Localization of Multiple GNSS Interference Sources Based on Target Detection in C/N₀ Distribution Maps

Abstract

The localization of multiple interference sources in Global Navigation Satellite Systems (GNSS) can be achieved using carrier-to-noise ratio (C/N₀) information provided by GNSS receivers, such as those embedded in smartphones. However, in increasingly prevalent complex scenarios—such as the coexistence of multiple directional interferences, increased diversity and density of GNSS interference, and the presence of multiple low-power interference sources—conventional localization methods often fail to provide reliable results, thereby limiting their applicability in real-world environments. This paper presents a multi-interference sources localization method using object detection in GNSS C/N₀ distribution maps. The proposed method first exploits the similarity between C/N₀ data reported by GNSS receivers and image grayscale values to construct C/N₀ distribution maps, thereby transforming the problem of multi-source GNSS interference localization into an object detection and localization task based on image processing techniques. Subsequently, an Oriented Squeeze-and-Excitation-based Faster Region-based Convolutional Neural Network (OSF-RCNN) framework is proposed to process the C/N₀ distribution maps. Building upon the Faster R-CNN framework, the proposed method integrates an Oriented RPN (Region Proposal Network) to regress the orientation angles of directional antennas, effectively addressing their rotational characteristics. Additionally, the Squeeze-and-Excitation (SE) mechanism and the Feature Pyramid Network (FPN) are integrated at key stages of the network to improve sensitivity to small targets, thereby enhancing detection and localization performance for low-power interference sources. The simulation results verify the effectiveness of the proposed method in accurately localizing multiple interference sources under the increasingly prevalent complex scenarios described above.

1. Introduction

GNSS generally refer to various global satellite navigation systems and their augmentation systems, including the U.S. Global Positioning System (GPS), China’s BeiDou, Russia’s Global Navigation Satellite System (GLONASS), and Europe’s Galileo, among others [1,2]. GNSS is the preferred technology for most location- and time-dependent applications [3], supporting critical infrastructure closely tied to national operations and people’s livelihoods [4]. The spread-spectrum modulation techniques employed in satellite navigation provide GNSS signals with a certain level of anti-jamming capability [5]. However, the low received power of GNSS signals makes the system vulnerable to electromagnetic environments or interfering signals [6,7].

Currently, GNSS systems face increasingly severe interference threats, with interference incidents occurring more frequently. Therefore, rapid and effective interference source localization has become a crucial means of ensuring GNSS application security [8,9]. To enhance the effectiveness of interference while improving concealment, multiple interference sources are often employed cooperatively to target GNSS signals. Moreover, low-cost jammers are now readily available on the commercial market [10], making multi-source GNSS interference an even greater threat than single-source interference [11]. With the widespread adoption of null-steering antennas, an array antenna composed of M elements can suppress M-1 interference sources [12]. This has led to the evolution of interference methods from single-source to multi-source cooperative jamming [13]. Such cooperative interference typically employs distributed low-power directional jamming, leading to increasingly common yet challenging scenarios, including multiple directional interference sources, growing types and quantities of GNSS interference, and multiple weak interference signals. These developments pose significant challenges to existing GNSS interference localization methods. This paper focuses on addressing the multi-interference source localization problem in these challenging scenarios.

1.1. Related Work

Common GNSS interference source localization methods include Direction of Arrival (DOA)-based [14], Time Difference of Arrival (TDOA)-based [15,16] and Received Signal Strength (RSS)-based [17,18] approaches. Among these, DOA-based localization methods are limited by the number of antenna array elements, and the commonly used cross-localization technique suffers from positioning ambiguity when multiple interference sources coexist [19]. TDOA-based methods, however, face challenges in accurately extracting time differences between various interference sources and receivers due to the difficulty in separating multiple interfering signals. Additionally, commercial off-the-shelf (COTS) receivers typically cannot provide the raw data samples required for TDOA estimation [20].

Due to their low implementation complexity, RSS-based methods have been widely adopted for GNSS interference source localization [21]. The concept of Jamming 911 (J911) was first proposed by Logan Scott at the 2011 Institute of Navigation (ION) GNSS conference. It introduced a crowdsourcing-based approach for interference detection and localization, leveraging carrier-to-noise ratio (C/N₀) measurements reported by smartphones to identify and locate GNSS jammers [22]. Smartphones enjoy near-ubiquitous global penetration. For instance, China’s Ministry of Industry and Information Technology (MIIT) reported in its 2023 Communications Industry Statistical Bulletin (published in January 2024) that the country hosts 11.62 million mobile communication base stations and 1.727 billion mobile phone subscribers [23]. However, the developer of the J911 system did not specify how to efficiently process such massive-scale GNSS receiver data, nor did they provide an effective solution for multi-interference source localization [24]. Reference [25] proposed a centroid-based localization method for multiple GNSS interference sources using clustering and centroid convergence algorithms on C/N₀ data. While this method achieves multi-source localization, its performance degrades as the number and variety of interference sources increase. Moreover, it fails to address scenarios involving directional GNSS jammers or low-power interference sources.

1.2. Our Contribution

The paper begins by establishing a system and signal model tailored to the GNSS interference source localization scenario under study. Building upon this foundation, it proposes a multi-interference sources localization method based on object detection in GNSS C/N₀ distribution maps. The method innovatively constructs C/N₀ distribution maps by exploiting the similarity between GNSS receiver-reported data and image grayscale information. This transforms the GNSS multi-interference sources localization problem into an image processing-based object detection and localization task. To address key challenges in detecting and localizing target interference sources—including estimation of the main lobe direction for directional antennas, multi-target feature recognition and classification, and the detection/localization of low-power GNSS interference sources—the paper proposes an OSF-RCNN framework, an enhanced variant of Faster R-CNN, for processing C/N₀ distribution maps. The OSF-RCNN framework incorporates an Oriented RPN to regress the orientation angles of directional antennas, thereby enhancing both detection accuracy and localization precision for directional interference sources. The framework employs an FPN to achieve multi-scale feature fusion, which improves the method’s adaptability to various interference types and different transmission power levels. Additionally, SE blocks are strategically integrated at critical network nodes to intensify focus on low-power interference targets, consequently boosting detection precision for the small low-power targets.

