In recent years, the rapid evolution of low-altitude economy policies and the broad adoption of unmanned aerial vehicles (UAVs) have placed stringent requirements on the reliability of UAV communication links. The term “unmanned aerial vehicle” is no longer limited to military operations; it now encompasses logistics, agriculture, surveillance, and many civil applications. However, with the proliferation of wireless devices, UAV air-to-ground links face severe spectrum congestion and increasingly complex electromagnetic environments. Unmanned aerial vehicle communication systems are vulnerable to intentional or unintentional jamming, which can degrade the quality of the command, telemetry, and sensor data transmission. To guarantee safe flight and reliable data exchange, it is imperative to develop accurate jamming recognition techniques. Jamming recognition provides the basis for subsequent anti-jamming actions such as adaptive frequency hopping, power control, and interference suppression. This study focuses on the jamming recognition problem for UAV air-to-ground links, with emphasis on signal denoising, feature extraction, feature selection, and robust classification under harsh noise conditions.
The rest of this article is organized as follows. Section 2 describes the system model and the types of jamming signals considered. Section 3 presents the proposed multistep singular spectrum analysis (MS-SSA) algorithm and a matrix-extension variant (MEMS-SSA) for reducing computational complexity while preserving filtering performance. Section 4 introduces a divergence Pearson correlation coefficient feature selection (DPCC-FS) algorithm and a channel-space attention convolutional neural network (CSA-CNN) for jamming classification. Section 5 reports experimental results, and Section 6 concludes the paper.
1. System Model and Jamming Signal Characteristics
The UAV air-to-ground communication link considered in this work is based on a direct-sequence spread spectrum (DSSS) system with QPSK modulation. The ground station receives the transmitted signal from the UAV along with jamming signals and ambient noise. The received baseband-equivalent signal can be expressed as
$$
y(t) = h x(t – \tau_h) + k_j j(t – \tau_j) + n(t),
$$
where $x(t)$ is the UAV transmitted communication signal, $h$ is the channel gain of the desired signal, $\tau_h$ is its path delay, $j(t)$ is the jamming signal, $k_j$ is the jamming channel gain, $\tau_j$ is the jamming delay, and $n(t)$ is additive white Gaussian noise. The objective of jamming recognition is to determine which type of jamming is present in $y(t)$ before applying specialized suppression techniques.
In this article, ten common suppression jamming types are considered: continuous wave (CW), amplitude modulation (AM), noise amplitude modulation (NAM), binary phase shift keying (BPSK) jamming, binary frequency shift keying (BFSK) jamming, partial band noise jamming (PBNJ), multi-tone jamming (MTJ), linear frequency modulation (LFM) jamming, noise frequency modulation (NFM) jamming, and periodic pulse noise jamming (PPNJ). Table I summarizes the mathematical models and typical parameter settings.
| Type | Mathematical Model | Parameters |
|---|---|---|
| CW | $J(t) = U_0 \cos(2 \pi f_I t + \theta_I)$ | $f_I \in [2.4,2.5]$ GHz, $\theta_I \sim U[0,2\pi)$ |
| AM | $J(t) = U_0 [1 + \beta m(t)] \cos(2 \pi f_I t + \theta_I)$ | $m(t)$: sinusoidal, $\beta=0.5$ |
| NAM | $J(t) = [U_0 + U_n(t)] \cos(2 \pi f_I t + \theta_I)$ | $U_n(t)$: band-limited Gaussian, variance $\sigma_n^2$ |
| BPSK | $J(t) = U_0 \cos(2 \pi f_I t + \theta_I + \theta_P)$ | $\theta_P \in \{0, \pi\}$ |
| BFSK | $J(t) = U_0 \cos(2 \pi f_{I,k} t + \theta_I)$, $k \in \{1,2\}$ | $f_{I,1}=9$ kHz, $f_{I,2}=11$ kHz |
| PBNJ | $J(t) = U_n(t) \cos(2 \pi f_I t + \theta_I)$ | Bandwidth 63 kHz |
| MTJ | $J(t) = \sum_{i=1}^{N} U_0(i) \cos(2 \pi f_I(i) t + \theta_I(i))$ | $N=6$, frequencies 8–13 kHz |
| LFM | $J(t) = U_0 \cos(2 \pi f_I t + \pi \frac{f_H-f_L}{T_{sweep}} t^2)$ | $f_L=5$ kHz, $f_H=15$ kHz |
| NFM | $J(t) = U_0 \cos(2 \pi f_I t + 2 \pi K_{FM} \int_0^t U_n(\tau) d\tau)$ | $K_{FM}=1$ kHz/V |
| PPNJ | $J(t) = U_n(t) \sum_{k} \mathrm{rect}(\frac{t-kT}{T_p})$ | $T_p/T=0.5$ |
The spectral and constellation properties of these jamming types are notably different. For example, CW and MTJ produce strong discrete spectral lines, while PBNJ and NFM smear energy over a wide band. BPSK and BFSK introduce phase or frequency hopping patterns, whereas PPNJ occupies the channel only during short pulses. However, when the jamming-to-noise ratio (JNR) is very low, the distinction among these types becomes blurred, making the recognition task particularly challenging.
