*3.1. RF-Based*

RF is used for UAV remote command and control communication. RF-based detection technologies rely on real-time sensing, capturing, processing, analyzing, and retrieving data from UAV's RF-emitted signals. Acquired RF data are intended to identify, track and classify the detected UAV and localize the controller. RF-based techniques analyze the captured spectrum between the UAV and operators using circular or linear array antennas to detect both the drone and its controller in all-weather environments. As most of the communication between a drone and its controller occurs in the ISM band, around 2.4 GHz, the implementation cost of such a system is much lower compared to a radar-based solution [21–28,51].

In RF-based detection technologies, RF and WiFi-based fingerprinting techniques are major verification systems. RF-based techniques include studying and analyzing the characteristics of the captured transmitted RF signal from UAVs or UAVs' controllers. However, WiFi-based fingerprinting is related to the WiFi links and traffic between the UAV and its remote controller. The reviewed studies include the analysis of RF spectrogram (fingerprinting) [7,9,10,19], angle of arrival (AOA) (MUSIC) [12], and direction of arrival (DOA) [40] methods for the identification and localization of drones using conventional as well machine learning algorithms [9,10,13–15].

In [7], the technique proposes a complete UAV detection and identification system framework designed to work in the 2.4 GHz frequency band. The system starts with capturing the wireless signals in the test area. Then, the captured signal is processed based on a 4-level Haar wavelet transform analysis. The standard deviation of the processed signal is calculated to define the UAV detection condition. After the detection of the UAV, the RF fingerprinting stage is activated, and three main features are extracted: (1) fractal dimension (FD), (2) square integrated bispectra (SIB), and (3) axially integrated bispectra (AIB). These features are adjusted and weighted using principal component analysis (PCA) and neighborhood component analysis (NCA) algorithms. The final RF fingerprints

are stored as the training data for a set of machine learning algorithms used to classify the UAV.

Based on indoor and outdoor experimental scenarios, the average identification accuracy of UAVs is summarized with respect to three fingerprinting features. Furthermore, in [8], the WiFi network traffic is monitored, and the UAV detection method is based on WiFi fingerprint analysis. The extracted features are related to the captured traffic's duration, behavior, and distribution. Different scenarios are applied to evaluate the system's performance in UAV detection, where the average precision is about 96%.

The authors of [15,19,20] fed the extracted time-domain characteristics (shape factor, skewness, kurtosis, and variance) of recorded RF signals to the machine learning data processing units to detect and classify UAVs. In [9], indoor experimental testing is conducted for data collection using the RF fingerprints of the transmitted signal from the micro-UAV controller to the UAV for UAV detection and classification. Different micro-UAV controllers (a total of 14) operating at the 2.4 GHz frequency band were used to create the dataset (a total of 100 RF signals) and test the proposed detection and classification technique. Each micro-UAV controller has a different transmitted signal, categorized with its unique transmitter characteristics, excluding the traditional threshold-based detection technique. The Markov model algorithm is later used for UAV detection and energy transient signal approach for feature extraction and UAV classification. The performance and accuracy of the system were found to be 96.3%.

UAVs use the Industrial Scientific and Medical (ISM) frequency bands, i.e., 2.4 GHz and 5.8 GHz bands, to communicate with their remote controllers [11]. Multiple passive RF sensors support these frequency bands and are used for non-invasive surveillance operations, including UAV monitoring, detection, localization, and tracking. In [11], the UAV detection system consists of a sensor node, Keysight RF sensor N6841A, operating in the range of 20 MHz–6 GHz, broadband antenna, and GPS tracker linked with geolocation software, N6854A. The RF signals are detected and collected within a radius of 2 km from the sensor node. A GPS antenna also records the time stamps for these collected signals. The localization of the UAV is performed using a detection algorithm and time difference of arrival (TDOA) measurements. Extended Kalman filter (EKF) framework and fitting motion models (MM) address these errors and improve localization performance.

Furthermore, the research work in [10] illustrates a system model and architecture followed by experimental validation of the proposed direction finding (DF) method of sparse de-noising auto-encoder (SDAE) for UAV surveillance. This method consists of a single channel for a receiver and a directional phased array antenna. The mechanism of the system works as follows. First, the transmitted signal from the drone to its ground controller gets processed using an RF switching mechanism to measure the received signals output power at each directional phased array antenna. Next, the acquired output power values from the N-antennas of the phased directional array are input to the proposed SDAE-based deep neural network (DNN). The first network layer extracts received wattage values. Then the remaining network utilizes sparse representation to categorize UAVs' signal directions. The system diagram of the proposed method in [10] is depicted in Figure 2. To summarize, the wattage power values are passed to the proposed deep network, followed by the DF method, which exploits both the sparsity parameter of the transmitted UAV signal and the gain variation parameters of the directional antenna array.