The study utilized the established system and signal model to construct a dedicated dataset, followed by comprehensive simulation tests to evaluate the proposed method’s performance. The experimental validation specifically assessed the method’s localization capabilities under the following challenging conditions: (1) directional interference, (2) multiple interference types, and (3) weak interference scenarios. The results demonstrate that the proposed OSF-RCNN method achieves effective detection and accurate localization of multiple GNSS interference sources across all these challenging operational conditions.

The remainder of this paper is organized as follows: Section 2 introduces the system and signal model of the scenarios discussed. In addition, this section describes the specific steps of the proposed multi-interference sources localization method based on target detection in C/N₀ distribution maps. It also presents the construction method for C/N₀ distribution maps, explains the dataset generation process, and analyzes the remaining challenges in detecting and localizing target interference sources in complex scenarios. In response to these challenges, the OSF-RCNN framework is designed and proposed. Section 3 conducts simulation tests to evaluate the method’s performance and analyzes the test results. Section 4 presents the conclusions of this paper.

2. Materials and Methods

The paper proposes a localization method for multiple GNSS interference sources based on target detection in C/N₀ distribution maps. The method begins by constructing C/N₀ distribution maps, utilizing the similarity between GNSS receiver C/N₀ measurements and image grayscale features. Then, an OSF-RCNN framework is introduced to process the C/N₀ distribution map, enabling effective GNSS interference localization in challenging scenarios such as the presence of multiple directional interference sources, diverse and numerous interference types, and multiple weak interference signals. A detailed description of the proposed method is provided in this chapter.

2.1. System Model

The GNSS multi-interference sources localization scenario in this study is illustrated in Figure 1, which includes both directional and omnidirectional interference sources. The proposed method enables effective GNSS multi-interference sources localization by processing the GNSS data collected from numerous smartphones within the monitored area.

In Figure 1, the number of smartphones is M; the C/N₀ returned by the k-th GNSS receiver and its own location information form a three-dimensional vector dataset ${\mathit{X}}_k=\left[x_k,y_k,{\left(C/N_0\right)}_{eq,k,dB}\right]$, where $x_k$ and $y_k$ are the two-dimensional coordinates of the receiver, and ${\left(C/N_0\right)}_{eq,k,dB}$ is the C/N₀ measurement of the smartphone. This paper utilized the data ${\mathit{X}}_k,1\le k\le M$ returned by smartphones in the scenario to confirm the location of interference sources, denoted as ${\mathit{C}}_i=\left[x_i,y_i\right],1\le i\le N$, with the number of interference sources being N. The aim of this paper is to determine the locations ${\mathit{C}}_i=\left[x_i,y_i\right](1\le i\le N)$ of multiple interference sources when they coexist, using the data ${\mathit{X}}_k(1\le k\le M)$ returned by receivers in the scenario. The problem can be described as follows:

$$\left({\mathit{C}}_i,i=1,\dots ,N\right)=T\left({\mathit{X}}_k,k=1,\dots ,M\right)$$

(1)

where T represents the multi-interference sources localization method in this paper. For the scenario described in this paper, the system model was further developed to account for the presence of directional interference sources. Reference [25] provides a method for calculating the C/N₀ of navigation receivers under the simultaneous influence of multiple GNSS interference sources. The formula for calculating the C/N₀ of the k-th receiver in this scenario is as follows.

$${\left(C/N_0\right)}_{eq,k}={\displaystyle \frac{\left(C/N_0\right){{\displaystyle \int }}_{-\infty }^{\infty }{\left|H_{R_k}\left(f\right)\right|}^2{S}_f\left(f\right)df}{{{\displaystyle \int }}_{-\infty }^{\infty }{\left|H_{R_k}\left(f\right)\right|}^2{S}_f\left(f\right)df+{{\displaystyle \sum }}_{i=1}^N\frac{J_{i,k}}{N_0}{{\displaystyle \int }}_{-\infty }^{\infty }{\left|H_{R_k}\left(f\right)\right|}^2{S}_{J_i}\left(f\right)S_f\left(f\right)df}}$$

(2)

where ${\left(C/N_0\right)}_{eq,k}$ represents the equivalent C/N₀ of GNSS the k-th receiver under the influence of multiple interferences; C is the satellite signal power; $C/N_0$ is the C/N₀ without interference; $J_{i,k}$ is the interference signal power of the i-th interference source at the k-th receiver; $H_{R_k}\left(f\right)$ is the transfer function of the k-th receiver’s filter; $S_f\left(f\right)$ is the power spectral density of the satellite signal, normalized to the unit area over infinite bandwidth; and $S_{J_i}\left(f\right)$ is the signal power spectral density of interference source i, normalized to unit area over infinite bandwidth. We define the spread-spectrum processing anti-interference quality factor $Q_{i,k}$ related to the interference signal and satellite navigation signal type as:

$$Q_{i,k}={\displaystyle \frac{{{\displaystyle \int }}_{-\infty }^{\infty }{\left|H_{R_k}\left(f\right)\right|}^2{S}_f\left(f\right)df}{R_C{{\displaystyle \int }}_{-\infty }^{\infty }{\left|H_{R_k}\left(f\right)\right|}^2{S}_{J_i}\left(f\right)S_f\left(f\right)df}}$$