2. Proposed MS-SSA Filtering Algorithm
Singular spectrum analysis (SSA) is a powerful technique for extracting the main components of a time series. Nevertheless, standard SSA constructs a full Hankel trajectory matrix whose adjacent rows are shifted by only one sample. This leads to high computational complexity and large memory consumption when processing long signals, which is problematic for real-time unmanned aerial vehicle applications. To address this issue, we introduce the multistep singular spectrum analysis (MS-SSA) algorithm, which inserts a delay step $S$ into the trajectory matrix construction and reconstructs the sequence using an oblique averaging operation.
2.1 MS-SSA Formulation
Given a discrete-time signal sequence $X = \{X(1), X(2), \dots, X(W)\}$, the MS-SSA algorithm constructs the trajectory matrix as
$$
\mathbf{X} = \begin{bmatrix}
x_1 & x_2 & x_3 & \cdots & x_K \\
x_{S+1} & x_{S+2} & x_{S+3} & \cdots & x_{S+K} \\
x_{2S+1} & x_{2S+2} & x_{2S+3} & \cdots & x_{2S+K} \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
x_{(L-1)S+1} & x_{(L-1)S+2} & x_{(L-1)S+3} & \cdots & x_{(L-1)S+K}
\end{bmatrix},
$$
where $L = \lfloor (W-K)/S \rfloor + 1$ is the number of rows, $K$ is the window length, and $S$ is the delay step. The singular value decomposition (SVD) of $\mathbf{X}$ is given by
$$
\mathbf{X} = \mathbf{U} \boldsymbol{\Sigma} \mathbf{V}^T = \sum_{i=1}^{r} \sqrt{\lambda_i} \, \mathbf{u}_i \mathbf{v}_i^T,
$$
where $\lambda_i$ are the eigenvalues of $\mathbf{X}\mathbf{X}^T$, $\mathbf{u}_i$ and $\mathbf{v}_i$ are the corresponding left and right singular vectors, and $r$ is the rank of $\mathbf{X}$. The dominant singular values represent the main signal components, whereas the smaller ones are associated with noise. By setting a threshold $p$ such that
$$
\frac{\sum_{j=1}^{p} \lambda_j}{\sum_{i=1}^{r} \lambda_i} \geq \eta,
$$
with $\eta=0.85$ in this work, we retain the first $p$ components and discard the rest. The reconstruction step uses an oblique diagonal averaging formula that accounts for the delay $S$:
$$
\hat{x}_t = \begin{cases}
\frac{1}{\lceil t/S \rceil} \sum_{m=1}^{\lceil t/S \rceil} T_{m, t-(m-1)S}, & 1 \leq t < K, \\
\frac{1}{L} \sum_{m=1}^{L} T_{m, t-(m-1)S}, & K \leq t \leq W-K+1, \\
\frac{1}{\lfloor (W-t)/S \rfloor + 1} \sum_{m= \lceil (t-K)/S \rceil +1}^{L} T_{m, t-(m-1)S}, & W-K+1 < t \leq W.
\end{cases}
$$
The final filtered signal is obtained by summing the reconstructed components $\hat{x}^{(i)}_t$ for $i=1,\dots,p$.
2.2 Matrix-Extension MEMS-SSA
When the delay step $S$ is small, the reduction of the trajectory matrix dimension is moderate, and the oblique averaging overhead may outweigh the savings. Therefore, we propose a matrix-extension variant (MEMS-SSA) that pads extra rows with zeros to maintain the same dimension as the standard Hankel matrix. The extended trajectory matrix $\tilde{\mathbf{X}}$ is built by inserting $S-1$ zero rows between each pair of original rows. This approach enables the use of a conventional anti-diagonal averaging procedure, thereby reducing the time complexity of the reconstruction step for small $S$. The resulting filtering performance is identical to that of MS-SSA, but the computational time is lower when $S$ is small.
