An Intelligent Passive System for UAV Detection and Identification in Complex Electromagnetic Environments via Deep Learning

Abstract

With the rapid proliferation of unmanned aerial vehicles (UAVs) and the associated rise in security concerns, there is a growing demand for robust detection and identification systems capable of operating reliably in complex electromagnetic environments. To address this challenge, this paper proposes a deep learning-based passive UAV detection and identification system leveraging radio frequency (RF) spectrograms. The system employs a high-resolution RF front-end comprising a multi-beam directional antenna and a wideband spectrum analyzer to scan the target airspace and capture UAV signals with enhanced spatial and spectral granularity. A YOLO-based detection module is then used to extract frequency hopping signal (FHS) regions from the spectrogram, which are subsequently classified by a convolutional neural network (CNN) to identify specific UAV models. Extensive measurements are carried out in both line-of-sight (LoS) and non-line-of-sight (NLoS) urban environments. The proposed system achieves over 96% accuracy in both detection and identification under LoS conditions. In NLoS conditions, it improves the identification accuracy by more than 15% compared with conventional full-spectrum CNN-based methods. These results validate the system’s robustness, real-time responsiveness, and strong practical applicability.

1. Introduction

Once confined to military applications, unmanned aerial vehicles (UAVs) are now transforming a wide range of civilian fields. These include logistics [1,2], agriculture [3], surveillance [4], and emergency response [5,6]. Recent research on UAV-assisted communications and spectrum mapping has further highlighted their value in low-altitude networks, e.g., UAV-assisted informative path planning for 3D spectrum mapping [7] and sparse Bayesian learning-based 3D radio environment construction [8]. However, alongside these benefits, the malicious or unauthorized use of UAVs has emerged as a growing security threat [9], posing serious risks to public safety, personal privacy, and the operational continuity of critical infrastructure. Several high profile incidents have exemplified these challenges, such as the 2018 disruption at Gatwick Airport, the 2019 interference at Newark Liberty International Airport, and repeated low-altitude intrusions near Shanghai Pudong International Airport in China. In response to these threats, regulatory initiatives such as the mandated Remote ID policy in the United States [10] have further stimulated the development of advanced UAV identification technologies [11]. These incidents underscore the urgent need for robust, reliable, and real-time UAV detection and identification systems to ensure the security and orderly management of low-altitude airspace.

To address the challenges introduced by unauthorized UAVs, numerous detection and identification technologies have been proposed in recent studies, including radar sensing [12,13], optical imaging [14,15], and acoustic sensing [16,17]. A summary of the advantages and limitations of these methods is provided in Table 1. For instance, a distributed K-band radar system capable of high-sensitivity detection and velocity estimation of small UAVs was introduced in [18]. An empirical mode decomposition-based method was developed for radar-based automatic multicategory UAV classification, demonstrating promising accuracy on real measurement data [19]. Reference [20] proposed the IRSDD-YOLOv5 framework, which incorporates an infrared small target detection module, a customized prediction head, and optical feature enhancement strategies, demonstrating superior performance. A residual image prediction model based on the U-Net architecture was presented in [21] for infrared UAV detection. In addition, a distributed system integrating a wireless acoustic sensor network with machine learning algorithms was introduced in [22] to detect and localize unauthorized UAVs. Although each method performs well under specific conditions, they exhibit inherent limitations. Radar systems often miss small UAVs due to their low radar cross-section. Optical techniques are highly sensitive to illumination, occlusion, and adverse weather. Acoustic sensing, meanwhile, is vulnerable to environmental noise and suffers from limited detection range.

In contrast, communication-assisted sensing, particularly radio frequency (RF) sensing, identifies UAVs by passively monitoring their communication signals, as illustrated in Figure 1. This approach offers several advantages, including all-weather applicability, non-line-of-sight (NLoS) capability, and ease of deployment. Recently, integrated sensing and communication (ISAC), including resource-efficient UAV systems and near-field modeling for reconfigurable intelligent surface-assisted air-to-ground links, has attracted significant research attention [23,24]. In particular, authors in [25] proposed a reinforcement learning-based ISAC framework for low-altitude UAV detection, achieving adaptive parameter optimization for robust detection under multipath and low-signal-to-noise ratio conditions. Similarly, an orthogonal time–frequency and space-based ISAC system that exploits micro-Doppler signatures to distinguish UAV echoes from static clutter, demonstrating high detection accuracy with a software-defined radio prototype, was proposed in [26].

