Corner Reflector Array Jamming Discrimination Using Multi-Dimensional Micro-Motion Features with Frequency Agile Radar

Corner Reflector Array Jamming Discrimination Using Multi-Dimensional Micro-Motion Features with Frequency Agile Radar 1st Jie Yuan, 2nd Lei Wang*, 3rd Yanhao Wang , 4th Yimin Liu

arXiv:2604.16008v1 [cs.LG] 17 Apr 2026

Department of Electronic Engineering, Tsinghua University Beijing, China Abstract—This paper introduces a robust discrimination method for distinguishing real ship targets from corner-reflectorarray jamming with frequency-agile radar. The key idea is to exploit the multidimensional micro-motion signatures that separate rigid ships from non-rigid decoys. From Range–Velocity maps we derive two new hand-crafted descriptors—mean weighted residual (MWR) and complementary contrast factor (CCF)—and fuse them with deep features learned by a lightweight CNN. An XGBoost classifier then gives the final decision. Extensive simulations show that the hybrid feature set consistently outperforms state-of-the-art alternatives, confirming the superiority of the proposed approach. Index Terms—frequency agile radar, corner reflector array, micro-motion features, target recognition, XGBoost

I. I NTRODUCTION Effectively countering various passive and active interferences is a key capability of radar systems [1], [2]. Corner reflectors represent a typical type of passive interference false target in radar electronic countermeasure environments. Leveraging their metallic composition and unique geometric structure, corner reflectors can achieve a radar cross section (RCS) comparable to that of a large ship, despite their small physical size. Multiple corner reflectors can be arranged into an array deployed around a vessel, simulating ship targets through various configurations, thereby effectively deceiving anti-ship guidance radar detection and tracking systems. Consequently, analyzing the differences in radar echoes between ships and corner reflector arrays, and utilizing these distinctions for target identification, is of significant research importance. A unique correspondence exists between a target’s micromotion features and its structural and kinematic properties. Based on this principle, ships and corner reflectors exhibit notable structural differences and demonstrate distinct kinematic behaviors under identical sea conditions, allowing for their discrimination using micro-motion characteristics. Reference [2] investigated the motion model of sea-surface corner reflectors and the extraction of micro-Doppler features, demonstrating that micro-Doppler can be regarded as a unique signature of the target that provides additional information for target recognition [3], [4]. Meanwhile, reference [5] proposed a method for extracting relevant time-frequency characteristic This work was supported by the National Natural Science Foundation of China under Grants 62301295 and the corresponding author is Lei Wang (email: [email protected]).

parameters from the time-frequency representation of target echoes to distinguish between corner reflector jamming and ships, where reliable micro-Doppler signature classification requires the use of robust features that are capable of uniquely describing the micromotion [6]. However, current research on discriminating ship targets from corner reflector jamming based on micro-motion features predominantly focuses on the impact of single corner reflectors, leaving a notable gap in mature methodologies for dealing with corner reflector array jamming. This paper proposes a recognition method for corner reflector array jamming based on multi-dimensional micromotion features. First, we establish a target echo signal model for frequency-agile radar [7] and generate Range-Velocity maps(RV Maps) of the target through signal processing. By analyzing the differences in micro-motion characteristics between corner reflector arrays and ship targets, we design and extract two key discriminative features: the mean weighted residual (MWR) and the complementary contrast factor (CCF). These handcrafted features are subsequently combined with deep representations automatically extracted by a convolutional neural network (CNN) to form a multi-dimensional fused feature vector. Finally, an XGBoost classifier is employed to achieve high-accuracy identification of ships versus corner reflector arrays.Extensive comparative simulation experiments demonstrate the effectiveness of the proposed method.The method proposed in [8], which addresses the ballistic target recognition using multi-dimensional micromotion features in frequency-agile radars, provides an important reference for this work. II. F REQUENCY-AGILE R ADAR E CHO M ODEL AND TARGET RV M AP G ENERATION A. Frequency-agile Radar Echo Signal Model To obtain high resolution for each pulse in frequency-agile mode and support subsequent synthetic wideband processing, we adopt the Linear Frequency Modulated (LFM) signal as the baseband waveform. The signal model for the n-th pulse can be expressed as:      t̂ 1 2 sn (t) = rect exp j2π fn t̂ + µt̂ , (1) Tp 2