(3)

After substituting Equation (3) into Equation (2), Equation (2) can be simplified as follows:

$${\left(C/N_0\right)}_{eq,k}={\displaystyle \frac{1}{\frac{1}{C/N_0}+{{\displaystyle \sum }}_{i=1}^N\frac{J_{i,k}/C}{Q_{i,k}{R}_c}}}$$

(4)

where the spread spectrum processing anti-jamming quality factor $Q_{i,k}$ is related to the types of interference signals and satellite navigation signals, and $R_c$ represents the pseudocode rate of the satellite guidance signal. Typically, GNSS receivers estimate the C/N₀ in units of $\mathrm{dB}\cdot \mathrm{Hz}$, and the calculation formula is as follows:

$${\left(C/N_0\right)}_{eq,k,dB}\triangleq 10lg{\left(C/N_0\right)}_{eq,k}$$

(5)

According to the classic narrowband radio propagation path loss model provided in reference [26], the interference signal power $J_{i,k}$ from interference source i at the k-th receiver can be related to the transmission power $J_{i,T}$ of interference source i, as shown in the following equation.

$${\left(J_{i,k}\right)}_{dB}={\left(J_{i,T}\right)}_{dB}-L_0-10\alpha lgd-v$$

(6)

where ${\left(J_{i,T}\right)}_{dB}=10lg\left(J_{i,T}\right)$, ${\left(J_{i,k}\right)}_{dB}=10lg\left(J_{i,k}\right)$, $L_0$ is the propagation loss at 1 m from the transmitter, $d$ is the distance between transmitter and receiver, $\alpha$ is the path loss index, and $v$ is a Gaussian random variable representing noise. Considering both omnidirectional and directional interference sources, Formula (6) can be expressed as:

$${\left(J_{i,k}\right)}_{dB}={\left(P_{i,T}\right)}_{dB}+{\left(G_{i,k}\right)}_{dB}-L_0-10\alpha lgd-v$$

(7)

where ${\left(P_{i,T}\right)}_{dB}$ represents the power of the interference source amplifier, and ${\left(G_{i,k}\right)}_{dB}$ denotes the antenna gain of the interference source, which differs between omnidirectional and directional interference. Using the aforementioned model, a simulation scenario was constructed where the positions of the GNSS jammers and GNSS receivers were randomly assigned within a specified area. Using Formula (6), the signal power of each interference source at the location of each GNSS receiver can be calculated in the presence of directional interference sources. Additionally, Formula (4) can be used to calculate the C/N₀ of each GNSS receiver under the influence of multiple interference sources when directional interferences are present.

2.2. C/N₀ Distribution Map Construction Method

Grayscale images are a common type of image in computer vision applications. In a grayscale image, the brightness of each pixel $\left(i,j\right)$ is represented by a grayscale value $Gray\left(i,j\right)$, which is typically expressed as an 8-bit integer ranging from 0 to 255. In this case, 0 represents black, 255 represents white, and intermediate values correspond to varying shades of gray [27]. Since grayscale images contain only brightness information, they simplify image representation complexity while retaining the essential features of the image. Grayscale images are often used in operations such as edge detection, image segmentation, and feature extraction, enhancing the efficiency and reliability of image processing and analysis [28].

The GNSS receiver information reported by smartphones in Figure 1 includes receiver location data and C/N₀ information, denoted as ${\mathit{X}}_k=\left[x_k,y_k,{\left(C/N_0\right)}_{eq,k,dB}\right]$, which exhibits a high degree of similarity to the grayscale information of images. In this study, we utilize the information ${\mathit{X}}_k$ relayed by the GNSS receivers embedded in smartphones to construct the C/N₀ distribution map. In this context, the terms “smartphone” and “GNSS receiver” refer to the same device. Therefore, a method is proposed to construct a C/N₀ distribution map by projecting the C/N₀ data collected from numerous GNSS receivers in the monitoring area onto a grayscale image. After generating the C/N₀ distribution map, target detection methods from image processing can be used to detect GNSS interference sources as targets, thereby achieving the effective localization of multiple GNSS interference sources in the scenario of this paper.

Algorithm 1 presents the pseudocode for constructing the C/N₀ distribution map, where the navigation data reported by receivers in the monitored area is used to define the map dimensions and pixel grayscale values. Steps 1 to 3 calculate the receiver density and interference source localization accuracy within the monitored area to obtain the width and height of the C/N₀ distribution map. In this algorithm, the C/N₀ value under interference-free conditions is set to ${(C/N_0)}_{\max }$, and the receiver’s C/N₀ decreases as the level of interference increases [29]. To ensure that the generated C/N₀ distribution map effectively reflects the degree of interference experienced by the receivers, a C/N₀ threshold is set to ${(C/N_0)}_{CT}$. Grayscale values for C/N₀ values below the threshold ${(C/N_0)}_{CT}$ are set to 0 and values under interference-free conditions are set to 255, so that the receiver’s C/N₀ value $\left[{(C/N_0)}_{CT},{(C/N_0)}_{\max }\right]$ can be linearly mapped to grayscale values from 0 to 255 using the following formula:

$$Gra{y}_l=⌊{\displaystyle \frac{255\cdot \left[{(C/N_0)}_l-{(C/N_0)}_{CT}\right]}{\left[{(C/N_0)}_{\max }-{(C/N_0)}_{CT}\right]}}⌋$$

(8)

where ${(C/N_0)}_l$ represents the C/N₀ measurement at a specific receiving point, and $Gra{y}_l$ is the corresponding grayscale value at that point.