A core technique in this domain is RF fingerprinting, which enables device identification by exploiting unique signal characteristics [27]. Commonly used RF fingerprint features include wavelet coefficients [28], I/Q imbalance parameters [29], fractal dimensions [30], carrier frequency offsets [31], bispectral features [32], and transient-based signal characteristics [33]. These inherent characteristics give rise to distinctive signal traits that can be exploited for UAV identification. Reference [34] proposed a method for detecting and identifying small UAVs by extracting cyclostationarity, kurtosis, and spectral features from downlink RF signals. A dual-source RF detection system was presented in [35], enabling simultaneous identification of both flight control and video transmission signals (VTSs). In [36], RF fingerprinting was combined with machine learning algorithms to classify UAV types, achieving high identification accuracy. Meanwhile, UAV communication-specific channel measurements and clustering methods have been investigated to support RF-assisted identification, such as space clustering based on full-domain channel characteristics [37] and comprehensive surveys on UAV channel sounding technologies [38]. Moreover, frequency-hopping signals (FHSs), commonly used in UAV communications, exhibit distinctive time–frequency characteristics. These spectrogram patterns have attracted increasing research interest in spectrogram-based UAV detection and identification methods. For example, an improved YOLO-based algorithm was proposed in [39] to identify UAV models using FHS spectrogram features.

Table 1. Advantages and disadvantages of different UAV detection methods.

Method	Advantages	Disadvantages
Radar-based method [12,13,18,19]	High performance	Difficult to detect small UAVs/High cost
Optical-based method [14,15,20,21]	Provides intuitive results	Short detection range/Susceptible to weather
Acoustic-based method [16,17,22]	Low cost	Short detection range/Easily affected by noise
RF-based method [34,35,36,39]	Low cost/Long range	Easily affected by interference

Table 1. Advantages and disadvantages of different UAV detection methods.

Method	Advantages	Disadvantages
Radar-based method [12,13,18,19]	High performance	Difficult to detect small UAVs/High cost
Optical-based method [14,15,20,21]	Provides intuitive results	Short detection range/Susceptible to weather
Acoustic-based method [16,17,22]	Low cost	Short detection range/Easily affected by noise
RF-based method [34,35,36,39]	Low cost/Long range	Easily affected by interference

In general, communication-assisted methods for UAV detection and identification offer several notable advantages. However, these approaches remain highly susceptible to interference from other electromagnetic signals, which severely constrains their detection accuracy and reliability in complex environments. To overcome these limitations, this paper develops a passive UAV detection and identification system that combines a high-resolution RF front-end with a spectrogram-based deep learning framework. Specifically, the YOLO network is first utilized to locate and extract the regions of frequency-hopping UAV signals from the spectrogram, thereby suppressing the influence of environmental noise and interference. Subsequently, a CNN classifier is employed to identify the UAV type based on the extracted signal regions. This two-stage design effectively decouples detection and identification, enabling robust and accurate performance under both line-of-sight (LoS) and NLoS conditions. The main innovations of this work are summarized as follows:

• A passive UAV detection and identification system is proposed, with a modular architecture comprising a multi-beam directional antenna, a wideband spectrum analyzer, and a computing unit. The system enables real-time acquisition, processing, and identification of RF signals and is specifically designed to ensure robust operation in complex electromagnetic environments.
• A two-stage deep learning framework is proposed, which integrates a YOLO-based detector for FHSs with a CNN-based UAV model classifier. The YOLO module enables rapid and accurate localization of FHS regions in spectrograms, while the CNN extracts discriminative features for precise UAV model identification, thereby enhancing both accuracy and robustness under severe interference conditions.
• Extensive measurements are conducted in diverse LoS and NLoS conditions across a campus. The results showed that the proposed system achieved over 96% detection and identification accuracy under LoS conditions and improved identification accuracy by more than 15% in NLoS conditions compared with conventional full-spectrogram classification methods.