where we define t̂ = t − tn as fast time, tn = nTr is slow time, n = 0, 1, ..., N − 1, N is the number of coherent processing pulses, Tr is the pulse repetition interval (PRI), Tp is the pulse width, rect(t) equals 1 when 0 ≤ t ≤ 1 and 0 otherwise, µ = B/Tp is the chirping rate and B is bandwith, fn = f0 + Cn ∆f is the carrier frequency for the n-th pulse, f0 is the radar’s initial carrier frequency, Cn is the random frequency-hopping code, which is a random integer from the set{0, 1, ..., M − 1}, M is the number of distinct frequency points, and ∆f is the frequency step. Considering the target as a composition of scattering centers, the final received target echo can be expressed as the linear superposition of echoes from all scattering centers: rn (t) =

K X

 σk rect

k=1

t̂ − τn,k Tp

 · (2)



  1 2 exp j2π fn (t̂ − τn,k ) + µ(t̂ − τn,k ) , 2 where σk and τn,k are the scattering amplitude and the time delay of the k-th scattering center, respectively, and τn,k is defined as τn,k = 2(Rk + vk nTr )/c. Rk and vk are the initial radial distance and radial velocity, respectively, between the k-th scattering center on the target and the radar [9], [10], and c is the speed of light. The received echo signal is downconverted to obtain: r̂n (t) =

K X

 σk rect

k=1

t̂ − τn,k Tp

 · (3)

  1 2 . exp j2π −fn τn,k + µ(t̂ − τn,k ) 2 

B. Echo Signal Processing and Target RV Map Generation According to (1), the frequency-domain representation of the transmitted signal sn (t) can be obtained as: r Sn (f ) =

1 rect µ



f B



  f2 · exp −jπ . 2µ

(4)

Therefore, the frequency-domain representation of r̂n (t) can be expressed as:

R̂n (f ) =

K X

r σk

k=1

1 rect µ



f B



  2  f exp −j2π + fn τn,k 2µ (5)

After pulse compression, the frequency-domain representation of the signal is obtained as:

Rpc (f ) =

K X k=1

1 σk rect µ



f B

 exp(−j2πfn τn,k ).

(6)

After Doppler compensation, the frequency-domain representation of the signal is:   2vfn f Rcomp (f ) = Rpc (f ) · exp −j2π cµ      (7) K X σk f 2vf rect exp −j2πfn + τn,k . = µ B cµ k=1

Considering range cell migration correction, the frequencydomain expression of the compensated signal is:   2vnTr Rrmc (f ) = Rcomp (f ) · exp −j2πfn c        K X σk f 2v f = rect · exp −j2πfn + nTr + τn,k . µ B c µ k=1 (8) Since the carrier frequency of each pulse varies and is typically non-continuous, each spectral segment needs to be resampled onto a corresponding unified frequency axis. According to [11], by accumulating signals in phase, we can i obtain S(fi ) = N1i ΣN ni =1 Rrmc,ni (fi ), where Rrmc,ni (fi ) is the spectrum of the echo signal from the ni -th pulse using carrier frequency fi after Doppler compensation and range migration correction, and Ni denotes the number of pulses using carrier frequency fi .Subsequently, the coherently integrated signal spectrum at each frequency point is resampled onto the synthesized frequency axis to obtain the complete target echo signal spectrum within one CPI. By performing an IFFT on this spectrum, the target’s high-resolution range profile(HRRP) can be obtained.The total bandwidth is Bsynth = M · ∆f , and the range resolution is ∆R = c/2Bsynth . Since the radial velocities of individual scattering points on the target vary,we set multiple velocity grid points vi and perform Doppler compensation and range migration compensation on the target echoes according to the value of each grid point.This allows us to extract the radial velocities of the scattering points on the target through peak matching. Each vi corresponds to a HRRPi [12].By concatenating this series of HRRPi along the dimension of the predefined velocity grid points, a RV map of the target can be generated: RV = |Joint(HRRPi )|, where “Joint” means concatenating the HRRP generated at each grid point along the velocity grid dimension. The velocity resolution is: ∆v = λ/2N Tr . This map explicitly represents the coupled distance and velocity information of each scattering center on the target as detected by the radar. . III. R ECOGNITION M ETHOD D ESIGN FOR M ARITIME TARGETS A. Maritime Target Micro-motion Modeling and Characteristic Analysis Micro-motion refers to the small-amplitude, high-frequency, or quasi-periodic self-motions of a target relative to its main translational motion. The micro-motion of maritime targets is induced by wave-driven disturbing forces that push the target

