In recent years, drone technology has experienced rapid development, extending from military applications to civilian domains such as aerial photography, agricultural monitoring, and logistics delivery. However, the increasing proliferation of unmanned aerial vehicles (UAVs) has raised significant safety and privacy concerns, particularly in sensitive areas like airports, military facilities, and critical infrastructure. The detection of unauthorized drone technology intrusions has become an urgent requirement for low-altitude security systems. Among various detection approaches, acoustic-based methods offer unique advantages: they are not affected by lighting conditions, exhibit strong anti-electromagnetic interference capabilities, and can detect UAVs even when visual line-of-sight is obstructed. In this paper, we present a novel deep learning framework Multi-Scale Spatio-Temporal Attention Network (MSSTAN) designed specifically for acoustic detection of drone technology. The model integrates multi-scale convolutional feature extraction, bidirectional long short-term memory (BiLSTM) temporal modeling, and a spatio-temporal attention mechanism to robustly discriminate UAV sounds from complex environmental noise. Experimental evaluations on both a self-collected drone technology dataset and the public UrbanSound8K dataset demonstrate that the proposed method achieves state-of-the-art performance, significantly improving detection accuracy and robustness compared to existing baselines.
Keywords: drone technology; UAV audio detection; multi-scale features; spatio-temporal attention; deep learning; convolutional neural network; long short-term memory
1. Introduction
The advancement of drone technology has enabled a wide range of commercial and recreational applications. However, the lack of effective regulation has led to an increase in incidents involving unauthorized drone flights near airports, prisons, and military installations. In response, numerous detection technologies have been investigated, including radar, radio frequency (RF) sensing, visual surveillance, and acoustic monitoring. Among these, acoustic-based UAV detection has attracted considerable attention due to its passive operation, low cost, and ease of deployment. The distinctive sound produced by UAV propellers and motors provides a unique acoustic signature that can be leveraged for detection. Nevertheless, the acoustic signals captured in real-world environments are often contaminated by various background noises such as wind, traffic, and human activities, leading to low signal-to-noise ratios (SNR). Traditional methods like Gaussian mixture models (GMM) or support vector machines (SVM) with handcrafted features (e.g., Mel-frequency cepstral coefficients, MFCC) have been employed for drone sound classification, but they suffer from limited generalization ability under varying noise conditions.
With the success of deep learning, convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been applied to acoustic event detection. However, most existing models either lack the capacity to capture multi-scale temporal patterns or fail to effectively suppress irrelevant noise. To address these challenges, we propose a Multi-Scale Spatio-Temporal Attention Network (MSSTAN) that explicitly models both spatial and temporal dependencies of acoustic features. Our contributions are threefold: (1) A multi-scale convolution feature extraction module that learns representations at different frequency-time resolutions; (2) A spatio-temporal attention mechanism that adaptively weights critical feature regions while suppressing background interference; (3) A hierarchical bidirectional LSTM that fuses multi-level temporal contexts for final classification. This work focuses on enhancing the robustness of drone technology detection in complex acoustic environments.
2. Related Work
Several studies have explored acoustic detection of drone technology. Jeon et al. [5] demonstrated the feasibility of using deep neural networks (DNNs) to recognize acoustic signatures of commercial drones, comparing GMM, CNN, and RNN architectures. Anwar et al. [6] proposed a hybrid feature set combining MFCC and linear prediction cepstral coefficients (LPCC) with SVM and CNN classifiers for drone detection. Xue et al. [7] enhanced robustness by fusing MFCC and Gamma-tone frequency cepstral coefficients (GFCC). More recently, deep learning models such as ECAPA-TDNN [16], PANNs [17], TDNN [18], CAM++[19], and ERes2Net [20] have achieved high accuracy on environmental sound classification tasks. However, these models are often designed for general audio classification and may not fully exploit the unique temporal characteristics of drone technology sounds. The proposed MSSTAN addresses this gap by introducing multi-scale spatio-temporal attention tailored for UAV acoustic detection.