In [16], the authors discuss UAV detection using RF-transmitted signals between UAVs and their remote controllers. Power spectrum cancellation and multi-hop autocorrelation are developed to achieve RF passive detection of UAVs and controllers to detect emitted signals. The multi-hop autocorrelation method can detect the cross-correlation signal if the Signal-to-Noise (SNR) ratio is small by applying an emitted remote-control signal. A limitation of the multi-hop autocorrelation is the low accuracy in the case of fixed frequency in remote-control signals. The calculated parameters significantly depend on the autocorrelation function, leading to false positives. Hence, the study of [16] used the power spectrum cancellation technique to eliminate the effect of fixed frequency signals. Power spectrum cancellation works by first finding the differences between the control signal power spectrum and fixed frequency signals over time. Once the differences are identified, the fixed frequency signal is eliminated, and the remote control signal is applied to multi-hop auto-correlation to finalize the parameters for UAV detection.

Furthermore, [17] stated that the RF passive detection method has the advantage of low cost, license-free, long-range distance coverage, and early warning capability. They also illustrated an RF passive system architecture, which analyzes the electromagnetic RF spectrum emitted from exchanged signals between the UAV and its controller. The passive RF detection algorithms analyze these signals to sense alternations in the frequency and time domain RF spectrum.

Various studies have reported promising results utilizing different algorithms and techniques for RF-based UAV detection. However, the presence of noise affects the accuracy and detection range. Table 2 summarizes the reviewed papers and tabulates the features and accuracy of the undertaken methodology for RF-based UAV detection.



#### *3.2. Radar*

Radar signal processing is among the classical approach for aircraft and drone detection as it can be used in all weather conditions with 24/7 operation [18,21,52] as compared to acoustic and visual detection methods. In this approach, the received signal is characterized to detect echo, doppler signature, or radar cross-section (RCS) for detecting and tracking the target [21,28,29,53]. The conventional radar signal processing techniques have the limitation of accurate distinction of mini UAVs from birds due to their smaller RCSs. AI-based techniques are proposed [28,29,53,54] to process the extracted features from the radar signals to address this issue to some extent.

In radar-based detection, radio energy is used to detect the target and define its position [21,23,55]. Typically, a radar-based detection system has three main components: RF radar, data acquisition, and signal processing. In RF radar, the electromagnetic energy radiates into space and encounters the UAV's body flying in the monitored area. The UAV's reflected wave is returned and received by the system, measured, and processed in real-time (data acquisition and signal processing). Hence, the UAV is successfully located, and its flight path is tracked by the system [30,36,55,56].

Frequency-modulated continuous wave (FMCW) and continuous wave (CW) radars are preferred to be used in UAV detection and identification, especially for their continuous pulsing, effective cost, and performance [21]. The FMCW radar contains a transmitter and a receiver antenna. The oscillator and the control signal produce the transmitted signal. After the backscattering/reflected signal is received, it gets passed to the I/Q demodulator for filtering. Power is equally distributed into two signals with 90 degrees phase shift to be forwarded to the low pass filter (LPF). The intermediate frequency (IF) signal, resulting from in-phase and quadrature-phase components, is directed to the analog-to-digital converter (ADC) and the digital signal processing (DSP), as depicted in Figure 3. The distance and velocity of the target can be defined by using the time delay and phase information of both the transmitted and received signals [21].

The studies of [24,25,28,54,57,58] employed the principal component analysis (PCA) [24], convolutional neural networks (CNN) [23,28,51,54], long short-term memory (LSTM) [28], and support vector machines (SVM) [57,58] techniques for the processing of extracted features from radar signals such as micro-doppler spectrogram [23,28,54,57,58] and rangedoppler signature [24] for the classification of drones. Recently authors in [13] used the hierarchical learning approach for the detection of the presence, type, and flight trajectory of a UAV. Due to the smaller size of most UAVs, wideband, high frequency, expensive radars are required for the accurate detection and tracking of mini UAVs [23,24,52,54], which increases the overall cost of the detection and localization system.