(a)

Fig. 1. The spatial relationship between the radar and the target [5]

into movements in different directions.In three-dimensional space, it encompasses six degrees of freedom of motion, including three translational degrees of freedom—surge, sway, and heave—and three rotational degrees of freedom—roll, pitch, and yaw.These variables describe the translational motions of the target along its longitudinal, transverse, and vertical axes (surge, sway, and heave, respectively), as well as the rotational oscillations around its longitudinal, transverse, and vertical axes (roll, pitch, and yaw, respectively).Among these, the three rotational degrees of freedom are directly driven by wave-induced torques and can generate significant micro-Doppler effects, while the influence of the translational degrees of freedom is relatively uniform and their impact is less pronounced.Therefore, the micro-motion of maritime targets can be simplified to a model comprising only the three rotational micro-motions(roll, pitch, and yaw) [13]. The spatial relationship between the radar and the target is shown in Fig.1. OXY Z is the radar observation coordinate system, and Ot Xt Yt Zt is the target’s body-fixed coordinate system.In the initial state, the planes OXY and Ot Xt Yt are parallel. The line connecting the origins of the two coordinate systems, OOt , represents the line-of-sight (LOS) direction of the radar observing the target. Its length is the radar-to-target distance R0 . The angle between OOt and the plane Ot Xt Zt (blue plane) is the azimuth angle θ of the observed target, while the angle between OOt and the plane Ot Xt Yt (orange plane) is the elevation angle φ of the observed target. The target attitude angle γ, induced by roll, is defined as the angle between the axis Ot Yt and the plane OXY . Micro-motions in different dimensions generate corresponding micro-motion velocities, and the micro-motion velocities of different parts of the target vary.Under radar observation, this manifests as a small incremental component ∆v superimposed on the overall translational velocity of each scattering center. This increment represents the projection of the total micro-motion velocity onto the radar radial direction, which differs for each scattering center. Evidently, ship is rigid body, while corner reflector array is not. For rigid target, there exists a well-defined kinematic con-straint relationship between the micro-motion velocities of scattering centers on it. Taking the pitch motion of the ship in Fig. 2 as an example: two scattering centers are selected at the bow (A) and the stern (B), respectively.When

(b)