The remainder of this paper is organized as follows: Section 2 introduces the RF spectrogram-based system framework for passive UAV detection and identification system. Section 3 describes the hardware architecture and algorithm design of the proposed system. Section 4 presents experimental validation and analysis under both LoS and NLoS conditions. Finally, conclusions are drawn in Section 5.

2. Spectrogram-Based Framework for Passive UAV Detection and Identification

2.1. Characteristics of Frequency Hopping and Video Transmission Signals

UAV communication links are generally categorized into two types. The first employs a dual-link configuration, in which FHSs is used for flight control, while an Orthogonal Frequency-Division Multiplexing (OFDM) signal is employed for video transmission. The second approach integrates both control and video data into a single OFDM-based communication link. However, this unified scheme suffers from reduced anti-jamming capability and a higher risk of link disruption in the presence of interference. Consequently, most modern UAV systems are designed with a dual-link architecture, leveraging FHS for robust control and OFDM for high-throughput video transmission.

The FHS of a UAV can be modeled as follows:

$$x_{FHS}\left(t\right)=A\sum _{k=0}^{N_h-1}{W}_T(t-k{T}_h)cos[2\pi f_k(t-k{T}_h)+{\phi }_n]$$

(1)

where $N_h$ denotes the number of frequency hopping points, A is the signal amplitude, $W_T(·)$ is the rectangular window function defined over the hopping period $T_h$, $f_k$ represents the frequency at the k-th hopping point, and ${\phi }_n$ is the initial phase. In the time–frequency domain, this signal appears in the spectrogram as a series of periodically spaced, stripe-like patterns with abrupt transitions in frequency. These pronounced structural features provide strong discriminative cues, making FHSs especially suitable for passive UAV detection and identification tasks.

Regarding video transmission, modern commercial UAVs such as those in the DJI (Shenzhen, China) series typically employ advanced OFDM-based communication systems with bandwidths of 10 MHz or 20 MHz. These systems are capable of dynamically selecting communication frequency bands to maintain reliable data throughput under diverse operating conditions. Although this mechanism improves transmission efficiency, it also increases signal diversity and complexity, which poses significant challenges for accurate UAV model identification. The OFDM signal used for video transmission can be mathematically described as the superposition of multiple modulated subcarriers. Assuming the transmission begins at time $t=t_s$, the signal can be represented as follows:

$$x_{ODFM}=\left\{\begin{array}{c}{\displaystyle \sum _{i=0}^{N-1}}{d}_i\mathrm{rect}\left(t-t_s-{\displaystyle \frac{T}{2}}\right)exp\left[j2\pi \left(f_c+{\displaystyle \frac{i}{T}}\right)\left(t-t_s\right)\right],\begin{array}{cc}& t_s\le t\le t_s+T\end{array}\\ 0,\begin{array}{cc}& t\le t_s{t}_s\ge t_s+T\end{array}\end{array}\right.$$

(2)

where $N_s$ denotes the number of subcarriers, $d_i$ is the complex modulated symbol on the i-th subcarrier, $f_c$ is the frequency of the first subcarrier, and T represents the OFDM symbol duration. The rectangular window function $\mathrm{rect}\left(·\right)$ is defined as follows:

$$\mathrm{rect}\left(t\right)=\left\{\begin{array}{c}1, \left|\tau \right|\le T/2\\ 0, \mathrm{otherwise}\end{array}\right.$$

(3)

In the spectrogram, the OFDM signal appears as a continuous and wideband structure with high temporal stability and consistent bandwidth.

To further investigate the radio frequency characteristics of the UAV communication signal in practical deployment scenarios, a representative RF spectrogram dataset is constructed using real measurement data from an open-source UAV signal repository [40]. All signal samples are collected in a microwave anechoic chamber under standardized experimental conditions, ensuring a high signal-to-noise ratio and reliable measurement fidelity. The two-dimensional spectrograms of six representative commercial UAVs operating under typical communication conditions are presented in Figure 2. The FHSs display substantial variation in their spectral patterns across different UAV platforms, which reflects differences in hopping algorithms, bandwidth allocation, and modulation schemes. In contrast, the VTSs exhibit uniform OFDM structures that closely resemble standard Wi-Fi signals in both spectral shape and bandwidth occupancy. These observations underscore the superior discriminative value of the FHS over the OFDM-based VTS for passive UAV detection and identification. The distinct time–frequency patterns of the FHS facilitate more robust detection and identification, particularly in complex electromagnetic environments where interference and overlapping signals are prevalent.