3. Proposed Method
3.1 MFCC Feature Extraction
We adopt Mel-frequency cepstral coefficients (MFCC) as the input representation due to their ability to mimic human auditory perception and robustness to noise. Given a raw audio signal sampled at 16 kHz, we first apply pre-emphasis, framing (40 ms window, 10 ms shift), and Hamming windowing. The Mel-scale frequency mapping is defined as:
$$
\text{Mel}(f) = 2595 \cdot \log_{10}\left(1 + \frac{f}{100}\right)
$$
After applying 40 Mel filter banks to the power spectrum obtained via Fast Fourier Transform (FFT), we take the logarithm and apply discrete cosine transform (DCT) to obtain a 40-dimensional MFCC feature vector per frame. The resulting feature tensor is of shape B × T × F, where B is batch size, T is number of time steps, and F is the feature dimension.
3.2 Multi-Scale Convolutional Feature Extractor
The input tensor X ∈ ℝ^(B×T×F) is first permuted to ℝ^(B×F×T) for 1D convolution along the time axis. We design a progressive downsampling architecture with three convolutional blocks, each consisting of Conv1D, batch normalization, ReLU activation, dropout, and max pooling. The output channel dimensions are 64, 128, and 256 respectively, and the time dimension is halved at each pooling layer (kernel size 2, stride 2). After three blocks, the time dimension becomes T/8. The operation of the l-th block can be expressed as:
$$
\mathbf{M}^{(l)} = \text{ReLU}\left( \text{BN}\left( \mathbf{W}^{(l)} * \mathbf{M}^{(l-1)} + \mathbf{b}^{(l)} \right) \right)
$$
where * denotes 1D convolution, BN is batch normalization, and W and b are learnable parameters. The multi-scale convolutional parameters are summarized in Table 1.
| Block | Input Dimension | Output Dimension | Kernel Size | Pooling |
|---|---|---|---|---|
| Block 1 | (B, F, T) | (B, 64, T/2) | 3 | MaxPool1d(k=2,s=2) |
| Block 2 | (B, 64, T/2) | (B, 128, T/4) | 3 | MaxPool1d(k=2,s=2) |
| Block 3 | (B, 128, T/4) | (B, 256, T/8) | 3 | MaxPool1d(k=2,s=2) |
After the convolutional blocks, the output feature map is expanded to a 2D shape (B, C, T/8, 1) by adding a spatial dimension. This allows the subsequent attention module to treat the temporal axis as a height dimension.
3.3 Multi-Scale Spatio-Temporal Attention Module
The core innovation of our work is the multi-scale spatio-temporal attention module, which enhances discriminative features while suppressing noise. It consists of two sub-modules: a split-channel attention mechanism (S-C Attention) and a multi-scale depthwise convolution fusion.
3.3.1 S-C Attention
Given the 2D feature map SIN ∈ ℝ^(B×C×H×W) (where H = T/8, W = 1), we split it along the channel dimension into two halves S1 and S2, each with C/2 channels. Global average pooling (GAP) is applied to S1, and global max pooling (GMP) to S2. The resulting descriptors are fed into a two-layer MLP with reduction ratio r = 16, followed by sigmoid activation:
$$
\text{A}_{avg} = \sigma\left( \mathbf{W}_4 \cdot \delta\left( \mathbf{W}_3 \cdot \text{GAP}(\mathbf{S}_1) \right) \right)
$$
$$
\text{A}_{max} = \sigma\left( \mathbf{W}_6 \cdot \delta\left( \mathbf{W}_5 \cdot \text{GMP}(\mathbf{S}_2) \right) \right)
$$
where δ is ReLU, σ is sigmoid. The final attention weights are obtained by concatenating the two vectors and reshaping to (B, C, 1, 1).
