Classification of Targets by Using Frequency Modulated Continuous Wave Radar Data With Machine Learning Techniques

Introduction

Nowadays, the use of drones has been steadily rising with advancements in drone technology and increased market accessibility. This situation brings security risks in many fields. For instance, drones can be employed to secure critical facilities. Additionally, drones originally designed for civilian purposes have the potential to be weaponized and utilized in various forms of warfare, terrorist activities, and illicit operations such as smuggling [1–4]. Such events reveal the critical importance of detecting and classifying these platforms. Since Radar Cross Section (RCS) signatures of drones are relatively similar to other biological targets in some cases, they had similar speed profiles, making them difficult to classify with traditional radar classification algorithms. The limitations of traditional RCS and object velocity-based classification methods for drones have prompted a growing body of research focused on alternative detection approaches, including machine learning techniques.

There are several studies for improving the classification accuracy of this kind of target. One of these studies is the study conducted by Karlsson et al. [5] uses the convolutional neural networks (CNN) algorithm by preprocessing 77 GHz Frequency Modulated Continuous Wave (FMCW) radar data for drone classification. The data set used in this study includes In-phase Quadrature (IQ) radar data, range information of targets, and time information. The dataset used in this study contains a total of 75,868 data points. These data include measurements of six different drone classes, two different human activities (walking and running), six different bird classes, and a corner reflector. In this study, first, after the preprocessing of the acquired data, classification was performed with CNN. As a result of this classification, 90% success was achieved. To increase this rate, it is aimed to increase the success by improving the signal-to-noise ratio of the detections. In addition to the dataset used in this study, synthetic data were generated, and these synthetic data were also included in the study to improve performance. The dataset used in this study is openly available and can be accessed via open source [5].

In a study by White et al. [6], the authors perform this study with two datasets collected from L Band radar. The first dataset includes 35 bird classes and 10 drone class signature collections. With this dataset, a study was performed on CNN, and the performance for single-target data was achieved with an accuracy of 82%. The second dataset includes 64 bird classes and 95 drone classes. With the usage of this complex dataset, the accuracy of CNN will increase up to 89%–90% [6].

In a study by Zhu et al. [7], the authors provide a detailed examination of the development of radar automatic target recognition (RATR) technology and the applications of deep learning algorithms within this field. This paper shows that the classification accuracy of space targets can be improved through the utilization of the AlexNet and SqueezeNet models. Furthermore, the use of CNN algorithms for drones and aerial targets achieved an accuracy rate of 96.86%. This study concludes that deep learning algorithms offer a high classification accuracy rate for RATR technologies and provides a foundation for future research.

Kumawat et al. [7] focus on the classification of smaller aerial targets (such as drones) by using micro-Doppler signatures. In this study, a dataset which is called “DIAT-μSAT” was used. This dataset includes 4,849 micro-Doppler signatures and six different classes. They collect this dataset with custom CW X-Band radar. They used pre-trained networks such as VGG16 and VGG19 for the classification. The classification accuracies were obtained as 95% and 97% for VGG16 and VGG19, respectively.

In Narayanan et al. [8], the authors primarily focus on classifying birds and drones by analyzing their micro-Doppler spectrums. They use a dataset collected with costume 10 GHz(X-Bant) radar. This dataset includes 2,279 micro-Doppler signatures. In this study, the authors employed Support Vector Machines (SVM) for classification, achieving an accuracy of 90% when classifying drones based solely on their size. The accuracy for distinguishing between birds and drones separately was 96%. In the final case, the authors employed five different classes for drones and birds, achieving a classification accuracy of 85.

In our study, the dataset collected using SAAB SIRS 1,600 FMCW 77 GHz Radar [5] was utilized for the classification of six distinct drone classes, two human activity classes, six bird classes, and one corner reflector class. This dataset includes 75,856 different measurements. For the preprocessing of those data, raw IQ data are converted to amplitude frequency. Subsequently, all data are transformed into spectrogram images. Those spectrogram images are fed to deep learning and machine learning algorithms to classify the drones and birds.

The main contribution of this study, classification accuracy, is higher than state-of-the-art studies in the literature. Additionally, by combining the CNN algorithm with the linear SVM method, the classification accuracy surpasses that reported in existing literature.

Materials and Methods/Methodology

2.1.