2.2. Architecture of the Passive UAV Detection and Identification System

For efficient passive detection and identification of UAV targets, this paper proposes a deep learning-based system that processes RF spectrograms. As illustrated in Figure 3, the system employs a modular framework consisting of three primary components: signal acquisition, UAV signal detection, and UAV type identification.

The raw radio frequency signal is first captured as in-phase and quadrature (IQ) samples in complex electromagnetic environments, where the received data contains both UAV signals and coexisting interference sources such as Wi-Fi, Bluetooth, and other ambient RF signals. These in-phase and quadrature components are subsequently combined to construct a complex-valued time-domain signal:

$$\mathrm{IQData}\left(n\right)=I\left(n\right)+jQ\left(n\right)$$

(4)

where $I\left(n\right)$ and $Q\left(n\right)$ denote the in-phase and quadrature components at time index n, respectively. To reduce spectral leakage during frequency analysis, a Hanning window is applied:

$$w\left(n\right)=0.5\left(1-cos\left({\displaystyle \frac{2\pi n}{N-1}}\right)\right),n=0,1,2,\dots ,N-1$$

(5)

where N is the signal length. The windowed IQ data is then computed as follows:

$${\mathrm{IQData}}_{\mathrm{windowed}}\left(n\right)=\mathrm{IQData}\left(n\right)·w\left(n\right)$$

(6)

The time–frequency representation of the signal is obtained using the short-time Fourier transform, defined as follows:

$$X(t,f)={\int }_{-\infty }^{\infty }x\left(\tau \right)w(\tau -t)e^{-j2\pi f\tau }d\tau$$

(7)

where $x\left(\tau \right)$ is the time domain signal, $w(\tau -t)$ is the window function, and $X(t,f)$ denotes the complex spectrogram at time t and frequency f. This transformation retains the temporal–frequency distribution of the energy of the signal, effectively highlighting structural patterns such as frequency-hopping trajectories. For practical implementation, the discrete Fourier transform is computed via a fast Fourier transform:

$$Y\left(f\right)=\sum _{n=0}^{N-1}{\mathrm{IQData}}_{\mathrm{windowed}}\left(n\right)e^{-j2\pi fn/N}$$

(8)

where $Y\left(f\right)$ represents the frequency-domain signal at frequency f, and N is the number of data points. The power spectral density (PSD) is then obtained by squaring the magnitude of the FFT result:

$$P\left(f\right)={\left|{\displaystyle \frac{Y\left(f\right)}{N}}\right|}^2$$

(9)

Finally, to convert the PSD to decibels (dBm), the following equation is used:

$$P_{\mathrm{dBm}}\left(f\right)=10{log}_{10}\left({\displaystyle \frac{P\left(f\right)}{0.001}}\right)$$

(10)

where $0.001$ represents the reference power (1 mW). To ensure consistent input quality and enhance the robustness of the deep learning model, the generated spectrograms undergo image normalization and augmentation.

The UAV detection module analyzes the preprocessed spectrograms using a YOLO-based detection framework to localize candidate regions corresponding to FHS. By generating accurate binary decisions on the presence of UAV-related signals, this module enables rapid and reliable screening of complex spectrogram data. In doing so, it significantly alleviates the computational burden on downstream identification processes, while preserving high detection accuracy and real-time responsiveness.

Figure 2. Spectrum of FHSs and VTSs from different UAV models: (a) DJI Inspire 2. (b) DJI Matrice 100. (c) DJI Matrice 210. (d) DJI Mavic Pro. (e) DJI Phantom 4. (f) DJI Phantom 4 Pro Plus.

Figure 2. Spectrum of FHSs and VTSs from different UAV models: (a) DJI Inspire 2. (b) DJI Matrice 100. (c) DJI Matrice 210. (d) DJI Mavic Pro. (e) DJI Phantom 4. (f) DJI Phantom 4 Pro Plus.

Figure 3. Architecture of the proposed deep learning-based UAV detection and identification framework based on RF spectrograms.

Figure 3. Architecture of the proposed deep learning-based UAV detection and identification framework based on RF spectrograms.