**Figure 3.** Architecture design of FMCW radar [21].

In [55], the authors proposed research and experiments for evaluating the data acquisition and signal processing algorithm in a CW radar system that supports C and X frequency bands operations. The radar system uses the micro-Doppler principle. The extracted signatures in the frequency and time domains are used in UAV classification for calculating the propeller blades' length and determining the rotation propellers' speed. For performance evaluation, the number of UAV propellers varies during the experiments while fixing the propellers' rotational speed and a maximum distance of 25 m between the radar and UAV. The classification and measurement of UAVs become complex with the increase in propellers.

In another work [56], simulation and analysis of continuous wave radar's echo signals are studied and presented in different conditions at an operating frequency of 35 GHz. Mainly UAV detection is based on the time-frequency characteristics of the Micro Doppler signal produced by the rotor rotation using singular value decomposition. Discrete wavelet transform is also used to remove environmental clutter from the radar echo signal, whereas the support vector machine (SVM) is used as a classifier. The detection accuracy of the developed system achieved 85%.

Another type of radar-based UAV detection mechanism, cylindrical phased array radar, was discussed in [27]. The system performs better for UAV detection when comparing the omnidirectional scanning to planner array radar due to the flexibility of changing the direction of the beam and illumination time to the target after the phased array was used. As for the operational norms, the system's hardware structure and signal processing flow are designed to get a strong clutter suppression specified in the investigation, and the result of the experiment shows potential for UAV detection. Authors in [27] developed a cylindrical phased array radar system and explored signal optimization by specifying signal processing flow with the moving target detection (MTD) based on the maximum signal-to-clutter ratio (SCR) criterion.

Tang et al. [24] explained the type x-band, a small phased array radar based on AD9361, an RF Agile Transceiver. The AD9361 is a highly integrated RF module with a high-performance agile transceiver for 3G and 4G base station applications. The reported radar system consists of a control module controlling the antenna beam pointing through the transmitter/receiver (T/R) module. The signal processor also sends waveforms as transmitted RF signals to AD9361 within the timing sequence. Then the corresponding waveform is generated and established by AD9361. The radar simulation detects a drone with a radar cross section (RCS) of 0.01 m−<sup>2</sup> within the range of 5 km. For radar detection, enhanced reflected signals are necessary to minimize the effect of noise. An SNR value greater than 14 dB indicates a highly accurate detection.

In the case of a reliable RCS, the chosen wavelength should not reach half of the detected object's dimension. It is critical to use a higher frequency while using Dopplerbased detection. As illustrated in [22], radar is used to detect smaller drones; however, it has an ill-prepared standard for UAV detection based on low air-velocity aircraft and weak radar signature. During target detection, the radars would receive reflections from clutterlike objects, landscapes, and precipitation, posing a challenge in detection. A target can only be detected if system noise due to clutter is minimized. A 30 × 30 rectangular phase array used in [22] detects the presence of drones in monostatic radar. It would continuously scan the predefined surveillance region, with the limitation of a 90-degree azimuth sector, to achieve 360 azimuth coverage at a low cost. Doppler estimation discussed in [22] can be described as a spectrum estimation process.

The reference [26] illustrates the new method based on 5G millimeter waves with an end-to-end network. It further explains the detection method done using 5G millimeterwave radar at rotors of UAVs. The high-resolution range profile (HRRP) can identify a UAV location, while micro-Doppler identifies the UAV. Moreover, the cepstrum method was used to extract any number and speed information of the detected UAV rotor. Multiple UAVs can be identified using the sinusoidal frequency modulation (SFM) parameter optimization method. The proposed method determines the following: the number of detected UAVs, the number of rotors, the rotation speed of all rotors, and the position of the UAVs. The proposed radar detection in [26] presents a UAV identification and detection study by providing a method for UAV tracking using the GPS-independent method, such as GPS signal failure, GPS signal interference, and satellite occlusion areas. HRRP technology and micro-Doppler provide a successful solution to detect and localize any rotating targets regardless of weather conditions. The presented simulated results showed high robustness and performance of the cepstrum method.

Authors in [59] presented a passive radio drone detection system that uses goodnessof-fit (GoF) based spectrum sensing and the MUSIC algorithm to detect the transmitted signal of a drone and its controller and estimate the DOA. Once a signal is detected, the DOA is estimated at the detected frequency. The MDL algorithm detects the number of targets and whether the source is a drone or controller. The detection system detected drones and controllers from different manufacturers with good sensitivity.