Fig. 2. Rigid/Non-rigid target RV Map: (a) ship; (b) corner reflector array

A experiences a wave-induced velocity Vp1 as shown in the figure, the rigid body constraints of the entire hull cause B to simultaneously exhibit a corresponding velocity Vp2 .In contrast, for non-rigid corner reflector arrays, since each corner reflector operates as an independent unit, the micro-motion states of scattering centers across individual reflectors exhibit random characteristics.In the target’s RV Map, this is reflected as follows: the range-velocity distribution of scattering points on a rigid ship exhibits a strong linear trend, as shown in Fig. 2a, whereas the corner reflector array demonstrates a more random and scattered distribution with poorer linearity, as illustrated in Fig. 2b. Since radar can only detect radial distance and velocity, for the same target in an identical scenario, the observed range and velocity spread of the target will vary with the radar’s viewing angle.For rigid target (ship), its overall range and velocity distributions exhibit a coupled relationship: when the radar-detected range spread of the target is large, its velocity spread is small, and vice versa.But for non-rigid target(corner reflector array), the coupling between these two distributions is significantly weaker: there is no strong correlation between the range spread and velocity spread—they can easily exhibit either large or small spreads simultaneously. B. Feature Extraction and Recognition Method Design Firstly, the range-velocity coordinates (ri , vi ) and the weight wi (calculated based on scattering intensity) of each scattering center on each RV diagram are extracted, where i ∈ {1, ..., Nsc }, Nsc is the number of coordinates of target scattering centers.The following two feature computation methods are designed based on the analysis in III-A: (1)Feature 1:Mean Weighted Square Residual of RangeVelocity Linear Fit: we calculate the r −v linear regression fitting for each map using the least squares method: v̂i = ari +b, both a and b are fixed parameters determined through the fitting calculation.The mean weighted square residual(MWR) of each map can be calculated: MWR =

Nsc 1 X wi (vi − v̂i )2 . Nsc i=1

(9)

MWR represents the degree of linearity among the scattering points in an RV Map. Typically, the RV Map for a ship

(a)

(b)

Fig. 4. Schematic of the target models:(a) ship; (b) corner reflector array

Fig. 3. The flowchart of recognize corner reflector array and ship based on feature extraction from target RV Maps

demonstrates significant linearity, corresponding to a smaller MWR value. In contrast, the pattern observed for a corner reflector array is the opposite. (2)Feature 2:Range-Velocity Distribution Complementary Contrast Factor:we calculate the respective standard deviations to characterize the spread distribution in the range and velocity dimensions(σr and σv ): v v u u Nsc Nsc u 1 X u 1 X t 2 σr = (ri − r) , σv = t (vi − v)2 . (10) Nsc i=1 Nsc i=1 Then we calculate a complementary contrast factor (CCF) as: CCF =

σr − σ v , σr + σ v + ϵ

(11)

where ϵ is a very small number to prevent division by zero, set here as ϵ = 10−6 .When σr and σv are simultaneously large or small, the CCF will be relatively small, indicating non-rigid characteristics (corner reflector array). Conversely, when the two exhibit a complementary relationship (one large and one small), the CCF becomes larger, representing rigid characteristics (ship). It is important to note that since ri (range) and vi (velocity) typically differ in order of magnitude, they need to be normalized to a comparable scale before calculating the CCF. Particularly, even when normalized to the same order of magnitude, the differences between the σr and σv cannot be readily dismissed, thus making them viable features for recognition. In addition, we employ a CNN for automatic feature extraction(denoted as λCN N ) and combine these learned features with the manually designed features as input to the recognition method [14]–[16].Considering the varying importance of these features for target recognition, we have selected XGBoost as our final classifier model [17]–[19].This model, through its powerful ensemble learning capabilities and built-in regularization, significantly enhances training speed and resistance to

overfitting while ensuring model accuracy. We assign the label ′ ′ 1 to ship targets and ′ 0′ to corner reflector arrays. The feature vector [MWR, CCF, σr , σv , λCN N ] extracted from the target RV Map is then used as input to the model for supervised training [20], [21].The model prediction output discriminates whether the target is a ship or a corner reflector array. Fig.3 presents the complete workflow of the recognition process for using the targets’ RV Maps. IV. E XPERIMENTS Simulated experimental data is used to validate the performance of the proposed method.The targets include one ship and one array composed of four linearly arranged icosahedral corner reflectors, as shown in Fig.4. The ship target has a length of approximately 144m, while each corner reflector has a diameter of about 1m, with an inter-element spacing of roughly 40m. We employ DualSPHysics (an open-source SPH-based fluid/multi-physics solver) to set up irregular waves with a JONSWAP spectrum and apply a second-order (Stokes II) nonlinear correction to simulate realistic ocean wave conditions [22].Considering that different wave amplitudes have distinct effects on the target’s motion state, we configured wave scenarios with significant wave heights of 0.1m, 1m, and 2m respectively. Motion simulations were conducted for both the ship and the corner reflector array under identical conditions, yielding attitude angle variation data (yaw, pitch, roll) for the targets.Based on the spatial relationship between the radar observation coordinate system and the target body coordinate system, we transformed the target’s intrinsic attitude angle variations into changes in the azimuth and elevation angles observed by the radar.The frequency response of the target at different aspect angles is then calculated using CST Studio Suite(CST) software, which implicitly contains information about the target’s motion state.The frequency band used here is the Ku-band. Particularly, according to Eq.5,we can reformulate it into the following form: r     1 f f2 rect exp −j2π R̂n (f ) = µ B 2µ ×

