoor environments when trained on half of the measured data. In [81], a method for localizing and classifying UAVs using mm-Wave FMCW radar is presented. The UAV’s height and distance are estimated, and the UAV’s activity is classified with lightweight models like logistic regression, SVM, and a CNN. In [82], a UAV classification method that utilizes a CNN and micro-Doppler signatures is presented. Micro-Doppler signatures and CVD are combined to create a merged Doppler image, which is processed by a CNN model. In [83], the effectiveness of W-band radar for detecting small UAVs and emphasizes radar’s reliability in challenging scenarios are discussed. The advantages of FMCW radar in the mm-Wave range are highlighted, including compactness and suitability for various surveillance applications, especially when other sensors may be less effective. In [84], a radar system for detecting and classifying micro UAVs based on Doppler signatures and spectral correlation functions is introduced. It employs a Deep Belief Network (DBN) for classification and achieves over $90\%$ accuracy in UAV detection and classification using a $2.4$ GHz CW Doppler radar in lab experiments.

4.3. RCS-Based UAV Detection and Classification

In [85], EM waves scattering from micro-UAV rotor blades are studied, considering factors like polarization, frequency, and azimuth angle. Both simulations and experiments reveal variations in scattering with azimuth and frequency. In [86], RCS-based statistical system for classifying UAVs is presented at microwave frequencies. It begins with RCS measurements of six commercial UAVs at various frequencies and different polarizations. It shows that the UAVs’ RCS depends on factors like shape, size, material, azimuth angle, frequency, and polarization. Statistical models are used to characterize the UAVs, leading to a recognition system with an average accuracy exceeding $97.60\%$ at an SNR of 10 dB. In [20], RCS measurements on different UAVs at 26–40 GHz are conducted for two categories: carbon fiber and plastic UAVs. It is shown that larger carbon fiber UAVs are easier to detect, while plastic UAVs are less visible. In [87], improved UAV classification using mm-Wave radar, used in [20], and deep learning is presented. It introduces an efficient Long Short-Term Memory (LSTM) model with a weight optimization technique and Adaptive Learning Rate Optimizing (ALRO). The LSTM-ALRO approach achieves a high UAV detection accuracy of $99.88\%$. In [21], a system for UAV detection and classification using persistent range-Doppler radar is presented. It combines a Constant False Alarm Rate (CFAR) detection stage with a CNN for classification. It classifies UAVs, cars, and people with an accuracy of $99.48\%$.

4.4. Multi-Static Radar Systems for UAV Detection

In [88,89], a multi-static radar system is used to study UAVs carrying payloads. The first study distinguishes between hovering and flying UAVs with different payloads, achieving over $96\%$ accuracy in payload weight classification. The second study employs domain transformations and CNN to classify UAVs with or without payloads, achieving approximately $95.1\%$ accuracy for hovering UAVs and $96.6\%$ for flying UAVs. In [90], a multi-static radar is used with three steps signal processing algorithm. A CFAR is used to detect the targets, micro-Doppler signatures are used to discriminate the detected targets, and finally, a Kalman filter is used to track the targets over time. In [91], radar measurements of small UAVs using a multi-static radar system at L- and X-bands are presented. It includes range-time plots, micro-Doppler signatures, and a quantitative analysis of signal quality. The study achieved over $90\%$ classification accuracy for distinguishing UAVs from other objects using simple features and classifiers.

4.5. UAV Detection and Classification in Low SNR Conditions

In [92,93], an investigation on the effect of SNR on UAV-bird classification using radar data are conducted. It is shown that classification accuracy drops with lower SNR. Data augmentation improves accuracy at low SNR levels, achieving an overall accuracy of $92\%$. In [94], a K-band radar system with high sensitivity for detecting low RCS targets and measuring their range and velocity is introduced. It employs fiber-optic links to reduce leakage and platform designs to minimize ground reflections. In [95], micro UAV detection in rocky terrain using a 24 GHz dual-polarized FMCW radar is explored. It deals with land clutter, using a CFAR detector, and enhances UAV discrimination by extracting and plotting the micro-Doppler signature of rotating propellers and the UAV’s trajectory in the time-frequency domain. In [55], a clutter suppression method using an Orthogonal Matching Pursuit (OMP) algorithm for LFMCW radar is presented. The method reduces clutter power and extracts micro-Doppler signals for distinguishing different UAVs in low SNR conditions.

4.6. MIMO Radars for UAV Classification

In [96,97,98,99,100,101], MIMO radars are used to improve the detection and classification of UAVs. The first work emphasizes enhancing UAV detection by utilizing range-velocity processing and micro-Doppler analysis for improved classification. The second work discusses the development of a mm-Wave FMCW radar system for detecting small UAVs, addressing technology challenges and performance expectations. The third work presents a low-cost $2\times 2$ MIMO radar prototype developed with Software Defined Radio (SDR) technology and evaluates its performance in detecting slow-moving, low RCS micro-UAVs. The fourth work introduces a low-cost S-band MIMO radar system designed for detecting small UAVs and employing calibration methods to improve target SNR. The fifth work proposes two methods for fast subspace computation and accurate AoA acquisition, leveraging randomized low-rank approximation for high-resolution environmental sensing. The sixth work introduces an end-to-end detection and tracking framework for multiple micro-UAVs using FMCW-MIMO radar, addressing low SNR challenges.

