The study, published in Frontiers in Signal Processing by Hao Yin, Yu Lu, Yunfen Chang, Kai Zhang, Yunhui Wu, Fan Yang, Xuexu Li, Xinliang Pang, and Peng Li of the State Key Laboratory of Chemistry for NBC Hazards Protection in Beijing, tackles a deceptively simple question that has long been answered by intuition rather than evidence: which mathematical method should be used to convert a one-dimensional infrasound recording into the two-dimensional image that a neural network actually sees? The researchers systematically compared three classical time-frequency analysis techniques—the Short-Time Fourier Transform (STFT), the Continuous Wavelet Transform (CWT), and the Hilbert–Huang Transform (HHT)—under identical conditions, and the results upend the assumption that higher theoretical resolution always translates into better machine learning performance.

The stakes are considerable. Infrasound monitoring is one of the core technologies of the International Monitoring System established under the Comprehensive Nuclear-Test-Ban Treaty, and it also underpins early warning systems for natural disasters such as volcanic eruptions and debris flows. Yet the signals arriving at monitoring stations are notoriously difficult to disentangle. Natural phenomena like lightning, earthquakes, and volcanic eruptions produce infrasound that can closely resemble, in duration, waveform morphology, and spectral character, the signatures of human-made events such as nuclear tests, large chemical explosions, and rocket launches. As global monitoring networks expand and data volumes surge, manual identification has become untenable, pushing researchers toward automated, deep-learning-based classification.

Previous approaches to this problem have generally followed two paths. One involves fusing multiple hand-crafted features—such as Mel-frequency cepstral coefficients, spectral entropy, or wavelet packet decomposition—and feeding them to classical classifiers like support vector machines, decision trees, or random forests. The other, more recent strategy embraces end-to-end learning, in which convolutional neural networks (CNNs) extract features directly from time-frequency images of the raw signals. Reported accuracies in earlier work ranged from roughly 81 percent for distinguishing earthquakes, tsunamis, and volcanic eruptions to about 96 percent for four-class problems. But a critical gap remained: no study had quantitatively compared how the choice of time-frequency transformation itself affects CNN performance when the input signals are noisy and unfiltered.

To close that gap, the team assembled a standardized dataset of 2,306 infrasound samples drawn from monitoring stations, spanning six categories: nuclear tests (257 samples), chemical explosions (1,084), volcanic eruptions (543), rocket launches (33), earthquakes (137), and lightning (252). Because the recordings originated from different stations with inconsistent sampling rates, the researchers applied shape-preserving piecewise cubic interpolation to unify everything at 100 hertz, followed by normalization. Crucially, they deliberately skipped the denoising step that most prior pipelines relied upon, choosing instead to test how each time-frequency method coped with signals in their raw, contaminated state—a decision that mirrors the realities of operational monitoring, where pristine data is a luxury.

Each method was then carefully tuned. For STFT, the team varied the window function (Hamming versus Hanning), window lengths of 0.32, 1, and 2.56 seconds, and fast Fourier transform sizes of 512, 1,024, and 2,048 points, with a uniform 50 percent overlap between sliding windows. For CWT, they tested two wavelet basis functions—the oscillatory Morlet wavelet and the pulse-sensitive Mexican Hat wavelet—at scale factors of 64 and 128. For HHT, which decomposes signals through empirical mode decomposition (EMD) before applying the Hilbert transform, they evaluated the standard EMD variant alongside the noise-assisted EEMD and CEEMDAN versions. All resulting time-frequency images were fed into an identical multi-scale CNN architecture comprising four convolutional modules with progressively shrinking kernels, batch normalization, max-pooling, Leaky ReLU activations, and three fully connected layers, ensuring that any performance differences could be attributed to the representations themselves rather than to model disparities.

The verdict was decisive. STFT with a Hanning window of 1-second length and an FFT size of 1,024 delivered the best results across every primary metric: 98.35 percent accuracy, 98.44 percent precision, 98.38 percent recall, an F1-score of 98.37, and a Cohen’s Kappa of 0.9751. Across all eighteen STFT parameter combinations, accuracy ranged from 94.86 to 98.35 percent, with the Hanning window generally outperforming the Hamming window and window length proving more influential than FFT size. By contrast, the best CWT configuration—Morlet wavelet at scale 64—reached only 94.02 percent, some 4.33 percentage points behind, while HHT managed 93.59 percent. The efficiency gap was even starker: generating the optimal STFT representation took just 0.85 seconds, whereas CWT required roughly 3.6 times longer on average, and HHT lagged at 65.3 times the STFT benchmark, with the CEEMDAN variant demanding more than two and a half minutes per signal. Models trained on STFT images also converged fastest, stabilizing within about 30 epochs.

The explanation for this counterintuitive outcome lies in how each method handles noise. EMD, the engine of HHT, decomposes signals based on their local extrema, so random noise introduces swarms of spurious extreme points that destabilize the entire decomposition—a phenomenon known as mode mixing, in which a single component scatters across multiple intrinsic mode functions or several frequencies collapse into one. The result is that even two recordings of the same event can yield wildly different time-frequency images depending on the noise they happen to contain. The wavelet transforms suffer from the opposite failure mode: their celebrated high resolution acts as a double-edged sword, faithfully capturing not only genuine signal features but also amplifying random spikes and periodic noise into cluttered, low-contrast backgrounds that confuse the classifier. STFT’s fixed window, meanwhile, sacrifices adaptive resolution but imposes a natural smoothing across the time-frequency plane, suppressing random noise fluctuations into a uniform, low-energy backdrop. Signals of the same category retain consistent overall contours even when contaminated, giving the CNN stable, learnable patterns and a clear gradient descent path.

The practical implications reach well beyond the laboratory. By demonstrating that denoising preprocessing can be skipped without penalty—indeed, that it may introduce information loss and computational overhead—the study simplifies the classification pipeline for real-time infrasound monitoring systems, from nuclear explosion verification to disaster early warning. The authors note that the choice of time-frequency method should match the application: STFT for real-time or large-scale classification, HHT for offline nonlinear analysis, and CWT for low-frequency local feature analysis. They also point toward future refinements, including attention mechanisms that could sharpen the network’s focus on key time-frequency regions and lightweight CNN architectures for further efficiency gains. For now, the message is clear: in the noisy real world, the century-old Fourier transform, applied with the right window, remains a formidable ally to modern deep learning.

Subject of Research: Comparison of time-frequency analysis methods for CNN-based classification of noisy infrasound signals

Article Title: Classification of infrasonic signals based on time-frequency representation

Article References: Yin, H., Lu, Y., Chang, Y., Zhang, K., Wu, Y., Yang, F., Li, X., Pang, X., & Li, P. (2026). Classification of infrasonic signals based on time-frequency representation. Frontiers in Signal Processing, 6, Article 1844194. https://doi.org/10.3389/frsip.2026.1844194

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1844194

Keywords: infrasound, time-frequency analysis, short-time Fourier transform, continuous wavelet transform, Hilbert-Huang transform, convolutional neural network, signal classification, nuclear explosion monitoring, disaster early warning, empirical mode decomposition, deep learning, acoustic signal processing