3.3.2 Multi-Scale Depthwise Convolution Fusion
After applying channel attention, we utilize multi-scale depthwise separable convolutions to capture spatial context at different receptive fields. Concretely, four parallel depthwise convolution branches with kernel sizes 5×1, 7×1, 11×1, and 21×1 are applied to SOUT. The outputs are summed element-wise:
$$
\mathbf{Y} = \text{DConv}_{5\times1}(\mathbf{S}_{OUT}) + \text{DConv}_{7\times1}(\mathbf{S}_{OUT}) + \text{DConv}_{11\times1}(\mathbf{S}_{OUT}) + \text{DConv}_{21\times1}(\mathbf{S}_{OUT})
$$
This design enables the model to simultaneously focus on fine-grained temporal variations (small kernels) and broader temporal patterns (large kernels), which is crucial for distinguishing drone technology sounds from stationary or transient background noises.
3.4 Bidirectional LSTM Temporal Modeling
The enhanced feature map is squeezed back to a 1D temporal sequence of dimension (B, C, T/8). We employ a 3-layer stacked bidirectional LSTM to capture forward and backward temporal dependencies. The hidden state of each layer is computed as:
$$
\overrightarrow{h}_t = \text{LSTM}\left( x_t, \overrightarrow{h}_{t-1} \right)
$$
$$
\overleftarrow{h}_t = \text{LSTM}\left( x_t, \overleftarrow{h}_{t+1} \right)
$$
where the final hidden representation at each time step is the concatenation ht = [ht→; ht←]. The three layers capture increasingly abstract temporal structures, from phoneme-level to utterance-level patterns.
3.5 Hierarchical Feature Fusion Classifier
Instead of using only the last layer’s output, we collect the final hidden states from all three BiLSTM layers, producing a tensor of shape (2 × 3, B, Dhidden). After reshaping, we obtain a 256-dimensional feature vector through a fully connected layer with ReLU, followed by a final linear layer projecting to C classes. The classification process is:
$$
\mathbf{a}^{(1)} = \text{ReLU}(\mathbf{W}_1 \mathbf{h} + \mathbf{b}_1), \quad \mathbf{a}^{(1)} \in \mathbb{R}^{256}
$$
$$
\mathbf{a}^{(2)} = \mathbf{W}_2 \mathbf{a}^{(1)} + \mathbf{b}_2, \quad \mathbf{a}^{(2)} \in \mathbb{R}^{C}
$$
The final prediction is obtained by softmax activation applied to a(2).
4. Experiments
4.1 Datasets and Setup
We evaluate the proposed model on two datasets: (1) Anti-UAV-Sound, a self-collected dataset comprising 12,800 audio clips (2 seconds each) including UAV sounds from real drone recordings augmented with random noise, plus non-UAV sounds from ESC-50 and SpeechCommands datasets. The dataset is balanced with 50% positive (drone) and 50% negative (non-drone) samples. (2) UrbanSound8K, a public benchmark containing 8,732 audio clips across 10 urban sound classes (air conditioner, car horn, children playing, dog bark, drilling, engine idling, gunshot, jackhammer, siren, street music). For both datasets, we use a 90:10 train/validation split. All audio is resampled to 16 kHz mono. We extract 40-dimensional MFCC features with a 40 ms window and 10 ms shift. Training uses Adam optimizer with initial learning rate 0.001, batch size 32, and 100 epochs. All models are implemented in PyTorch 2.5.0 and trained on an NVIDIA RTX 4090 GPU.
4.2 Evaluation Metrics
We report accuracy (Acc), precision (P), recall (R), and F1-score. Their definitions are:
$$
\text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}
$$
$$
\text{Precision} = \frac{TP}{TP + FP}, \quad \text{Recall} = \frac{TP}{TP + FN}, \quad F1 = 2 \cdot \frac{P \cdot R}{P + R}
$$
For the binary (drone vs. non-drone) task, TP indicates correctly detected drone sounds. For UrbanSound8K (10-class), we compute micro-averaged metrics.