K X

σk exp(−j2πfn τn,k )

(12)

k=1

PK where k=1 σk exp(−j2πfn τn,k ) the coherent sum of the frequency responses from all scattering centers on the target,

TABLE I: Radar Simulation Parameters Parameter Initial carrier frequency Carrier frequency number Frequency step ∆f Pulse width Sampling frequency Synthetic bandwidth PRI SCR Pulse number of a CPI

(a) θ = 80◦ , ship

Configuration 16GHz 8 25MHz 36µs 50MHz 200MHz 250µs 20dB 256

TABLE II: Hyperparameters of the XGBoost classifier Parameter

Description

nest depthmax learning rate subsample colsample bytree objective eval metric random state

Number of boosting trees Maximum depth of each tree Step size shrinkage Subsampling rate of training instances Fraction of features used per tree Classification objective Evaluation metric Random seed

Value 200 6 0.05 0.8 0.8 binary:logistic logloss 42

(b) θ = 80◦ , corner reflector array

Fig. 6. Variation of accuracy with significant wave height

(c) θ = 170◦ , ship

(d) θ = 170◦ , corner reflector array

Fig. 5. Comparison of RV Maps between ship and corner reflector array targets under two observation angles

which can be substituted by the target’s overall frequency response obtained through software CST simulations. The radar waveform used in the experiment is LFM. The specific simulation parameters are listed in Table I.With the target directly facing the radar (both the bow of the ship and each corner reflector aligned toward the radar) defined as the 0-degree azimuth angle and counterclockwise as the positive direction, we set up 17 observation perspectives with azimuth angles ranging from 10 to 170 degrees in 10-degree increments (θ = 10◦ → 170◦ , θstep = 10◦ ) and an elevation angle of approximately 2 degrees(φ ≈ 2◦ ).Under each wave height scenario and for every observation angle, 100 RV Maps were generated for both the ship and the corner reflector array targets. Since the corner reflector array is composed of multiple independent corner reflector units, the echoes from each individual corner reflector must be linearly superimposed during simulation before signal processing to generate the target RV Map. Fig.5 demonstrates a comparison of the RV maps generated by ship and corner reflector array targets under two different observation(θ = 80◦ , 170◦ ) angles in a scenario with a significant wave height of 1m. The red bounding box

encloses the (r, v) coordinate region of the scattering points on the target. Its horizontal extent represents the length spread of the detected target, while its vertical extent corresponds to the velocity spread. It can be visually observed that under the same observation angle, the RV Map of the ship target exhibits better linearity than that of the corner reflector array. Under different observation angles, the range spread and velocity spread of the ship target demonstrate a complementary relationship, whereas the corner reflector array does not. This observation aligns with our theoretical analysis. The generated RV Map dataset was divided into training and testing sets in a 7:3 ratio. Training and testing were conducted using the proposed feature extraction and recognition methods. The configuration of the XGBoost model is detailed in Table II.We conducted tests on data under different wave height conditions and performed a comparative study using three experimental groups: Handcrafted Features + XGBoost, CNN Features + XGBoost, and Fused Features + XGBoost. The corresponding accuracy results are presented in Fig 6.As can be intuitively observed, the recognition accuracy increases with the rising wave height. This trend is primarily attributable to the fact that higher waves enhance the manifestation of nonrigid characteristics in the arrangement of the corner reflector array.Furthermore, the method we adopted demonstrates higher accuracy compared to all other control groups. This not only validates the effectiveness of integrating handcrafted features with CNN-based features to form a fused feature vector but also indicates the superior classification performance of XGBoost for this specific problem. It is particularly noteworthy that using only our customdesigned feature vectors (without incorporating CNN-