2.3 Computational Complexity
The computational complexity of SVD is roughly $O(K^2 L)$ for an $L \times K$ trajectory matrix. For standard SSA, $L \approx W-K+1$, and if $K \approx W/2$, then $L \approx K$, giving $O(K^3)$. For MS-SSA, $L \approx (W-K)/S+1$, so the complexity becomes $O(K^2 L) \approx O(K^3 / S)$. Therefore, the speedup is roughly proportional to $S$. The reconstruction cost is also reduced because the number of rows in the trajectory matrix is smaller.
We evaluated the time and memory usage of SSA, MS-SSA, and MEMS-SSA for different sequence lengths (2500, 3000, 3500) and window lengths (500, 1000, 1500). Figure 1 (not shown here) shows that MS-SSA and MEMS-SSA reduce execution time by more than 98% and memory consumption by more than 99% compared with standard SSA. Furthermore, MEMS-SSA achieves lower execution time than MS-SSA for $S=10$ and $S=20$, at the cost of a slight increase in memory usage.
| Algorithm | Execution Time (s) | Memory (MiB) | RMSE |
|---|---|---|---|
| Original signal | – | – | 0.1008 |
| SSA | 89.3160 | 22408.7 | 0.1333 |
| MS-SSA ($S=100$) | 1.1492 | 129.4 | 0.0442 |
| MEMS-SSA ($S=100$) | 1.3832 | 160.5 | 0.0442 |
Table II shows the performance at JNR=0 dB. The root mean square error (RMSE) after MS-SSA filtering is reduced by 56.15% compared with the original noisy signal, whereas standard SSA actually increases the RMSE because it tends to reconstruct the noise as part of the meaningful components. This result indicates that the proposed MS-SSA filters the interference sequence more effectively and preserves the essential features of the jamming signal.
3. UAV Jamming Signal Data Processing
The overall processing chain for jamming data before classification involves the following steps: data normalization, MS-SSA denoising, sequence segmentation, and feature extraction. First, the received signal is normalized to have zero mean and unit variance:
$$
\hat{x}(n) = \frac{x(n) – \mu_x}{\sigma_x}, \quad X(n) = \frac{\hat{x}(n) – \mu_{\hat{x}}}{\sigma_{\hat{x}}},
$$
where $\mu_x$ and $\sigma_x$ are the mean and standard deviation of the raw signal. The normalized signal is then divided into sub-sequences of length 3000 with a sliding distance of 100 samples. Each sub-sequence is filtered by MS-SSA with window length $K=1000$ and delay step $S=100$. The filtered sub-sequences are concatenated into a matrix $\mathbf{X}_{\text{seq}}$ of size $m \times W$, where $m$ is the number of sub-sequences.
From each filtered sequence, we extract 15 statistical and spectral features, including time-domain envelope fluctuation $R$, third-order moment skewness $a_3$, fourth-order moment kurtosis $a_4$, frequency-domain kurtosis $b_4$, frequency peak-to-average ratio $\text{PAR}_F$, high-order peak energy concentration $C_2$, average spectral flatness $F_{se}$, carrier factor $C_\lambda$, power spectral variation coefficient $P_c$, power spectral entropy $H_s$, fractional-order energy accumulation $R_{ft}$, and correlation coefficient $C_r$. Some of these features are defined as follows:
$$
R = \frac{\sigma^2}{\mu^2}, \quad a_3 = \frac{E[(x(n)-\mu)^3]}{\sigma^3}, \quad a_4 = \frac{E[(x(n)-\mu)^4]}{\sigma^4},
$$
$$
b_4 = \frac{E[(X(k)-\mu_f)^4]}{\sigma_f^4}, \quad \text{PAR}_F = \frac{\max |X(k)|}{\frac{1}{N}\sum_{k=1}^N |X(k)|},
$$
$$
C_2 = \frac{\sum_{i=m-k}^{m+k} |Y(i)|^2}{\sum_{i=1}^{N} |Y(i)|^2}, \quad Y(n) = \mathrm{FFT}[x^2(n)].
$$
These features are chosen because they reflect different aspects of the jamming spectrum. For example, CW and MTJ have large $C_2$ because their energy is concentrated in discrete tones, while PBNJ and NFM have high spectral entropy because their energy is spread across the band.