Steps 5 to 13 use the C/N₀ measurements reported by the receivers in the monitored area to fill each pixel of the C/N₀ distribution map, in which Step 6 mainly involves searching for data points $\left\{X_{l,D1},X_{l,D2},\dots ,X_{l,DM}\right\}$ whose distance from the pixel is less than $Dp$, and DM is the number of points that satisfy the condition. Step 7 calculates the weighted average value ${(C/N_0)}_{lM}$ of the C/N₀ data for the DM data points, where the weight is the inverse of the distance to that pixel. If ${(C/N_0)}_{lM}>{(C/N_0)}_{CT}$, then the gray value corresponding to the value ${(C/N_0)}_{lM}$ needs to be calculated using Formula (8) and written to the corresponding pixel. If ${(C/N_0)}_{lM}\le {(C/N_0)}_{CT}$, then the gray value 255 is written to that pixel. Steps 14 to 16 apply dilation and median filtering to the generated grayscale image to suppress noise, thereby generating the C/N₀ distribution map.

Algorithm 1: Generation method of the CNR distribution map

INPUT: The number of receivers M

INPUT: Monitoring area width length L and width W

INPUT: The receiver returned data set $\left\{X_1,X_2,\dots ,X_M\right\}$

INPUT: The C/N0 threshold ${(C/N_0)}_{CT}$

INPUT: The C/N0 value under interference-free ${(C/N_0)}_{\max }$

INPUT: Set the distance threshold $Dp$ for selecting C/N0 data when assigning grayscale values to pixels

OUTPUT: The C/N0 distribution map ICN

1: Calculate receiver density as $Rd=M/L\times W$

2: Set the localization accuracy as $Lp=1/⌈Rd/40⌉\times 2$

3: Calculate the width of the C/N0 distribution map as $W_{ICN}=L/Lp$, the height of the C/N0 distribution map as $H_{ICN}=W/Lp$, and the number of pixels as $N=W_{ICN}\times H_{ICN}$

4: Initialize the gray values of all pixels in the C/N0 distribution map ICN to 0 corresponding to the C/N0 under interference-free conditions

5: While $l\le N$ do

6: Obtain the receiver returned data set $\left\{X_{l,D1},X_{l,D2},\dots ,X_{l,DM}\right\}$ that are within a distance of $Dp$ from the pixel

7: Calculate the weighted average ${(C/N_0)}_{lM}$ of the C/N0 data for the DM data points, where the weights are the inverse of the distance to the pixel

8: If ${(C/N_0)}_{lM}>{(C/N_0)}_{CT}$ then

9: Calculate the grayscale value $Gra{y}_l$ corresponding to the value ${(C/N_0)}_{lM}$ using Formula (8) and write it to the corresponding pixel

10: Else

11: Write the grayscale value 255 to the corresponding pixel

12: End if

13: End while

14: Perform a dilation operation on ICN

15: Perform a median filtering operation on ICN

16: Return the C/N0 distribution map ICN

2.3. Dataset Generation and Problem Description

2.3.1. Dataset Generation

According to the signal model outlined in Section 2 and the construction method of the C/N₀ distribution map described in Section 3.1, a dataset was generated. This dataset comprised 20,000 C/N₀ distribution maps, each corresponding to a single interference localization experiment. For each interference localization experiment, the following parameters were configured:

(1) The monitoring area was set to 60 × 60 km.

(2) The number of receivers was randomly set between 400 to 600 per square kilometer, with their positions following a random distribution.

(3) The maximum number of interference sources was set to 16. In each trial, the total number of interference sources (denoted as $N\_all$) was randomly selected as an integer between 1 and 16. Subsequently, an integer $N\_Od$ between 1 and $N\_all$ was randomly chosen to represent the number of omnidirectional interference sources. The number of directional interference sources for that trial was then calculated as $N\_all-N\_Od$. For instance, in one trial, a random integer between 1 and 16 was selected, resulting in 16 jamming sources. Among these, 7 were randomly designated as omnidirectional jammers, leaving 9 as directional jammers.

(4) The antenna type of each jamming source was randomly assigned using typical models provided by MATLAB R2024a’s Antenna Toolbox [30]. Figure 2 illustrates the radiation patterns of different antenna models at the navigation frequency. Among them, the typical omnidirectional antenna is the dipole antenna, while typical directional antennas include Yagi-Uda antennas, horn antennas, Vivaldi antennas, helical antennas, and patch microstrip antennas. Simultaneously, the power of each jammer and the jammer direction were randomly set sequentially. The jammer power was an arbitrary integer between −10 dBm and 10 dBm. The jammer direction was randomly set to an arbitrary angle within the range of 1° to 360°, with an adjustment accuracy of 1°. The jamming transmission was narrowband jamming, corresponding to a Q value of 1.

(5) Based on the system model in Section 2, the navigation signal C/N₀ for each receiver in multiple interference source scenarios can be calculated for each experiment. On this basis, the method in Section 3.1 can be used to generate a sample of the C/N₀ distribution map for this experiment.

The dataset generated through the aforementioned steps was utilized for training and testing the proposed target detection method. Representative examples of C/N₀ distribution maps from the dataset are illustrated in Figure 3. As shown, interference effects cause signal C/N₀ degradation, forming interference source shadow targets. Consequently, the interference source localization problem can be transformed into a target detection task within the C/N₀ distribution maps. In these maps, omnidirectional and directional interference sources exhibit distinct image characteristics under different parameter configurations. For the scenario addressed in this study, the objectives were to identify target types and precisely localize interference sources within the C/N₀ distribution maps. Deep learning-based target detection methods offer significant advantages in terms of both accuracy and computational efficiency. Therefore, this work employed a deep learning-based detection framework to achieve robust interference source detection and localization.