(a)

(b)

Fig. 7. T-SNE visualization of the features(blue for corner reflector array and orange for ship): (a) handcrafted features; (b) fused features

extracted features) in conjunction with XGBoost achieves a recognition accuracy of over 86%. This result demonstrates that our feature extraction method can effectively capture the distinguishing characteristics between the two types of targets. The t-SNE visualization of the features extracted from the two types of targets is shown in Fig7 (Fig 7a for handcrafted features, Fig 7b for fused features). As observed in Fig7, the feature vectors we extracted can effectively distinguish between the two types of targets to a considerable extent. The overlapping regions are primarily attributed to instances where the corner reflector array may exhibit rigid-body characteristics under certain scenarios, leading to misclassification. The tSNE visualization under the fused feature set demonstrates a clearer separation between the two classes. V. C ONCLUSIONS In this paper, we have developed an efficient radar target recognition framework that integrates handcrafted features with deep learning representations to discriminate between ships and corner reflector array jamming. The key innovation is our approach to leveraging the distinct multidimensional micro-motion characteristics of both targets. We design and extract two new features, MWR and CCF, directly from the target’s RV Map to capture discriminative information on structural rigidity and motion distribution.These interpretable handcrafted features are combined with deep representations extracted by a CNN and fed into an XGBoost classifier. Extensive simulations under various sea states and radar viewing angles demonstrate that the proposed method achieves high recognition accuracy, particularly in higher sea conditions where non-rigid micro-motion patterns are more pronounced. It is worth emphasizing that the handcrafted multidimensional micro-motion features alone achieve over 86% accuracy, highlighting both their strong interpretability and effectiveness. Future work will expand the application of our method to more diverse scenarios and evaluate the proposed multidimensional micro-motion features using real-world measured data. R EFERENCES [1] Y. Su, L. Wang, X. Lu, C. Liu, and Y. Liu, “Multiple mainlobe jamming reconstruction and suppression in wideband distributed radars,” IEEE Trans. Radar Syst., vol. 3, pp. 1362–1374, 2025.