Upon detection of a valid UAV signal, the localized FHS region is forwarded to the UAV identification module. To mitigate the influence of signal power variations, the FHS region is first converted into a binary image representation. A block-based detection algorithm is then applied to extract the centroids of individual hopping burst. Each hopping signal is thus represented as a set of two dimensional coordinates in the time–frequency domain, collectively forming a frequency-hopping point set. This point set is subsequently fed into a CNN-based classifier that exploits the spatial distribution of hopping patterns for UAV identification. By focusing on the geometric arrangement of hopping bursts rather than global spectrogram textures, the system achieves more robust and accurate UAV identification under complex electromagnetic conditions.

3. System Hardware and Algorithm Design

3.1. Hardware Platform

To satisfy the stringent requirements of high sensitivity, real-world deployability, and scalability, the proposed platform integrates high-performance commercial off-the-shelf hardware with custom-designed components to form a comprehensive RF signal acquisition and processing system. As illustrated in Figure 4, the hardware architecture consists of a multi-beam directional antenna for spatially selective signal reception, a wideband spectrum analyzer for real-time RF signal monitoring and digitization, and a local computing unit for signal decoding and spectrogram generation. This integrated design ensures accurate signal capture and efficient processing, thereby providing a reliable foundation for passive UAV detection and identification in complex electromagnetic environments.

A multi-beam directional antenna with six directional elements is designed to enhance the directional sensing capability of the UAV detection and identification system. The array provides full 360° horizontal coverage and employs high-gain, narrow-beam antennas to ensure strong spatial selectivity. The spacing and angular placement of the antennas are carefully optimized to achieve a balance between high-resolution multi-target detection in short-range scenarios and the structural compactness required for mobile and portable deployment.To facilitate high fidelity spectrum acquisition, the system employs the Spectran V6 PLUS spectrum analyzer developed by AARONIA GmbH (Euscheid, Germany). This advanced receiver supports an ultra-wide frequency range from 9 kHz to 8 GHz, with excellent dynamic range and high-speed sampling capabilities. Its maximum instantaneous scan rate of 1 THz/s effectively covers the primary UAV control bands at 2.4 GHz and 5.8 GHz. These characteristics make it particularly suitable for real-time monitoring of FHSs. The analyzer incorporates a built-in real-time spectrum analysis engine, enabling precise capture of short duration, burst like spectral transitions. This capability is critical for detecting dynamic hopping sequences and rapidly switching communication channels commonly used in UAV systems. Owing to its high performance and reliability, the Spectran V6 PLUS has been widely used in radio surveillance and counter-UAV applications, thus providing a robust foundation for real-world deployment. The detailed technical specifications are summarized in

Table 2. Spectran V6 PLUS Parameters.

Parameters	Value
Real-time Receiving Bandwidth	Max 245 MHz
Real-time Transmitting Bandwidth	120 MHz
Minimum Detectable Signal Interval	Minimum 10 ns
Maximum Receiving Power	+23 dBm
Maximum Transmitting Power	+20 dBm
Internal Amplifier	−170 dBm/Hz
Amplitude Accuracy	Typical ±0.5 dB
Frequency Reference Accuracy	0.5 ppm
Resolution Bandwidth	62 MHz to 200 MHz
Attenuator Adjustment Range	50 dB/70 dB (0.5 dB steps)
FPGA Model	XC7A200T-2
GPS Synchronization	Timestamp precision ±10 ns

. On the data processing side, the system integrates a high-performance local computing terminal, specifically a laptop equipped with an NVIDIA GeForce RTX 4060 GPU (Santa Clara, CA, USA). This terminal manages the entire pipeline from signal acquisition to spectrogram generation. It connects to the spectrum analyzer via a cloud-based communication protocol provided by the AARONIA device and runs custom MATLAB 2024 scripts for real-time IQ data acquisition, buffering, and decoding.

3.2. Algorithm Design

The detection and identification framework developed in this paper consists of two core modules. The first is a YOLO based detection module that automatically localizes UAV frequency-hopping regions in spectrograms. The second is a CNN-based identification module that further identifies UAV models from the detected regions. These two modules operate sequentially in a process of localization and identification, providing an efficient and automated solution for UAV detection and identification in challenging environments.