Dataset

In this study, the dataset provided by Karlsson et al. [5] is used for the classification. This dataset includes drone, bird, human, and corner reflector measurements, which are collected by SAAB SIRS 1600 FMCW 77 GHz Radar. This dataset is published as open source by KTH Royal Institute of Technology. The parameters of the SAAB SIRS 1600 FMCW radar are described in Table 1.

Radar Type	SAAB SIRS 1600, FMCW
Center Frequency	77 GHz
Bandwidth	160 MHz
Range Resolution	1 m
Pulse Repetition Frequency (PRF)	17 kHz
Azimuth Beam width	1°
Scan Rate	10 Hz, mechanical
Field Of View	±9°
Output Power	10 mW

Table 1.

Radar specifications [5].

This dataset was published in the format of MAT-File, which is a file format for MATLAB software. This MAT-File includes a 130 × 6 cell array. In each row of 130 cells, the format of each measurement is saved. This dataset includes 75,868 measurements. The classes included in the dataset are shown in Table 2 and Figure 1. The distribution of each measurement for classes is shown in Table 3.

Class	Description
“D1”	Drone, DJI matrice 200 v
“D2”	Drone, DJI Mavic 2 Pro
“D3”	Drone, DJI Phantom 3
“D4”	Drone, custom-built FPV drone
“D5”	Drone, Custom built with a Tarot 680 Pro frame
“D6”	Drone, Syma X23W
“human_walk”	Different humans walking, ranging from almost still to “normal” walking
“Human_run”	Different humans running
“seagull”	Flying Seagull
“pigeon”	Flying Pigeon
“raven”	Flying Raven
“black-headed gull”	Flying Black-headed gulls
“seagull and black-headed gull”	A mixture of flying seagulls and black-headed gulls.
“heron”	Flying Heron
“CR”	Triangular Corner Reflector, 100 m2 RCS at 77 GHz

Table 2.

Classes in the dataset [5].

Class	Number of Measurements
Drone, DJI matrice 200 v	6,921
Drone, DJI Mavic 2 Pro	9,555
Drone, DJI Phantom 3	10,212
Drone, custom-built FPV drone	8,735
Drone, Custom built with a Tarot 680 Pro frame	12,093
Drone, Syma X23W	12,093
Human walking and running	6,028
Flying Birds. With and without wing flaps.	7,792
Trihedral Corner Reflectors	3,280

Table 3.

Distribution of measurements by class.

The second column in the dataset contains the IQ data, which represent the combined values of five scan segments. Since the radar has the capability of mechanical scan, each measurement includes five scan segments. Each scan segment is a vector of size 1 × 256. In the case of combining five scans, each set of five scan segments is represented as a vector of size 1 × 1,280. An example of five scan segments for the “D1” class is presented in Figure 2.

Of these data, 57,868 were used for training, 9,000 for validation, and the remaining 9,000 for testing the model. This split was provided by the dataset creators, and the data were downloaded without any modification to ensure comparability with the literature.

2.2.

Preprocessing Stage

Before applying machine learning and deep learning techniques, the dataset needs to be rearranged and preprocessed. As described in the dataset section, the dataset includes raw IQ data. First, all IQ data was converted frequency domain with the usage of Short Time Fourier Transform (STFT). For this purpose, Eq. (1) and Eq. (2) are used.

(1)

(2)

Here, x(t) is the input signal, w(t) is the window function, and t is the time at which the window is centered. f is the frequency, and STFT (t, f) is the STFT of the input signal. By taking the square of the magnitude of STFT (t, f), a power spectrogram is obtained. Since the pulse repetition frequency is 17 kHz according to radar specifications, the FFT sampling frequency was selected to be 17 kHz. Additionally, the IQ data consists of 256 samples for each scan segment; the FFT sampling point is therefore set to 256. A Hamming window of length 256 is selected for the STFT. After computing the STFT, the resulting spectrogram is plotted and saved in “.jpeg” image format. Examples of the spectrogram of different classes are shown in Figure 3.

After the spectrograms were saved, the dataset classes were rearranged following the approach of Karlsson et al. [5] to facilitate a comparison of those of our study. Two human classes are combined as one class. Also in bird class, six bird classes are combined as one class as shown in Figure 4.

Experimental Studies

3.1.