A challenge associated with UAV detection is the presence of aircraft and birds in the background [60–62]. Hence, clutter suppression and target detection algorithms are needed to overcome this complex issue, as stated in [26]. Rationally, object detection of possible UAVs comes first, followed by classification to separate UAVs from other detected objects. In addition, the purpose of these classifications and identifications can be used to extract many unique features of these UAVs [23]. As stated previously, the effects of Doppler radar are used to determine the velocity of a distant object more accurately. This is obtained from the radial component of a target velocity in relation to radar. Using stepped frequency waveform (SFW), an HRRP can be obtained. Due to HRRP and Doppler information from a wide-band Doppler radar, detected objects scanned using wide-band are identified and classified. Millimeter wave base stations and 5G network systems can be used as detection network channels for UAV detection using the data from the processing center of 5G base stations. The process includes extracting essential parameters from multipath locations through 5G bases.

Many factors must be considered during the development to enhance the radar systems' performance, such as operating frequency, data acquisition, processing algorithms, classification techniques, and environmental clutter. The summary of reviewed studies of this technique is given in Table 3.


**Table 3.** Summary of reviewed radar-based techniques for UAV characterization.

#### *3.3. Acoustic*

Acoustic sensors, such as microphone arrays, capture the generated audio from the rotors and propellers of the drone and then compare the extracted features, including mel- frequency cepstral coefficients (MFCC) and short-time Fourier transform (STFT), with acoustic signature databases for the detection and classification of drones and UAVs using conventional and AI-based architectures. MFCC is a set of reflected human perception features of sounds, which is used in audio classification when paired with machine learning approaches. STFT is considered an intermediate feature compared to MFCC. MFCC compresses signals while representing them with coefficients set. On the other hand, STFT features contain more information and noise than MFCC, giving STFT an advantage. Deep learning models can easily adopt STFT and manage it given more complex data [37].

Authors proposed a machine learning framework in [38], shown in Figure 4, to detect and classify ADr sounds in a noisy environment, among other sounds. The required features are extracted from ADr sound using the feature extraction techniques of MFCC and linear predictive cepstral coefficients (LPCC). Following the feature extraction process, these sounds are then identified using SVMs. The results show that the SVM cubic kernel with MFCC outperforms the LPCC technique by detecting ADr sounds with 96.7% accuracy.

Acoustic-based technologies are effective for detecting UAVs since they are not affected by the UAV's frequency range, weather fluctuations, e.g., fog, environmental disturbance, and noise. Hence, such technologies do not block the acoustic sensors' earshot to detect the UAV's acoustic signals. Acoustic signals produced by the engine and propeller blades of the UAV are collected and processed to classify the UAV and calculate its distance, direction, and location [33].

Authors in [39] proposed a CNN-based system to detect drones using acoustic signals received by a microphone. STFT magnitude is used as the two-dimensional feature in the study since drones' harmonic properties differ from those of other devices that make a similar noise. The dataset comprised 68,931 and 41,958 frames of drone and non-drone sounds collected using DJI Phantom 3 and 4 drones flying outdoors. The proposed approach has a detection rate of 98.97% for the 100-epoch model and a false alarm rate of 1.28. Figure 5 illustrates the system overview of the proposed approach.

**Figure 4.** Overall system diagram of the approach presented in [38].

**Figure 5.** System overview of the method introduced in [39].

An acoustic-based detection system was designed and implemented in [30] to detect and locate the UAVs efficiently. The acoustic sensor array configuration comprises two tetrahedron-shaped microphones. The system uses multiple algorithms for data and features extraction from the collected acoustic signals: cepstral coefficients (CC) for extracting the harmonics' features, SVM to classify and distinguish between the extracted features' vectors related to UAV or background noise, and TDOA based on Bayesian framework. Signal processing is concluded with the temporal and dimensional features' vectors calculations to acquire the accurate UAV classification and localization path. They also study the contribution of the SNR in detecting the UAVs against the detection rate.

In [36], multi-label UAV sound classification is examined using stacked bidirectional long short-term memory (BiLSTM), an advanced, recurrent neural network (RNN) capable of handling sequence or multiple classification tasks and avoiding long-term dependency issues. The proposed BiLSTM model is 94.02% successful in UAVs' sound classification.