2.3.2. Problem Description

This method in Section 3.1 constructs a C/N₀ distribution map by utilizing the similarity between GNSS receiver return information and image grayscale information. Once the C/N₀ distribution map is obtained, object detection methods from image processing can be employed to detect and locate GNSS interference sources as the targets of interest. Traditional object detection algorithms typically achieve target detection by extracting image features and combining them with classification methods such as Support Vector Machine (SVM). However, these methods are computationally intensive and suffer from poor generalization performance. With the rapid development of deep learning, deep learning-based object detection methods have been widely adopted, offering advantages in accuracy, efficiency, and automation compared to traditional methods. In scenarios such as the coexistence of multiple directional interferences, increased diversity and density of GNSS interference, and the presence of multiple low-power interference sources, further solutions are needed to address the following issues in target localization using the C/N₀ distribution map:

(1) The estimation of the main lobe direction of the directional interference source antenna.

Figure 4 presents illustrative examples of targets in the C/N₀ distribution map under different pointing angles of directional interference using a Horn antenna. As shown, when the interference source uses a directional antenna, its radiation pattern is non-uniform and non-omnidirectional, with the main lobe exhibiting significantly higher gain than other directions. As a result, a shadow region with reduced C/N₀ appears in the direction of the antenna’s main lobe. To accurately localize directional interference sources, it is necessary to identify the main lobe direction of the directional antenna. Therefore, the target detection algorithm used must be capable of recognizing rotated targets and outputting their rotation angles.

(2) The problem of multi-type antenna target feature recognition and classification.

Figure 5 presents a schematic diagram of different types of interference source antennas at an equivalent radiated power of 0 dBm in the C/N₀ distribution map. As illustrated, interference sources with different antenna types exhibit distinct target characteristics. Therefore, to enable the effective localization of multiple GNSS interference sources under various scenarios, it is essential to first learn the features of each target type and then classify them accordingly.

(3) The problem of detecting and localizing weak GNSS interference source targets in the distribution map.

Figure 6 illustrates targets of interference sources with different antenna types at an equivalent radiated power of −15 dBm in the C/N₀ distribution map. As shown, weak GNSS interference sources typically affect only a limited number of navigation receivers, resulting in small-sized targets within the distribution map. To improve the accuracy of detection and localization for such weak interference sources, it is essential to enhance the deep learning model’s ability to learn and distinguish the features of small targets associated with various types of weak GNSS interference.

2.4. OSF-RCNN

Deep learning-based object detection methods can be divided into two categories: two-stage approaches based on region proposals and one-stage approaches based on regression. The two-stage method first generates candidate bounding boxes in the processed image, and then classifies these proposals and performs precise positional regression to obtain detection results. Representative algorithms of this approach include the Region-Based Convolutional Neural Network (R-CNN) series [31]. In contrast, one-stage methods treat all possible locations in the image as potential candidate bounding boxes, directly performing object classification and bounding box regression to obtain detection results, with the You Only Look Once (YOLO) series being representative examples [32]. YOLO is a single-stage object detection algorithm that formulates detection as a regression problem. It processes the entire image in a single forward pass and directly predicts bounding boxes along with class probabilities, enabling real-time performance. Since the one-stage method exhibits lower localization accuracy and poorer performance in multi-object detection [33], and this study requires the precise localization of multiple GNSS interference sources, we propose improvements to the Faster R-CNN algorithm to adapt it to the application scenario of this research.

To address the problems outlined in Section 2.3.2, this paper proposes an OSF-RCNN framework based on the Faster R-CNN algorithm for processing C/N₀ distribution maps to achieve the effective localization of multiple GNSS interference sources. Built upon the Faster R-CNN framework, the proposed OSF-RCNN can effectively learn and recognize the features of different types of interference sources. To account for the rotational characteristics of directional antennas, an Oriented RPN was introduced to regress the orientation angles of directional antennas [34]. Additionally, the SE mechanism [35] was integrated into key positions of the network to enhance attention to weak interference signals. Furthermore, FPN was employed to fuse multi-type and multi-scale feature information, thereby improving the detection accuracy of small targets and making the method more suitable for weak GNSS interference detection and localization. The experimental results demonstrate that the proposed method achieves accurate and reliable localization of multiple GNSS interference sources, effectively addressing challenging scenarios such as the presence of multiple directional interferences, increasing diversity and quantity of GNSS interference, and the detection of multiple weak interference signals.

2.4.1. Algorithmic Framework

The overall framework of the proposed OSF-RCNN is illustrated in Figure 7. It first employs a backbone network to extract multi-scale feature maps from the input image. Specifically, the backbone network integrates an SE-enhanced ResNet50 (SE-ResNet50) to extract image features, followed by an FPN that processes and fuses the feature maps output from different stages of the SE-ResNet50. As a result, multi-scale feature maps with rich semantic information and high resolution are generated [36]. Subsequently, an Oriented RPN processes the multi-scale feature maps to generate candidate bounding boxes that encode both direction and rotation information. After obtaining the proposals, both the candidate box information and feature maps are fed into a Rotated RoIAlign layer to achieve precise alignment between the Regions of Interest (RoIs) and the feature maps. Finally, the feature maps corresponding to the candidate boxes are passed through fully connected (FC) layers to perform the classification of the target boxes and regression of their positions and boundaries, thereby achieving accurate target localization.