[2] J. Huang, J. Chen, Z. Zhao, and J. Zhao, “Micro-Doppler modeling and simulating of corner reflector in sea surface,” Syst. Eng. Electron., vol. 34, no. 9, pp. 1781–1787, Sep. 2012. [3] C. Clemente, A. Balleri, K. Woodbridge, et al., “Developments in target micro-Doppler signatures analysis: Radar imaging, ultrasound and through-the-wall radar,” EURASIP J. Adv. Signal Process., vol. 2013, no. 47, 2013. [4] V. C. Chen, “The micro-Doppler effect in radar”. Norwood, MA, USA: Artech House, 2019. [5] Z. Hong, W. Qing-ping, P. Yu-jian, T. Ning, and Y. Nai-chang, “A sea corner-reflector jamming identification method based on time-frequency feature,” in Proc. IEEE Int. Conf. Signal Process., Commun. Comput. (ICSPCC), Ningbo, China, 2015, pp. 1–6. [6] C. Clemente, L. Pallotta, A. De Maio, J. J. Soraghan, and A. Farina, “A novel algorithm for radar classification based on Doppler characteristics exploiting orthogonal pseudo-Zernike polynomials,” IEEE Trans. Aerosp. Electron. Syst., vol. 51, no. 1, pp. 417–430, Jan. 2015. [7] L. Wang, T. Huang, and Y. Liu, “Randomized stepped frequency radars exploiting block sparsity of extended targets: A theoretical analysis,” IEEE Trans. Signal Process., vol. 69, pp. 1378–1393, 2021. [8] C. Zhao, L. Wang, and Y. Liu, “Ballistic target recognition based on 4D point cloud using randomized stepped frequency radar,” IEEE Trans. Aerosp. Electron. Syst., vol. 58, no. 6, pp. 5711–5729, Dec. 2022. [9] L. M. Shen and Y. Z. Ma, “An algorithm based on HRRP for target recognition,” in Proc. Int. Conf. Image Anal. Signal Process., Hangzhou, China, 2011, pp. 71–73. [10] X. Hu, F. Lu, C. Liang, J. Liu, and Y. Wang, “Range-Doppler spectrum estimation based on matrix completion for frequency agile radar,” in Proc. IEEE Radar Conf. (RadarConf20), Florence, Italy, 2020, pp. 1–5. [11] B. Fan, X. Du, and W. Hu, “Coherent accumulation of random frequency and pulse repetition interval agile radar,” in Proc. Int. Radar Conf. (RADAR), Rennes, France, 2024, pp. 1–5. [12] C. Zhao, L. Wang, Z. Yue, and Y. Liu, “Enhanced rotor blade length extraction using multicarrier-frequency radar observations,” IEEE Trans. Radar Syst., vol. 3, pp. 1273–1286, 2025. [13] S. Peter and V. V. Reddy, “Extraction and analysis of micro-Doppler signature in FMCW radar,” in Proc. IEEE Radar Conf. (RadarConf21), Atlanta, GA, USA, 2021, pp. 1–6. [14] Y. Diao, S. Liu, X. Gao, A. Liu, and Z. Zhang, “CNN based on multiscale window self-attention mechanism for radar HRRP target recognition,” in Proc. Int. Conf. Signal Image Process. (ICSIP), Suzhou, China, 2022, pp. 281–285. [15] Y. Zhong, L. Shi, Z. Kuang, X. Tu, Y. Huang, and X. Ding, “Radar HRRP signal re-ID via deep learning,” in Proc. CIE Int. Conf. Radar, Haikou, China, 2021, pp. 3202–3205. [16] Q. Lv, H. Fan, Y. Zhao, Y. Quan, and M. Xing, “Agile frequency RCSbased deep fusion network for ship and corner reflector identification,” IEEE Geosci. Remote Sens. Lett., vol. 21, pp. 1–5, 2024. [17] S. Ramraj, N. Uzir, R. Sunil, S. Banerjee, and S. C. Patel, “Experimenting XGBoost algorithm for prediction and classification of different datasets,” Int. J. Control Theory Appl., vol. 9, no. 40, pp. 651–662, 2016. [18] J. J. B. Jayachandran, K. Kirubasankar, and K. Varsini, “Enhancing handwritten character recognition with XGBoost: A machine learning approach,” in Proc. Int. Conf. Data Sci., Agents Artif. Intell. (ICDSAAI), Chennai, India, 2023. [19] W. Huang, R. Tang, G. Qu, and F. Zhang, “An XGBoost-based method for improved orbit prediction with an orbit-separate modeling strategy,” IEEE Trans. Aerosp. Electron. Syst., vol. 60, no. 4, pp. 4887–4895, Aug. 2024. [20] Y. Kong, W. Chen, and H. Leung, “Target type recognition algorithm for SAR image based on multi-feature fusion classifier of KPFD,” in Proc. IEEE Int. Conf. Electron. Inf. Emerg. Commun., Beijing, China, 2015, pp. 435–439. [21] J. Duan, D. Han, and W. Li, “Non-cooperative ship target fusion-based recognition with deep learning,” in Proc. Chin. Control Conf. (CCC), Tianjin, China, 2023, pp. 3439–3443. [22] J. M. Domı́nguez, G. Fourtakas, C. Altomare, et al., “DualSPHysics: From fluid dynamics to multiphysics problems,” Comput. Part. Mech., vol. 9, pp. 867–895, 2022.