Several types of research and studies aim to investigate and evaluate different algorithms used in data acquisition, processing, and classification of the collected acoustic signals. In [31], the system's performance level varies using different audio processing algorithms for characteristic feature vector extraction. The extracted features are inputted

into concurrent neural network (CoNN) for classification. The results confirm better accuracy when integrating CoNN with the Wigner-Ville dictionary than MFCC and mean instantaneous frequency (MIF).

The authors in [35] gathered the acoustic data from a local suburban airport for the five samples of commercial multirotor UAVs to establish the performance based on passive acoustic detection. The study characterizes the emitted noise of UAVs of different levels in an anechoic chamber at the airborne time. The microphone array was arranged within two circular tiers, each 1-m in radius, and separated vertically by 1.6 m to collect data from the local airports. The generalized cross-correlation (GCC)-based algorithm is used to find direction by fusing the time difference of both arrivals and steered power response with phase transform (SRP-PHAT). The smallest UAV with a 294 m detection distance was tested and demonstrated. Differential Doppler is used to overcome the decorrelation effect for better accuracy, as stated in [35].

In [32], the authors used classical detection and direction-finding methods using an array of microphones. There had been a physical investigation of the UAVs through experiments on acoustic emission with two signal models presented in harmonic signal and broadband signal for open area and indoor environments, respectively. The spectral signs are used for detecting and recognizing the UAVs in a noisy environment by incorporating the effect of noises in urban transport, speech signals, and environment noises. The result gives the same quality as the MFCC method, where acoustic portraits are unnecessary. The cross-correlation function is efficient in the direction-finding of the UAV. The study of [32] concludes with the following points: (1) high-pass filters are effective in the processing stage of UAV acoustic emission; (2) taking a noisy environment as a background experiment while detecting and recognizing UAVs by spectral signs performs similarly to the MFCC method, excluding acoustic portraits; (3) it is suggested to improve the efficiency of the CCFM algorithm in acoustic signals to filter out low-frequency noise; (4) MFCC and CCFM can be used to create an effective counter-action system against UAVs.

Yang et al. [37] researched the utilization of acoustic nodes in the UAV detection system. The proposed system finds the best configuration of the node for deploying the UAV acoustic detection system using machine learning models. The study was designed to investigate the best combination of acoustic features, STFT and MFCC, machine learning algorithms, SVM and CNN, for node optimization. After integrating the sensing nodes in four different configurations among the test sets, the one that maximizes the detection range without blind spots is selected. A semi-circle by the STFT-SVM model with a 75-m distance between the protected area and node has the best performance for configuration optimization. Demonstrating machine learning in the audio signal domain with different learning algorithms was used for detection module development. The study [37] focused on event sound detection using binary classification with MFCC features in an urban area.

A drone acoustic detection system (DADS) is proposed and demonstrated experimentally to detect, classify, and track airborne objects in [33]. They used a Phantom 4 UAV for testing, which reached 350 m with four degrees as an average precision to track a maneuvering UAV with compact acoustic nodes. This test also implemented the classification algorithm to detect a multirotor UAV based on a specific sound inherent in the flight control mechanism. The Steven Institute of Technology has developed the DADS to detect, track, and classify anonymous UAVs by propeller noise. The proposed system has three or more microphone nodes in a tetrahedron configuration. The communication between the microphone nodes and the central computer is done through WiFi for processing. The orientation calibration for the DADS system is performed by emitting white noise from a speaker and tracking the GPS position for several minutes. Based on the difference between the detected direction and computed ones from the surveyed GPS, the orientation can easily be corrected in the case of detection and tracking. Establishing a tracking process can be predicted using collected data and parameters. Node placement, the direction-finding probability that depends on precision and range for a given target, and ambient conditions with the tracker association threshold are among the collected data.

Two main components affecting the system's overall performance are (1) hardware specifications, including acoustic sensors and data acquisition tools, and (2) software tools and algorithms, including acoustic fingerprints and features extraction, classification, and localization. Table 4 summarizes the recent studies for acoustic-based UAV detection, classification, and localization.


**Table 4.** Summary of reviewed acoustic-based techniques for UAV characterization.

Unlike radar and RF approaches, the acoustic solution does not require a line of sight (LOS). However, this solution has challenges of a short range, the need for an extensive large signature database, and vulnerability to ambient environmental noise and clutters, particularly in urban areas [14,30,31,40], and quiet operation of the drone [9,30,38]. The detection of the drone pilot could be very difficult, too, using acoustic sensors.