2.4.2. Backbone Structure

The backbone network adopts a design that combines SE-ResNet50 with FPN [37], and the overall architecture of the network is shown in Figure 8. ResNet-50 is a deep convolutional neural network architecture renowned for its unique residual learning and efficient training capabilities. It has demonstrated outstanding performance in image classification tasks on benchmark datasets like ImageNet, achieving high accuracy. Due to its residual connection design, ResNet-50 is easier to train and optimize compared to other networks of similar depth [38]. The SE-ResNet50 in the backbone network consists of five stages: Conv1, Conv2_x, Conv3_x, Conv4_x, and Conv5_x. The configuration of each stage is consistent with that described in reference [35]. The outputs of each node in the processing pipeline of the backbone network, which combines SE-ResNet50 and FPN, are shown in Figure 8. This network integrates the feature learning capability of ResNet50 and the multi-scale feature representation ability of FPN, significantly improving target detection performance. This design approach not only enhances the network’s adaptability to different interference intensities but also strengthens its ability to extract weak interference features.

2.4.3. SE Mechanism

In Section 2.4.2, ResNet-50 consists of two fundamental block structures, the Identity Block and the Conv Block, collectively referred to as the Residual Block. SE-ResNet50 introduces two FC layers into each block of ResNet-50 to recalibrate feature responses, enhancing attention to critical features while suppressing less relevant ones, thereby improving the network’s performance more efficiently. This structure is known as the SE mechanism, and its specific architecture is illustrated in Figure 9.

The goal of the SE mechanism is to introduce an adaptive mechanism that enables the model to dynamically learn and adjust the importance of each feature channel, thereby enhancing its focus on critical features. The Squeeze operation achieves this by using Global Average Pooling to capture global information for each feature channel. The Excitation operation employs a multilayer perceptron (MLP) to learn excitation weights for the feature channels. This MLP typically consists of two FC layers: the first layer reduces the dimensionality, while the second FC layer restores the reduced dimension back to the original size. The learned excitation weights can be regarded as an attention mechanism that assigns importance to the feature responses of each channel in the feature map. Figure 9 illustrates the processing flow of the SE mechanism on the feature map and the corresponding changes in feature data dimensions. The output of the final Sigmoid function ranges between 0 and 1, representing the importance level of the corresponding feature channel.

2.4.4. Oriented RPN

The Oriented RPN enhances the original RPN by introducing a center-offset representation for rotated bounding boxes. It extends the original four-variable representation to six variables, enabling the inclusion of rotational attributes in the candidate boxes, as illustrated in Figure 10.

The Oriented RPN improves the regression from anchor boxes to proposal boxes from $\left(x,y,w,h\right)$ to $\left(x,y,w,h,\Delta \alpha ,\Delta \beta \right)$ compared to RPN. The illustration of each variable is shown in Figure 10, where $\left(x,y\right)$ represents the center coordinates of the proposal box, w is the width of the horizontal proposal box, h is the height of the horizontal proposal box, and $\Delta \alpha$ and $\Delta \beta$ represent the horizontal and vertical distances from the center of the horizontal proposal to the two vertices of the rotated proposal, respectively. By converting angle regression into distance regression, this approach avoids the non-differentiability of polygonal Intersection over Union (IoU) calculations and the discontinuity issues associated with angular representations.

Based on the above representation, the offset vector between the proposal boxes generated by the Oriented RPN and the anchor boxes is denoted as $\left({\delta }_x,{\delta }_y,{\delta }_w,{\delta }_h,{\delta }_{\alpha },{\delta }_{\beta }\right)$. From this, the representation vector $\left(x,y,w,h,\Delta \alpha ,\Delta \beta \right)$ of the rotated proposal can be calculated using the following formulas:

$$\left\{\begin{array}{c}\begin{array}{c}x=a_w\cdot {\delta }_x+a_x\\ y=a_h\cdot {\delta }_y+a_y\end{array}\\ \begin{array}{c}w=a_w\cdot e^{{\delta }_w}\\ h=a_h\cdot e^{{\delta }_h}\end{array}\\ \begin{array}{c}\Delta \alpha ={\delta }_{\alpha }\cdot w\\ \Delta \beta ={\delta }_{\beta }\cdot h\end{array}\end{array}\right.$$

(9)

It is important to note that the resulting shape may appear as a parallelogram. To rectify this, the shorter diagonal is extended to equal the length of the longer diagonal. The Oriented RoI generated by the Oriented RPN is processed using the Rotated RoIAlign method. Since the output of Oriented RPN consists of proposals with rotation angles, these proposals need to be rotated into horizontal ones before being aligned. This is the specific operation flow of the Rotated RoIAlign method.

3. Results and Discussion

This section presents a simulation-based analysis of the proposed method. Experiments 1, 2, and 3 evaluate its performance in directional interference scenarios, multi-type interference scenarios, and weak interference scenarios for the detection and localization of multiple GNSS interference sources, respectively. The results demonstrate that the proposed method can achieve the effective localization of GNSS interference sources under these challenging conditions.

3.1. Experiment 1: Simulation of Detection and Localization Performance for Directional Interference Sources

This section primarily evaluates the performance of the proposed method under directional interference scenarios with varying numbers of interference sources. The comparative methods include the approach from Reference [25] and the Faster R-CNN algorithm. Specifically, both Faster R-CNN and the proposed method were evaluated using two representative network architectures: ResNet-50 and VGG-16. ResNet-50 and VGG-16 are currently mainstream backbone networks for image recognition. Both achieve a good balance among accuracy, efficiency, and generalization ability. VGG-16 constructs a deep architecture by continuously stacking small convolutional kernels, featuring a simple and regular structure. ResNet-50 introduces a residual learning mechanism and employs skip connections to effectively alleviate the degradation problem in deep networks [39]. The experimental data were acquired following the methodology described in Section 3.1, with the scenario configuration detailed below:

(1) The GNSS interference monitoring area was set to 60 × 60 km.

(2) Receiver density varied randomly between 400 and 600 units per square kilometer, with spatial positions assigned randomly.