4. Feature Selection Using DPCC-FS
Although the extracted multi-domain features provide rich information, they may still contain redundant and irrelevant components, which can impair the classification accuracy and increase computational burden. To overcome this issue, we propose a filter-based feature selection algorithm that combines the divergence (scatter) criterion and the Pearson correlation coefficient. The class separability criterion based on within-class and between-class scatter matrices is defined as
$$
J_1 = \frac{\mathrm{tr}(\mathbf{S}_b)}{\mathrm{tr}(\mathbf{S}_w)},
$$
where $\mathbf{S}_w$ is the within-class scatter matrix and $\mathbf{S}_b$ is the between-class scatter matrix. A larger $J_1$ indicates that the feature has a better ability to separate different classes. The Pearson correlation coefficient between two features $f_i$ and $f_j$ is given by
$$
J_2 = \left| \mathrm{Corr}(f_i, f_j) \right| = \left| \frac{\mathrm{cov}(f_i, f_j)}{\sqrt{\mathrm{var}(f_i)\mathrm{var}(f_j)}} \right|.
$$
The proposed divergence Pearson correlation coefficient (DPCC) evaluation function is defined as
$$
\mathrm{DPCC}(f_i, f_j) = \alpha J_1(f_i) + \beta \left(1 – J_2(f_i, f_j)\right),
$$
with $\alpha=\beta=0.5$. The algorithm starts with an empty selected set $\mathcal{S}$ and iteratively selects the feature that maximizes the DPCC score while minimizing the redundancy with the already selected features. Table III shows the pseudocode of DPCC-FS.
| Step | Description |
|---|---|
| 1 | Initialize $\mathcal{S} = \emptyset$, $\mathcal{T} = \mathcal{F}$, $\mathcal{H} = \emptyset$. |
| 2 | Compute $J_1(f)$ for every $f \in \mathcal{T}$. |
| 3 | Move the feature with the largest $J_1$ into $\mathcal{S}$. |
| 4 | While $\mathcal{T} \neq \emptyset$ and $\max \mathrm{DPCC} > 0$: For each $f_i \in \mathcal{T}$, for each $f_j \in \mathcal{S}$, compute $\mathrm{DPCC}(f_i, f_j)$. Find the feature $f_m$ that maximizes $\mathrm{DPCC}$. If $\max \mathrm{DPCC} > 0$, move $f_m$ to $\mathcal{S}$. |
| 5 | Return $\mathcal{S}$ as the selected feature subset. |
In our experiments, DPCC-FS reduced the number of features from 15 to 8 while preserving or even improving classification accuracy. This reduction not only lowers the input dimension for the neural network but also enhances the discriminative information of the remaining features by removing redundant parameters.
5. CSA-CNN for Jamming Recognition
5.1 Channel-Space Attention Mechanism
Convolutional neural networks (CNNs) are widely used for classifying jamming signals because they can automatically learn discriminative representations from raw feature vectors. However, a standard CNN treats all channels and spatial locations equally, which may not be optimal when some features are more relevant than others. To overcome this limitation, we introduce a channel-space attention (CSA) module that adaptively recalibrates the feature maps. The channel attention is computed by aggregating spatial information using both average and max pooling:
$$
\mathbf{M}_C = \sigma\left( f^{3\times3}\left( \mathrm{avgpool}(\mathbf{F}) \oplus \mathrm{maxpool}(\mathbf{F}) \right) \right),
$$
where $\sigma$ is the sigmoid function, $f^{3\times3}$ denotes a convolution operation, and $\oplus$ represents concatenation. The spatial attention is computed as
$$
\mathbf{M}_S = \sigma\left( f^{3\times3}\left( \mathrm{avgpool}_c(\mathbf{F}) \oplus \mathrm{maxpool}_c(\mathbf{F}) \right) \right).
$$
The CSA module first applies spatial attention to enhance important spatial positions, and then applies channel attention to emphasize important channels. The complete operation is
$$
\mathbf{F}_S = \mathbf{M}_S \otimes \mathbf{F}, \quad \mathbf{F}_C = \mathbf{M}_C \otimes \mathbf{F}_S,
$$
where $\otimes$ denotes element-wise multiplication.