Convolution Neural Networks From Scratch

Different CNN architectures were designed to identify the model with the highest performance. Many hyperparameters affect the success of the CNN algorithm. Upon reviewing the literature, four parameters (input data size, filter size, learning rate, and convolution layer number) considered to influence performance have been addressed in this study. First, the architecture consisting of five-layer convolutional blocks, as presented in Figure 5, was constructed from scratch, and the change in performance of this architecture was analyzed as the input data size, filter size, and learning rate varied.

First, the accuracy and performance achieved by using five scan segments as input are compared to those obtained with a single scan segment. Upon examining the results presented in Table 4, it was determined that the use of five scan segments yielded successful outcomes. Consequently, five scan segments were applied as input throughout the remainder of the study.

Input Data	Training Accuracy	Accuracy and Standard Deviation	Precision	Recall	F1-Score	KAPPA	Process Time
Five Scan Segment (100 × 100 pixel size)	97.99	97.72 ± 1.32	97.74	97.72	97.73	0.9744	837 min 50 sec
One Scan segment (100 × 100 pixel size)	91.88	91.69 ± 0.94	92.07	91.69	91.72	0.9065	824 min 24 sec

Table 4.

Comparison of the results when five scan segments and one scan segment are used as input.

Following this step, the first technique involves modifying the input image pixel sizes in the five-layer CNN algorithm. In this method, the input sizes are altered to 100 × 100, 300 × 300, 500 × 500, and 600 × 600. Example images with modified input pixel sizes are presented in Figure 6. In this context, a comparison is made between training time and overall accuracy based on different success criteria. Based on the results (Table 5) obtained from the experiments, the input data size yielding the highest model performance was determined. The study continued with this data size.

Pixel Size	Training Accuracy	Accuracy and Standard Deviation	Precision	Recall	F1-Score	KAPPA	Process Time
100 × 100	97.99	97.72 ± 1.32	97.74	97.72	97.73	0.9744	837 min 50 sec
300 × 300	96.84	96.9 ± 1.45	96.94	96.9	96.9	0.9651	857 min 50 sec
500 × 500	96.8	96.44 ± 1.24	96.46	96.44	96.44	0.96	2005 min 5 sec
600 × 600	96.92	94.53 ± 1.08	94.86	94.53	94.58	0.9385	3,016 min 8 sec

Table 5.

Comparison of the effect of changing the pixel size of the input data on the results.

The performance of different adaptive learning rate algorithms was evaluated. In the approach utilizing adaptive learning algorithms, the OneCycleLR and ReduceLROnPlateau techniques are employed. For both techniques, two cases are considered. In the first case, the minimum learning rate is set to 10^–3 and the maximum learning rate to 10^–2. In the second case, the minimum learning rate is set to 10^–4 and the maximum learning rate to 10^–3. After applying these methods, they were compared based on their training time and overall accuracy according to various success criteria in Table 6.

Adaptive Learning Rate Algorithm	Training Accuracy	Accuracy and Standard Deviation	Precision	Recall	F1-Score	KAPPA	Process Time
OneCycleLR (0.01–0.001 learning rate)	97.88	97.47 ± 0.97	97.56	97.47	97.48	0.9715	1,025 min 42 sec
OneCycleLR (0.001–0.0001 learning rate)	96.34	95.96 ± 1.07	96.1	95.96	95.98	0.9545	1,063 min 29 sec
ReduceLROnPlateu (0.01–0.001 learning rate)	96.88	96.48 ± 1.16	96.51	96.48	96.47	0.9604	810 min 53 sec
ReduceLROnPlateuOnLR (0.001–0.0001 learning rate)	96.3	95.66 ± 1.18	95.83	95.66	95.69	0.9511	1,053 min 29 sec

Table 6.

Comparison of the results when different adaptive learning rate algorithms are applied.

One of the methods involves changing the filter size in the five-layer CNN. Initially, filter sizes of 3 × 3, 5 × 5, and 7 × 7 are used for all layers. The main objective of altering the filter size is to improve overall performance. To evaluate the impact of changing filter sizes on the performance of the five-layer CNN, training time and overall accuracy are compared in Table 7.