In the hopping signal detection stage, the YOLOv3 architecture is employed and customized to accommodate the unique characteristics of RF spectrograms. The choice of YOLOv3 is motivated by its favorable trade-off between detection accuracy and processing speed, making it particularly suitable for real-time applications with constrained computational resources. As shown in Figure 5, the network consists of three main components: feature extraction, feature fusion, and multi-scale prediction. The input to the YOLO-based detection network is the time–frequency spectrogram generated from the raw IQ RF signals after a short-time Fourier transform. A deep-feature extraction module, based on the Darknet backbone, is employed to learn local textures and hopping patterns. By leveraging its strong encoding capability, the backbone effectively captures weak yet critical time–frequency signatures. The extracted features are then passed into a feature pyramid fusion module, which incorporates upsampling, residual connections, and multi-scale pooling. Specifically, a spatial pyramid pooling structure with three max-pooling layers at different scales expands the receptive field, improving adaptability to local variations and scale changes. The multi-path fusion strategy ensures effective information flow between high-level semantic features and low-level spatial details, thereby enhancing detection robustness across various hopping signal sizes. Finally, three detection heads predict small, medium, and large hopping regions by regressing bounding box coordinates and confidence scores, while simultaneously classifying UAV control signals.

Traditional spectrogram identification methods typically treat the entire spectrogram as input to a deep neural network, directly learning spatial textures and energy distributions [39]. However, in real-world scenarios, spectrograms often contain substantial non-target noise and overlapping signals, which increases the risk of overfitting to background patterns and leads to degraded performance under non-ideal conditions. To address these challenges, this study introduces a specialized CNN-based UAV model identification module, as shown in Figure 6. The CNN classifier employs a five-layer convolutional architecture consisting of three convolutional layers followed by two fully connected layers. Each convolutional layer utilizes 3 × 3 kernels with ReLU activation and max-pooling operations to extract hierarchical spatial features, while the final softmax layer performs multi-class classification of UAV types. Hyperparameters are optimized through a combination of grid search and validation-based fine-tuning. Unlike conventional methods, the proposed method uses the coordinate-based structural information of the FHS as the primary input feature. Inspired by keypoint-based geometric relationship modeling in face identification, this approach overcomes the limitations of traditional spectrogram-based methods, which heavily rely on global texture distributions and have weak interpret ability. For the CNN-based identification module, the input consists of coordinate matrices representing the spatial distribution of the FHS in the time–frequency domain. Each spectrogram is first processed by a block-based detection algorithm to extract the two-dimensional coordinates of FHS points. These coordinates are then normalized with respect to time and frequency scales and structurally encoded to preserve the geometric relationships among hopping trajectories. The network not only considers the absolute positions of individual points but also captures their spatial relationships, such as hopping intervals, trajectory patterns, relative frequencies, and density distributions. The extracted structural features are fused through multi-layer fully connected layers and passed to a classifier that outputs the UAV model label. By incorporating position normalization and structural encoding mechanisms, the model learns an invariant and discriminative structural representation, thereby significantly improving identification accuracy.

4. Experimental Validation and Analysis

4.1. Measurement Setup

To validate the proposed system under realistic conditions, a comprehensive training and testing dataset is constructed by combining a public UAV spectrogram dataset with custom-measured UAV spectrograms. In addition to six mainstream commercial UAV models from an open-source RF signal database [40], two representative UAV types are selected for controlled experiments to evaluate the model’s robustness under practical conditions. The first UAV is an assembled platform, as shown in Figure 7a, equipped with a 2.4 GHz remote control link that employs a typical frequency-hopping communication mechanism. This configuration offers high protocol flexibility and a fully controllable signal structure, enabling in-depth analysis of hopping behavior under different configurations.The second UAV is a commercial DJI Mini, as shown in Figure 7b, whose control link also operates primarily in the 2.4 GHz band but utilizes a proprietary closed-source communication protocol with a more complex and stable frequency-hopping pattern. These two UAVs represent open and closed communication systems, respectively, providing complementary test cases for assessing the proposed framework’s adaptability and robustness in diverse communication environments.