5.2 CSA-CNN Architecture
The proposed CSA-CNN model consists of multiple encoding-decoding stages. Each encoding stage includes a convolutional block (convolution + batch normalization + ReLU) followed by a CSA module. The decoding stage uses residual connections from the original features to prevent degradation. The model architecture is summarized in Table IV.
| Layer | Output Shape | Description |
|---|---|---|
| Input | (batch, 8) | Selected features after DPCC-FS |
| Conv1 | (batch, 16, 16) | 1D convolution, kernel 3, padding 1 |
| CSA1 | (batch, 16, 16) | Channel-spatial attention |
| Conv2 | (batch, 32, 8) | 1D convolution, stride 2 |
| CSA2 | (batch, 32, 8) | Channel-spatial attention |
| Decoder Residual | (batch, 32, 8) | Add original processed features |
| Conv3 | (batch, 64, 4) | 1D convolution, stride 2 |
| CSA3 | (batch, 64, 4) | Channel-spatial attention |
| Flatten | (batch, 256) | Flatten and linear layer |
| Softmax | (batch, 10) | Probability over 10 jamming types |
The residual connection feeds the original feature representation directly to the decoder, which helps mitigate the vanishing gradient problem and prevents information loss. The model is trained with the cross-entropy loss using the Adam optimizer with a learning rate of $10^{-3}$ and a batch size of 1000. Early stopping with a patience of 50 epochs is applied to avoid overfitting.
6. Experimental Results
6.1 Dataset and Experimental Setup
The experiments were conducted using a simulated DSSS QPSK communication link in MATLAB Simulink, with parameters matching realistic UAV communication systems. The carrier frequency was 2.4 GHz, the chip rate was 1 kHz, and the sampling frequency was 100 kHz. Ten types of jamming signals were generated with the parameters listed in Table I. The JNR was varied from -20 dB to 20 dB in steps of 2 dB. For each JNR and each jamming type, 2000 samples were collected. Half of the samples were used for training and the remaining half for testing, yielding a test set of 1000 samples per class per JNR. Additionally, a real-data collection experiment was carried out using USRP B210 and ANTSDR E310 software-defined radios. The jamming power was varied from 0.01 W to 0.2 W, and the received signals were processed by the same chain.
6.2 Hyperparameter Sensitivity
We first analyzed the influence of the number of encoder-decoder layers and the batch size on the classification accuracy of CSA-CNN. Table V reports the average accuracy over five independent runs. It can be observed that two layers give the best trade-off between model capacity and generalization. Deeper models tend to overfit, especially at low JNR. Batch size has a mild effect, but 1000 provides stable results and faster computation.
| Layers | Batch | JNR=-20 dB | JNR=-10 dB | JNR=0 dB | JNR=10 dB | JNR=20 dB |
|---|---|---|---|---|---|---|
| 1 | 500 | 43.8 | 78.1 | 92.8 | 99.9 | 100.0 |
| 1 | 1000 | 42.8 | 78.4 | 94.4 | 99.9 | 100.0 |
| 2 | 1000 | 46.8 | 80.2 | 96.3 | 100.0 | 100.0 |
| 3 | 1000 | 44.8 | 77.1 | 90.7 | 99.9 | 100.0 |
| 4 | 1000 | 43.8 | 77.9 | 89.2 | 98.7 | 100.0 |
Based on these results, we set the number of layers to 2 and the batch size to 1000 for all subsequent experiments.
6.3 Effect of MS-SSA Filtering
To evaluate the impact of the proposed filtering, we compared the CSA-CNN performance on the raw dataset and on the MS-SSA filtered dataset. Figure 2 shows the accuracy curves. At low JNR (below -10 dB), filtering improves the accuracy by more than 5 percentage points. At medium JNR (between -10 dB and 0 dB), the improvement is about 3 percentage points. This demonstrates that MS-SSA effectively extracts the underlying jamming components from the noisy received signal, thereby facilitating the subsequent classifier.
We also compared MS-SSA with standard SSA, windowed SSA (WSSA), overlap SSA (ov-SSA), wavelet denoising, and Kalman filtering. The average RMSE over JNR from -20 dB to 20 dB is listed in Table VI. MS-SSA achieves the smallest RMSE, confirming its superiority in denoising jamming signals.
| Algorithm | Unfiltered | Wavelet | Kalman | WSSA | ov-SSA | MS-SSA |
|---|---|---|---|---|---|---|
| Average RMSE | 0.2315 | 0.1991 | 0.1418 | 0.0736 | 0.0690 | 0.0880 |
Although WSSA and ov-SSA also reduce the RMSE, MS-SSA lowers the execution time by more than 95% compared with WSSA and by more than 98% compared with ov-SSA while achieving a comparable or even lower RMSE. This makes MS-SSA particularly suitable for real-time unmanned aerial vehicle communications.
