← Back to Gallery

Interpretable Spectral Representations of Planetary Signals in Vector Spaces (Spec2VEC)

The application of Machine Learning (ML) to planetary science is often bottlenecked by a scarcity of labeled data and the complexity of non-stationary seismo-acoustic signals. While deep learning offers powerful feature extraction, it requires massive datasets that are rarely available for phenomena like Martian quakes or deep-ocean volcanism. Consequently, re- searchers rely on “hand-engineered” features from Time-Frequency Representations (TFRs), but these are often impoverished (simple statistics) and brittle (highly sensitive to subjective TFR hyperparameters). To resolve this, we introduce Spec2VEC, a framework that constructs interpretable spectral representations in vector space. Analogous to generating embeddings in natural language processing, Spec2VEC projects raw, subjective TFRs into a high-dimensional, discriminative feature space. We achieve this by treating the TFR not just as a probability distribution, but as a textured image. Our framework systematically extracts a feature vector using two novel pathways: (1) a Textural path that uses Hilbert space-filling curves to map 2D spectral patterns into 1D sequences for information-theoretic analysis (e.g., Entropy, Complex- ity); and (2) a Spatial path that uses a novel probabilistic segmentation (PDFSI) to quantify the energy concentration, connectivity, and coherence of signal components. We validate this vector space on a comprehensive benchmark of synthetic geophysical signals, demonstrating that it creates a distinct separation between signal classes where traditional metrics fail. Fi- nally, we demonstrate its utility for label-scarce discovery through the unsupervised clustering of hydroacoustic signals from the 2021 Fukutoku-Oka-no-Ba eruption. The results show that Spec2VEC generates a robust, interpretable feature space that enables the automated discovery of complex planetary processes, independent of the underlying TFR choice.

Session: Data-Driven and Computational Characterization of Non-Earthquake Seismoacoustic Sources [Poster]

Type: Poster

Room: Exhibit Hall A+B

Date: 4/16/2026

Presentation Time: 08:00 AM (local time)

Presenting Author: Tolulope Olugboji

Student Presenter: Yes

Invited Presentation: 

Poster Number: 37

Additional Authors