4.7. SPC Techniques for Noise Reduction

In [102,103,104], a Stationary Point Concentration (SPC) technique is introduced in the first work to address challenges related to phase noise in radar systems. The SPC technique reduces noise by concentrating it on a stationary point, and it can be implemented using DSP without additional hardware. In the second work, issues in heterodyne FMCW radar systems are addressed, including increased noise and unwanted Doppler shifts in the range-Doppler map. The authors introduce an enhanced SPC technique that mitigates noise and corrects Doppler shifts, improving SNR and velocity information for small moving UAVs, while in the third work, limitations in the SPC technique are identified and an A-SPC technique is introduced. The A-SPC method is designed to address the drawbacks of the SPC technique while retaining its advantages.

4.8. Radar Digital Beamforming Technology

In [105,106,107,108], Active Electronically Scanned Array (AESA) technology and its advantages are introduced in the first work for airborne radar systems, emphasizing its flexible antenna design and adaptable configuration. In the second work, a radar system for commercial UAV detection in the X-band is presented, achieving detection up to 2 km and discussing detection figures and RCS studies. The third work introduces an X-band radar system demonstrator and its signal processing, demonstrating UAV detection and tracking up to 3 km, highlighting its advantages for surveillance tasks. In the fourth paper, a new method that combines moving surveillance radar with computer vision, YOLO network, is presented to jointly detect and classify UAVs. The approach uses full spatial coverage radar data and achieves over $99\%$ in accuracy. All four works focus on radar digital beamforming technology, including radar detection and capabilities.

4.9. Full-Wave EM CAD Tools for UAV Classification

Recent research, as shown in [33,34,35,36,37,38,39,40,41], highlights that ML models can be trained on synthetic datasets and effectively applied to real-measurements datasets, resulting in successful UAV classification. Furthermore, the development of new ML models and signal processing techniques has been facilitated by leveraging full-wave EM CAD tools. For instance, in [33], the Mechanical Control-Based Machine Learning (MCML) algorithm is proposed to address the diminished classification accuracy observed when ML models are trained and tested on datasets featuring identical motions. Additionally, ref. [34] investigates the impact of antenna field of view (FOV) on UAV classification, revealing a decrease in classification accuracy with an increase in the relative angle between antennas and UAVs, providing insights into classifier performance. The exploration of multi-label classification in UAV detection and classification is initiated in [39], marking a novel investigation in the field. Furthermore, ref. [41] delves into the classification of UAVs alongside avian entities, demonstrating a decrease in classification accuracy with an increasing number of birds in a scene. Introducing a beamforming-based method termed Range-Doppler Integration while Steering (RDIwS), ref. [35] significantly enhances the signal-to-noise ratio (SNR) of UAVs at wide angles, thereby bolstering the detection and classification of UAVs. An ISAR-based classification algorithm [36] is proposed for improving the detection and classification of UAVs, specifically to accurately identify UAVs carrying explosive payloads, thereby enhancing radar security against UAV-borne threats.

The integration of full-wave EM CAD tools represents a pivotal advancement in the realm of UAV detection and classification. These tools offer a solution to the challenge of generating large volumes of complex datasets swiftly and efficiently. By leveraging full-wave EM CAD tools, researchers can simulate a wide array of scenarios encompassing diverse UAV configurations, radar parameters, and environmental conditions. This capability not only expedites the dataset generation process, but also enables the exploration of intricate interactions between UAVs and radar systems. Consequently, the utilization of full-wave EM CAD tools holds immense promise in advancing the development of robust detection and classification algorithms, ultimately enhancing the effectiveness and reliability of UAV surveillance and security systems.

5. Challenges, Future Directions, and Conclusions

While prior research efforts have made strides in addressing certain constraints within the realm of radar detection and classification of UAVs, some key aspects require attention for future research endeavors. Firstly, the utilization of real measurements in previous studies has been constrained by practical limitations and complexities, resulting in a limited number of UAVs used and reliance on radar-based digital twins for bird data. Overcoming this limitation necessitates a broader inclusion of UAVs in real-measurements, along with exploring methods to control bird flight paths for more accurate measurements. Furthermore, the datasets generated using full-wave EM CAD tools have lacked consideration for noise and clutter, thereby necessitating the integration of noise sources into simulations to obtain more realistic results. Additionally, while deterministic methods have been primarily employed, future research should incorporate statistical approaches to account for uncertainties, enhancing the validity and applicability of outcomes. Moreover, while ML methods have been compared in terms of classification accuracy, future studies should also consider factors such as memory usage, computational cost, and processing time for real-time operation. These considerations will contribute to overcoming existing limitations and advancing the field of UAV detection and classification.

In conclusion, this paper has provided a comprehensive primer aimed at researchers embarking on investigations into UAV detection and classification, with a specific focus on the integration of full-wave EM CAD tools. By elucidating radar’s pivotal role in UAV detection and systematically navigating through fundamental FMCW radar principles, methodologies, and case studies, this primer has underscored the significance of radar technology in enhancing aerial surveillance and security systems. Furthermore, the discussion on emerging trends and potential research directions emphasizes the indispensable nature of full-wave EM CAD tools in propelling radar techniques forward. As researchers continue to navigate the complex terrain of radar-based UAV detection and classification, this primer serves as an invaluable roadmap, fostering advancements and innovation in the field.