6.4 Classification Performance
Table VII lists the classification accuracy of CSA-CNN compared with a standard CNN and a multi-branch CNN (MB-CNN) when using the filtered dataset. CSA-CNN consistently outperforms both baselines. In particular, at JNR=8 dB, CSA-CNN reaches 100% accuracy for all ten jamming types. When JNR is low, CSA-CNN still maintains a noticeable advantage. For example, at JNR=-20 dB, CSA-CNN achieves 46.8% accuracy, whereas CNN and MB-CNN achieve 39.5% and 42.9%, respectively.
| JNR (dB) | CNN | MB-CNN | CSA-CNN |
|---|---|---|---|
| -20 | 39.5 | 42.9 | 46.8 |
| -10 | 74.6 | 75.8 | 80.2 |
| 0 | 89.6 | 91.6 | 96.3 |
| 8 | 94.3 | 95.8 | 100.0 |
| 20 | 100.0 | 100.0 | 100.0 |
Table VIII shows the impact of DPCC-FS feature selection. After applying DPCC-FS, the CSA-CNN accuracy slightly improves at low JNR and remains identical at high JNR. For CNN and MB-CNN, feature selection also yields better performance at most JNR values. This confirms that removing redundant features reduces overfitting and helps the network generalize better.
| JNR (dB) | CNN (No-FS) | CNN (DPCC-FS) | MB-CNN (No-FS) | MB-CNN (DPCC-FS) | CSA-CNN (No-FS) | CSA-CNN (DPCC-FS) |
|---|---|---|---|---|---|---|
| -20 | 39.5 | 43.3 | 42.9 | 44.6 | 46.8 | 46.8 |
| -10 | 74.6 | 77.1 | 75.8 | 81.0 | 80.2 | 82.7 |
| 0 | 89.6 | 88.8 | 91.6 | 92.9 | 96.3 | 96.9 |
| 10 | 96.4 | 94.4 | 95.0 | 94.8 | 100.0 | 100.0 |
The confusion matrices at different JNR values show that the main confusions occur between CW and NAM, and between PBNJ and PPNJ, especially at low JNR. This is expected because CW and NAM both occupy narrow bands, and PBNJ and PPNJ both have random amplitude fluctuations. However, as JNR increases, these confusions reduce rapidly, and at JNR=8 dB the confusion matrix becomes diagonal.
6.5 Real-Data Validation
To further validate the proposed method, we performed experiments with real UAV communication signals collected using software-defined radios. The jamming power was set to 0.01 W, 0.05 W, 0.1 W, 0.15 W, and 0.2 W under a fixed noise floor. Table IX presents the classification accuracy of the three models. CSA-CNN achieves more than 96% accuracy when the jamming power is 0.15 W and more than 99% at 0.2 W. These results confirm the practical utility of the proposed approach for unmanned aerial vehicle systems.
| Jamming Power (W) | CNN | MB-CNN | CSA-CNN |
|---|---|---|---|
| 0.01 | 41.7 | 46.9 | 50.1 |
| 0.05 | 72.9 | 77.8 | 83.5 |
| 0.10 | 80.4 | 82.5 | 86.9 |
| 0.15 | 92.3 | 88.2 | 96.9 |
| 0.20 | 94.1 | 92.8 | 99.3 |
7. Conclusion
In this article, we addressed the jamming recognition problem for unmanned aerial vehicle air-to-ground communication links. We proposed a multistep singular spectrum analysis (MS-SSA) algorithm and its matrix-extension variant MEMS-SSA to reduce the computational cost of standard SSA while preserving excellent denoising capability. We also introduced a divergence Pearson correlation coefficient feature selection (DPCC-FS) method to eliminate redundant features and improve class separability. Finally, we designed a channel-space attention convolutional neural network (CSA-CNN) that adaptively emphasizes important features in both channel and spatial dimensions. Extensive experiments on simulated and real datasets demonstrated that the proposed CSA-CNN model consistently outperforms conventional CNN and multi-branch CNN in terms of classification accuracy across a wide range of JNR values. The integration of MS-SSA filtering and DPCC-FS selection further enhances the robustness and efficiency of the entire recognition system. Future work will extend the method to cover smart jamming types and composite jamming scenarios, and we will investigate the fusion of the feature selection and recognition modules into a single end-to-end network to further reduce latency and computational burden for real-time unmanned aerial vehicle applications.