For dataset construction, the strategy employed in this study involves training the model on clean spectrograms and subsequently testing it with more complex spectrograms. This approach ensures both effective feature learning and adaptability to real-world deployment scenarios. The training data is collected entirely in a microwave anechoic chamber under strictly controlled RF-shielded conditions. The frequency hopping signals obtained are free from external interference, with minimal background noise and clear spectral boundaries, ensuring that the model could effectively learn intrinsic characteristics such as UAV control signals, hopping trajectories, and modulation patterns during the initial phase, thereby enhancing feature extraction and identification performance. The dataset consists of two parts. The first part includes spectrograms of multi-class UAV RF signals obtained from an open-source database, with 600 spectrogram samples generated for each UAV type. The second part comprises spectrograms of the self-assembled UAV and the DJI Mini, collected using the measurement platform in a microwave anechoic chamber as illustrated in Figure 8, with 600 spectrogram samples extracted for each UAV in this private measured dataset. For both datasets, all spectrograms are standardized to a resolution of 800 × 800 pixels and divided into training and validation sets with a 2:1 ratio.

To evaluate the performance of the proposed system under different propagation path conditions, measurements are conducted at six representative locations around the Technical University of Madrid, as shown in Figure 9. The experimental setting represents a typical urban low-altitude propagation scenario, incorporating various interference sources and obstructions typically encountered during UAV flight, such as low-rise buildings, vehicles, and trees. In addition, multiple RF sources, including Wi-Fi, Bluetooth, and other wireless devices, contributed to the complex and non-ideal electromagnetic environment. To assess the system’s adaptability for mobile monitoring applications, the receiving equipment is mounted on a vehicle to emulate a mobile monitoring terminal, which could be deployed for applications such as urban security patrols and emergency communication management, as shown in Figure 10a. The UAV and its controller are placed on a fixed platform on the rooftop of a building approximately 20 m above ground, serving as the transmission source, as illustrated in Figure 10b. The horizontal distance between the receiver and transmitter varied from 50 m to 150 m, covering typical low-altitude, short-range communication link characteristics. Specifically, test positions 1, 2, 3, and 6 correspond to LoS conditions, where no obstacles existed between the receiver and the UAV, and signal propagation occurred primarily along the direct LoS. In contrast, test positions 4 and 5 correspond to NLoS conditions, where the propagation path is partially obstructed by buildings, trees, or other obstacles. In these cases, the receiver mainly captured reflected or diffracted signals, resulting in greater environmental complexity and unpredictability. At each position, 500 spectrograms are collected for both the assembled UAV and the DJI Mini. As shown in Figure 11, the spectrograms display clear FHSs of the assembled UAV. In addition, significant RF interference is present in the environment. This interference manifests in the spectrograms as background noise patterns, broadband pulse bursts, or spurious frequency hopping signatures, which overlapped with and in some cases interfered with the target signal in the time–frequency domain, thereby complicating the detection task. Particularly at some NLoS locations, characterized by long reflection paths, high attenuation, and dense interference sources, the target signal is easily obscured.

4.2. Performance Evaluation

To evaluate the performance of the proposed YOLO-based UAV detection algorithm, validation measurements are conducted at six positions and the results are compared with those obtained using traditional methods that detect signals from the entire spectrogram. The results are summarized in

Table 3. Accuracy comparison of UAV detection methods.

Test Poistions	Traditional Method Accuracy [39]	Proposed Method Accuracy	Difference
1 (LoS)	87.6%	96.3%	+8.7%
2 (LoS)	77.4%	89.3%	+11.9%
3 (LoS)	75.7%	86.5%	+10.8%
6 (LoS)	84.5%	92.2%	+7.7%
4 (NLoS)	61.0%	81.1%	+20.1%
5 (NLoS)	57.9%	72.8%	+14.9%

. Under LoS conditions, the proposed method achieved detection accuracies of 96.3%, 89.3%, 86.5%, and 92.2% at test positions 1, 2, 3, and 6, respectively, which represents improvements of 8.7%, 11.9%, 10.8%, and 7.7% over the traditional method [39]. Under NLoS conditions, the performance gap between the two methods becomes more pronounced. The traditional method achieved accuracies of 61.0% and 57.9% at test positions 4 and 5, whereas the proposed method reached 81.1% and 72.8%, corresponding to improvements of 20.1% and 14.9%, respectively. These results indicate that traditional methods are more susceptible to background interference in complex electromagnetic environments and obstructed conditions, which leads to degraded detection performance. In contrast, the proposed method, by extracting FHS regions, effectively isolates non-target interference and captures key structural features. This enables stable detection and high-precision detection of UAV hopping signals across various environmental conditions, demonstrating superior robustness and adaptability.