Filter Size	Training Accuracy	Accuracy and Standard Deviation	Precision	Recall	F1-Score	KAPPA	Process Time
3 × 3	97.99	97.72 ± 1.32	97.74	97.72	97.73	0.9744	837 min 50 sec
5 × 5	97.47	97.22 ± 1.49	97.23	97.22	97.22	0.9688	890 min 18 sec
7 × 7	96.38	96.34 ± 1.46	96.45	96.34	96.35	0.9589	890 min 36 sec
9 × 9	96.98	96.98 ± 1.38	9	96.98	96.97	0.966	910 min 49 sec

Table 7.

Comparison of the results when different filter sizes in the convolution layers are selected.

After the input data size, learning rate, and filter size were determined through experimental studies, the number of convolutional layers was varied using the specified parameters to evaluate its impact on the model’s performance.

Initially, the number of convolutional layers was set to 5, but it was reduced to 4 and then increased to 6. The effect of the number of convolutional layers on model performance was examined as shown in Table 8. The block diagrams of the models with four and six convolutional layers are presented in Figure 7 and 8, respectively.

Convolutional Layer Number	Training Accuracy	Accuracy and Standard Deviation	Precision	Recall	F1-Score	KAPPA	Process Time
Six-layer CNN	98.26	97.77 ± 0.79	97.78	97.77	97.77	0.9749	1,000 min 41 sec
Five-layer CNN	97.99	97.72 ± 1.02	97.74	97.72	97.73	0.9744	837 min 50 sec
Four-layer CNN	97.28	97.44 ± 1.21	97.54	97.44	97.46	0.9713	787 min 29 sec

Table 8.

Comparison of the results when the number of convolution layers is changed.

Results and Discussions

With the usage of the methods described in the methods and experimental studies section, all results are shown in Table 4–10. In the first analysis, the main objective was to compare the accuracy difference between using five scan segments and one scan segment. As shown in the results in Table 4, using five scan segments significantly outperforms the one scan segment approach in classification. Therefore, subsequent methods continued with the five scan segments. In the second analysis, the effect of input spectrogram image pixel size on classification performance was evaluated using a five-layer CNN. The filter sizes were adjusted as described in the Methods and Experimental Studies section. As shown in Table 5, a pixel size of 100 × 100 yielded the highest classification accuracy. Adaptive learning rate algorithms were applied to the five-layer CNN. In this process, the OneCycleLR and ReduceLROnPlateau algorithms were used with similar minimum and maximum learning rates as specified in the Methods section. As shown in Table 6, the OneCycleLR algorithm outperformed the other adaptive learning rate algorithms with a minimum learning rate of 0.01 and a maximum of 0.001. However, since the accuracy of the OneCycleLR algorithm was still lower than that of the five-Layer CNN with a 100 × 100 pixel size input, the goal was to improve the five-layer CNN structure by modifying the filter sizes. Following the instructions in the Methods and Experimental Studies section, the five-layer CNN structure was modified with different filter sizes (5 × 5, 7 × 7, and 9 × 9). However, changing the filter sizes did not improve the accuracy. The five-layer CNN with a 3 × 3 filter size provided better accuracy than the models with larger filter sizes shown in Table 7. As a result, a new approach was tested to improve accuracy by modifying the CNN structure by adding and removing layers. This method demonstrated that a six-layer CNN improved overall classification accuracy which is shown in Table 8, although not to the desired extent. Consequently, the decision was made to train and classify the dataset using pre-trained networks, as described in the Methods and Experimental Studies section. The results given in Table 9 showed that pre-trained networks delivered good classification performance (over 95% accuracy); however, this approach did not lead to a significant improvement in overall accuracy. Based on these findings, a hybrid method combining CNN with classical machine learning techniques was explored. In this approach, a six-layer CNN was used for feature extraction, and classical machine learning algorithms were employed for training and testing the dataset. The results given in Table 10 indicate that the combination of the six-Layer CNN and the Linear SVM algorithm achieved an accuracy of 98.04%. The confusion matrix for this result is shown in Figure 9.

The proposed study integrates both conventional machine learning algorithms and a novel methodological perspective to assess radar-based object classification performance. All models were implemented and analyzed within the MATLAB environment to ensure consistent evaluation conditions. Experimental results indicate that the effectiveness of each algorithm may vary depending on the dataset characteristics and radar signal structure. The proposed approach can be further examined by adapting datasets of different scales and radar platforms to evaluate its generalization capability. Given the increasing computational power of modern radar systems, the suggested methods are expected to be practically implementable and computationally efficient for real-time or near-real-time operations.