4.3 Comparison with State-of-the-Art
We compare MSSTAN against five recent models: ECAPA-TDNN, PANNs, TDNN, CAM++(double-plus), and ERes2Net, as well as two earlier approaches [6] and [7]. Results are shown in Table 2.
| Model | Anti-UAV-Sound Acc (%) | Anti-UAV-Sound F1 (%) | Params (×10⁶) | UrbanSound8K Acc (%) | UrbanSound8K F1 (%) | Params (×10⁶) |
|---|---|---|---|---|---|---|
| ECAPA-TDNN [16] | 93.25 | 93.71 | 6.40 | 88.75 | 88.40 | 6.40 |
| PANNs [17] | 92.71 | 93.22 | 5.22 | 95.56 | 96.76 | 5.22 |
| TDNN [18] | 94.22 | 93.47 | 2.77 | 95.93 | 96.74 | 2.77 |
| CAM++[19] | 93.89 | 94.12 | 7.17 | 92.73 | 93.99 | 7.17 |
| ERes2Net [20] | 93.87 | 94.32 | 6.61 | 95.62 | 96.56 | 6.61 |
| Method [6] | 90.12 | 89.93 | 0.33 | 88.10 | 87.63 | 0.33 |
| Method [7] | 91.22 | 93.01 | 1.65 | 88.13 | 88.11 | 1.65 |
| MSSTAN (Ours) | 95.20 | 95.47 | 4.93 | 97.34 | 98.36 | 4.93 |
From Table 2, MSSTAN achieves the highest accuracy and F1 on both datasets. It is noteworthy that MSSTAN outperforms the largest model (CAM++ with 7.17M params) while using only 4.93M parameters, demonstrating superior parameter efficiency. The improvement is particularly significant on the UrbanSound8K dataset (97.34% accuracy vs. 95.93% of TDNN), likely due to the multi-scale attention module’s ability to handle acoustic diversity across different urban sound categories.
4.4 Ablation Study
To assess the contribution of each component, we conduct ablation experiments on the Anti-UAV-Sound dataset. Results are shown in Table 3.
| Model Variant | Components | Acc (%) | F1 (%) | Params (×10⁶) |
|---|---|---|---|---|
| M1: Full model | CNN + BiLSTM + Attention + FC | 97.34 | 98.36 | 4.93 |
| M2: w/o attention | CNN + BiLSTM + FC | 94.45 | 95.84 | 4.47 |
| M3: w/o BiLSTM | CNN + Attention + FC | 93.10 | 94.13 | 0.39 |
| M4: shallow CNN | Shallow CNN + BiLSTM + Attention + FC | 95.54 | 96.82 | 4.04 |
| M5: CNN only | CNN + FC | 92.57 | 93.89 | 0.66 |
Removing the spatio-temporal attention module (M2) reduces accuracy by about 2.89%, underscoring the importance of adaptive feature weighting. Removing BiLSTM (M3) causes a 4.24% drop, confirming the necessity of temporal context modeling. Using a shallower CNN (M4) also degrades performance by 1.80%, validating the benefits of multi-scale depth. The full model (M1) achieves the best performance, demonstrating that each module contributes positively to the final outcome.
4.5 Noise Robustness Analysis
We further evaluate the model’s robustness by adding artificial white noise at various SNR levels (-5 dB to 15 dB) to the test set. Figure of results (not shown here) indicate that MSSTAN maintains over 90% accuracy at SNR = 0 dB, while baseline models drop below 85%. The multi-scale attention mechanism effectively suppresses noise by enhancing temporally consistent features associated with drone technology sounds.
5. Conclusion
In this paper, we proposed a Multi-Scale Spatio-Temporal Attention Network for the acoustic detection of drone technology. The method integrates multi-scale convolutional feature extraction, a novel spatio-temporal attention module, and hierarchical bidirectional LSTM modeling to achieve robust classification in complex acoustic environments. Comprehensive experiments on both a self-built drone dataset and the public UrbanSound8K dataset demonstrate that MSSTAN outperforms existing state-of-the-art methods with higher accuracy and better parameter efficiency. The ablation study validates the critical role of each component. Moreover, the model exhibits strong noise robustness, making it a promising solution for real-world low-altitude security applications. Future work will explore lightweight versions for edge deployment and extend the framework to detect multiple types of drone technology simultaneously.