To further validate the effectiveness of the proposed UAV identification model, comparative analyses are conducted on the identification performance of the proposed system under multi-source data conditions. Initially, in an ideal environment, spectrograms of six representative commercial UAVs from the open-source database, together with spectrogram samples of the assembled UAV and DJI Mini collected in a microwave anechoic chamber, are selected to construct the training and testing datasets. A confusion matrix is employed to compare performance of the model across different UAV types. As a commonly used tool in multi-class classification tasks, the confusion matrix provides a visual representation of identification accuracy and misclassification rates for each UAV category. Each column of the matrix corresponds to predicted labels, while each row corresponds to true UAV labels. The identification results are shown in Figure 12. The proposed model achieved 100% accuracy across all test objects, thereby fully confirming the theoretical performance limits of the method in interference-free environments and demonstrating its precise capability for extracting and classifying frequency-hopping features.

To further validate the effectiveness of the proposed UAV identification module, two UAV models, the assembled UAV and DJI Mini, are selected as the target objects. The performance of the traditional identification method [39] is shown in Figure 13. Under LoS conditions, the identification accuracies for the assembled UAV and DJI Mini are 82.6% and 84.3%, respectively, demonstrating limited discriminative ability. However, under NLoS conditions, the accuracies dropped sharply to 61.4% and 69.3%, with misidentification rates reaching 38.6% and 30.7%, respectively. These results indicate that traditional methods, which lack structural constraints during feature extraction, are highly susceptible to interference from video transmission signals, Wi-Fi, and other non-target RF sources, leading to model confusion and performance degradation. The performance of the proposed identification method is shown in Figure 14. Under LoS conditions, the identification accuracies for the assembled UAV and DJI Mini improved significantly to 96.8% and 94.1%, respectively, compared to the traditional approach. In addition, the proposed framework achieves an average processing time of approximately 34 ms per detection–identification cycle, demonstrating millisecond-level, real-time performance suitable for practical UAV monitoring. More importantly, the proposed method maintained high accuracy and low misidentification rates under NLoS conditions, where traditional full image identification approaches perform poorly due to multipath interference and background noise. By focusing on frequency-hopping structures, the proposed system effectively isolates target features from interference, demonstrating superior robustness and adaptability. These results further demonstrate the effectiveness of the proposed two-stage YOLO–CNN framework. In this design, the YOLO network rapidly locates frequency-hopping signal regions within the spectrogram, significantly reducing the impact of environmental interference, while the CNN classifier performs fine-grained feature extraction for UAV type identification. Compared with single-stage deep learning models, this architecture achieves a better trade-off between real-time performance and identification accuracy, especially under non-stationary or overlapping signal conditions commonly encountered in complex electromagnetic environments.

5. Conclusions

This paper addresses the challenge of UAV detection and identification in complex electromagnetic environments by designing and implementing a spectrogram-based detection and identification system. The hardware platform integrates a high-sensitivity multi-beam directional antenna, a wideband spectrum analyzer, and a computing unit, providing strong spatial directivity and flexible deployment capabilities. On the algorithmic side, a joint identification framework combining YOLO and CNN is developed to enable accurate localization of FHS. Extensive measurements are conducted under both LoS and NLoS conditions around a university campus, focusing on two UAV targets: a custom-built drone and the DJI Mini. Experimental results show that the proposed system achieves over 96% identification accuracy under LoS conditions and improves accuracy by more than 15% over traditional methods in NLoS scenarios. Overall, the system demonstrates significant advantages in terms of identification accuracy, response speed, and interference resilience, highlighting its strong engineering applicability and potential for real-world deployment. Moreover, in our future work, we plan to integrate the RF sensing and identification modules onto UAV platforms to enable airborne passive monitoring and dynamic spectrum mapping, thereby enhancing environmental coverage and response speed in real-world deployments. In addition, we will further extend the framework to support simultaneous detection and identification of multiple UAVs in complex electromagnetic environments.