When the results of all experimental studies obtained in Table 4 through Table 10 are examined, the highest success was achieved when the features were extracted with the six-layer CNN algorithm created from scratch and classified with the SVM method.

The performance of all analyses is summarized in Figure 10. It can be seen from Figure 10 that the parameter that most affects the model’s accuracy is the scan size of the input data. The next most influential parameter is the pixel size of the input data. It is understood from Figure 10 that other parameters also have an impact on accuracy. Based on these results, the proposed architecture employs 5-scan input data with a resolution of 100 × 100 pixels, utilizes the OneCycleLR learning rate scheduling algorithm, adopts a 3 × 3 filter size, and consists of six layers. The features extracted from the proposed architecture were applied to different machine learning algorithms, and the results are given in Table 10.

Compared with the studies in the literature (Table 11), our method achieved higher accuracy. For example, while previous works using CNNs and transfer learning on different radar types reported accuracies between 90% and 97%, our hybrid CNN–SVM model reached 98.8% on the Karlsson et al. [5] dataset. This demonstrates that combining deep feature extraction with SVM classification provides state-of-the-art performance for drone and bird classification.

However, certain limitations must be acknowledged. The study relies exclusively on a single publicly available dataset, which restricts generalizability to other radars and operational environments. Additionally, small differences in accuracy (e.g., 97.7% vs. 98.0%) may not be statistically significant. Statistical analyses could not be performed in this study because the data sets were different. Future work should include statistical validation (confidence intervals and hypothesis testing) to reinforce these findings. Moreover, while training times are reported, the practical deployment of CNN–SVM models in real-time systems requires analysis of computational cost and scalability. Graphics Processing Unit (GPU) acceleration or embedded platforms can enable near real-time inference, while FPGA implementations could further reduce latency and energy consumption.

Conclusion(s)

Nowadays, due to the widespread production of drones and their easy accessibility, their use in different areas has become common. Because of the risks these uses may create, the importance of detecting drones with different sensors is increasing day by day. This study addressed this problem by exploiting micro-Doppler signatures and evaluating multiple machine learning approaches. The results demonstrated that most tested methods achieved accuracies above 90%, confirming the promise of CNN architectures, pre-trained models, and their combination with conventional classifiers for small air target classification.

The main contribution of this work lies in the proposed hybrid CNN–SVM approach, which achieved up to 98.8% accuracy. Beyond accuracy, this hybrid design improved stability on relatively small datasets by reducing overfitting and demonstrated state-of-the-art performance compared to standalone CNN or SVM models. Nevertheless, the study has limitations: it relied exclusively on a single publicly available dataset, and the observed accuracy improvements may not be statistically significant without further validation. Future research should therefore evaluate the approach with multiple radar datasets, different operating frequencies, and real-world collected signals to ensure broader generalizability.

From an application standpoint, the practical deployment of the proposed approach in real-time systems requires careful consideration of hardware efficiency and processing constraints. Recent advances in embedded GPU and FPGA technologies offer promising opportunities for achieving low-latency inference, which is essential for time-critical radar applications such as airport surveillance, infrastructure protection, and defense monitoring. Nevertheless, operational deployment may face challenges arising from noise, clutter interference, multiple moving targets, and adverse environmental conditions. Future work could explore optimized implementations and adaptive signal-processing techniques to enhance system robustness under real-world condition.

In conclusion, this work provides a strong foundation for radar-based drone and bird classification using micro-Doppler signatures. With validation on diverse datasets and optimization for real-time hardware, the proposed CNN–SVM approach has significant potential for practical use in modern security systems.

Acknowledgments

AI tools have been used to improve the spelling of some sentences.

Author Contributions

Emre Can Ertekin: Conceptualization, Writing – original draft; Selda Güney: Conceptualization, Writing – review & editing.

Funding

This research did not receive external funding from any agencies.

Ethical statement

Not applicable.

Data availability statement

Source data presented in this article is available in the study by Karlsson A., “Radar measurements on drones, birds and humans with a 77GHz FMCW sensor.” Sep. 2021. [Online]. Available: https://doi.org/10.5281/zenodo.5,845259.

Conflict of Interest

The authors declare no conflict of interest.