(3) Two types of interference sources were implemented: standard omnidirectional dipole antennas and directional horn antennas. The transmission power of each source was set to 0 dBm. The direction of each interference source was randomly assigned within the range from 1° to 360°, with a resolution of 1°. All sources emitted narrowband interference signals with a quality factor of Q = 1.

(4) Localization performance was evaluated across varying interference quantities (2–16 sources). For each source count, 100 Monte Carlo trials were conducted with randomized combinations of omnidirectional and directional sources.

(5) The detection accuracy rate $ac{c}_{ce}$ was calculated as $ac{c}_{ce}=T_{ce}/100$, where $T_{ce}$ represents successful detections across 100 trials. Meanwhile, the localization error was also statistically measured when the interference was correctly detected. The localization error was quantified as the mean root mean square error (RMSE) computed across multiple interference localization trials, measured in kilometers. The statistic $RMS{E}_M$ was calculated using formula $RMS{E}_M={{\displaystyle \sum }}_{k=1}^{T_{ce}}{{\displaystyle \sum }}_{i=1}^{N_{jammmer}}RMS{E}_i$, where $N_{jammer}$ represents the number of interferences.

In each Monte Carlo trial, the locations of the interference sources, as well as the number and locations of the receivers, were randomly generated. An example of such a trial is illustrated in Figure 11. Figure 11a shows the spatial distribution of the interference sources and receivers, while Figure 11b presents the corresponding C/N₀ distribution map generated based on the configuration in Figure 11a.

The experimental results are shown in Figure 12, where Figure 12a presents the statistical results for ${acc}_{ce}$, while Figure 12b presents the statistical results for $RMS{E}_M$. In the figures, CCC represents the method from Reference [25], while our proposed method is denoted as OSF-RCNN. The curves labeled Faster-RCNN-VGG16 and Faster-RCNN-ResNet50 correspond to the Faster R-CNN algorithm implemented with VGG-16 and ResNet-50 backbones, respectively. Similarly, OSF-RCNN-VGG16 and OSF-RCNN-ResNet50 represent the simulation results of our method using VGG-16 and ResNet-50 architectures, respectively.

The simulation results demonstrate that the CCC method exhibits significant degradation in interference detection accuracy as the number of directional interference sources increases, while ResNet-50 outperforms VGG-16 in both interference identification and quantity estimation. When using the ResNet-50 network, both the OSF-RCNN and Faster R-CNN methods can perform interference detection and estimate the number of interference sources, with OSF-RCNN achieving better performance. When different networks are used, the detection and recognition performance of Faster-RCNN-ResNet50 is significantly better than that of OSF-RCNN-VGG16, indicating that the choice of network has a considerable impact on the effectiveness of interference detection and recognition. OSF-RCNN-ResNet50 achieved the best interference detection and localization performance in this experimental scenario. In this experiment, when the number of interference sources reaches 18, the proposed method achieves an interference detection rate $ac{c}_{ce}$ of 77% and a positioning accuracy $RMS{E}_M$ of 0.5 km.

Figure 13 presents the experimental comparison between the OSF-RCNN-ResNet50 and OSF-RCNN-VGG16 methods. The OSF-RCNN-VGG16 method misidentifies the directional interference highlighted by the red circle as omnidirectional interference. Compared to the VGG-16 network, the ResNet-50 network has advantages such as a deeper architecture and better gradient propagation, making it more suitable for detecting and localizing directional interference in scenarios involving multiple interference sources.

Figure 14 presents the experimental comparison between the OSF-RCNN-ResNet50 and Faster-RCNN-ResNet50 methods. Figure 14c compares the localized results of the horn directional interference (highlighted in the top-left corners of Figure 14a,b with the ground-truth interference source position). It can be observed that the Faster-RCNN-ResNet50 method fails to accurately estimate the main lobe direction of the directional antenna, resulting in the deviation in the interference source position estimation.

As shown in Figure 15, with an increasing number of interference sources, phenomena such as target proximity to the edge of the C/N₀ distribution map and overlap between closely spaced targets lead to aliasing effects. In Figure 5a, where the number of interference sources is six, a localization error of up to 350 m occurs due to a target being near the edge of the C/N₀ map. In another case shown in Figure 15b, with 13 interference sources, localization errors of 538 m and 570 m are caused by overlapping targets. As the number of targets continues to increase, these issues become more pronounced, resulting in further degradation in the localization accuracy.

In conclusion, the CCC method is not applicable for interference source localization in scenarios with directional interference. The OSF-RCNN-ResNet50 method, with ResNet-50 as its backbone network, demonstrates superior performance in interference detection and identification, as well as interference source localization.

3.2. Experiment 2: Simulation of Detection and Localization Performance for Multi-Type Interference Sources

This section primarily evaluates the effectiveness of the proposed method in detecting and localizing multiple interference sources under multi-type interference scenarios using the optimized ResNet-50 network. The experimental data were acquired following the methodology described in Section 3.1, with the scenario configuration detailed below:

(1) The GNSS interference monitoring area was set to 60 × 60 km.

(2) Receiver density varied randomly between 400 and 600 units per square kilometer, with spatial positions assigned randomly.

(3) The number of interference types as set between two and six, with the typical omnidirectional antenna being a dipole antenna. Typical directional antennas include Yagi-Uda, horn, Vivaldi, helix, and patch microstrip antennas. The transmission power of each source was set to 0 dBm. The direction of each interference source was randomly assigned within the range from 1° to 360°, with a resolution of 1°. All sources emitted narrowband interference signals with a quality factor of Q = 1.

(4) The localization performance of the proposed method was tested under each interference type configuration with varying numbers of interference sources. The number of interference sources was set sequentially from 2 to 16, with 100 Monte Carlo trials conducted for each case. In each trial, the number of omnidirectional and directional interference sources was assigned randomly.

(5) The detection and recognition accuracy of interference sources, $ac{c}_{ce}$, was calculated, along with the localization error, $RMS{E}_M$, when the interference was correctly detected and recognized. The statistical methods are consistent with those described in Section 3.1.

The experimental results are shown in Figure 16, where Figure 16a presents the statistical results for ${acc}_{ce}$, while Figure 16b presents the statistical results for $RMS{E}_M$. The curves labeled NoT_2 to NoT_6 correspond to the cases where the number of interference types is set from two to six, respectively. The results demonstrate that the proposed method effectively adapts to complex scenarios with multiple interference types. However, as the number of interference types increases, both detection and localization performance exhibit slight degradation. In this experiment, under the condition of interference type 6 and 18 interference sources, the proposed method achieves an interference detection rate ${acc}_{ce}$ of 70% and a positioning accuracy $RMS{E}_M$ of 1.05 km.

Figure 17 illustrates examples of detection and localization performance for multi-type interference. Figure 17a–c show the interference detection and localization results in complex scenarios with interference types set to four, five, and six, respectively. It can be seen that the method proposed in this paper effectively achieves accurate detection and high-precision localization for multiple types of interference. In scenarios where the interference type is set to six and the number of interference sources reaches up to 18, this method can still achieve a detection recognition rate of over 70% and a localization accuracy within 1 km.

3.3. Experiment 3: Simulation of Detection and Localization Performance Under Varying Interference Transmission Powers

This section evaluates the performance of the proposed method in detecting and localizing multiple interference sources using the optimized ResNet-50 network under varying transmission power levels. The experimental data were acquired following the methodology described in Section 3.1, with the scenario configuration detailed below:

(1) The GNSS interference monitoring area was set to 60 × 60 km.

(2) Receiver density varied randomly between 400 and 600 units per square kilometer, with spatial positions assigned randomly.

(3) The number of interference types was set to three and six, with the typical omnidirectional antenna being a dipole antenna. Typical directional antennas include Yagi-Uda, horn, Vivaldi, helix, and patch microstrip antennas. The direction of each interference source was randomly assigned within the range from 1° to 360°, with a resolution of 1°. All sources emitted narrowband interference signals with a quality factor of Q = 1.

(4) For each interference type setting, the interference source power was set to 0 dBm, −5 dBm, and −15 dBm, respectively.

(5) For each combination of interference type and power, the localization performance of the method was tested with varying numbers of interference sources. The number of interference sources was set sequentially from 2 to 16, with 100 Monte Carlo trials conducted for each case. In each trial, the numbers of omnidirectional and directional interference sources were assigned randomly.

(6) The detection and recognition accuracy of interference sources, $ac{c}_{ce}$, was calculated, along with the localization error, $RMS{E}_M$, when the interference was correctly detected and recognized. The statistical methods are consistent with those described in Section 3.1.

The experimental results are shown in Figure 18, where Figure 18a presents the statistical results for ${acc}_{ce}$, while Figure 18b presents the statistical results for $RMS{E}_M$. The result curves under different interference source power levels and interference source type settings were plotted. For instance, 0 dBm_NoT_3 represents the curve where the number of interference source types is three and the interference source power is 0 dBm. The results demonstrate that the proposed method effectively adapts to complex scenarios with weak interference sources. However, as the number of interference types increases and the interference power decreases, the detection and localization performance of the proposed method exhibits a slight decline. As the interference transmission power decreases, changes in transmission power have a greater impact on the localization performance of the proposed method than the increase in the number of interference types. In this experiment, under the condition of an interference power of −15 dBm, interference type six, and 18 interference sources, the proposed method achieves an interference detection rate $ac{c}_{ce}$ of 60% and a positioning accuracy $RMS{E}_M$ of 1.8 km.

Figure 19 illustrates the interference detection and localization performance under different interference transmission powers when the interference type is set to six. Figure 19a–c show the detection and localization performance of the proposed method when the interference transmission power is 0 dBm, −5 dBm, and −15 dBm, respectively. The method can effectively detect and accurately locate multiple types of weak interference. In a challenging scenario where the interference type is set to six, the interference transmission power is −15 dBm, and the number of interference sources is as high as 18, the proposed method can still achieve a detection recognition rate of over 60% and an interference localization accuracy within 1.8 km.

4. Conclusions

This study proposes a multi-interference source localization method based on target detection in GNSS receiver C/N₀ distribution maps, specifically addressing challenging scenarios, including the coexistence of multiple directional interferences, increased diversity and density of GNSS interference, and the presence of multiple low-power interference sources. Building upon established system and signal models suitable for GNSS interference source localization, we innovatively exploit the similarity between GNSS receiver navigation information and image grayscale information to construct C/N₀ distribution maps, thereby transforming the multi-interference source localization problem into a computer vision-based object detection task. To address the specific challenges in target detection and localization for our application scenario, we developed the OSF-RCNN framework by enhancing the Faster R-CNN architecture. The proposed method integrates an Oriented RPN to regress the orientation angles of directional antennas, effectively addressing their rotational characteristics. The FPN is used to fuse multi-scale feature information, enhancing the method’s adaptability to different types of interference sources and varying transmission powers. The SE mechanism is introduced at key positions within the network to enhance focus on weak interference targets, improving the detection accuracy of small targets. The paper presents the corresponding system and signal models, constructs training and testing datasets, and conducts simulation tests to evaluate the detection and localization performance of the proposed method. The results demonstrate that the method effectively detects and localizes GNSS interference sources across a range of challenging scenarios.