ÃÈÃÃÉçÇø

When I first started working with integral field spectroscopic (IFU) data, I was struck by how much complexity was being averaged out or masked by traditional processing techniques. Most segmentation methods in astronomy—especially those designed for IFU data cubes—rely either on predefined morphological components or on signal-to-noise heuristics. Among the most common is Voronoi binning, which prioritizes the signal-to-noise ratio at the expense of preserving the underlying spectral variation.

I began exploring an alternative path: Instead of applying predefined morphological models or relying on signal-to-noise-driven techniques like Voronoi binning—which often obscure spectral diversity—I wanted to see what would happen if we let the spectra themselves define the structure.

This motivation led to Capivara, an unsupervised segmentation algorithm that groups regions of a galaxy based on their full spectral similarity. It doesn't assume what a structure "should" look like—no prior templates, no imposed geometry. Instead, it clusters regions whose spectra are alike, allowing the data to guide the segmentation.

We applied Capivara to galaxies from the MaNGA survey, and without manual intervention or supervised training, the algorithm successfully delineated physically coherent structures—including tidal tails, bars, rings, spiral arms and even foreground stars. These features emerged naturally from spectral correlations, rather than from spatial proximity or brightness thresholds.

Technically, Capivara is built for scalability. It uses spectral clustering implemented with Torch and GPU acceleration, making it efficient for modern IFU datasets that can contain tens or hundreds of thousands of spectra per galaxy. Importantly, Capivara avoids the spectral averaging inherent in many binning approaches—a critical feature when studying subtle processes like metallicity gradients, shock fronts or ionization structure.

Our goal wasn't to replace existing tools, but to provide a complementary framework—one that maintains spectral integrity and supports exploratory, data-driven science. Capivara outputs include cluster centroids, confidence scores, and masks that integrate easily with existing analysis pipelines.

The software is open-source and under an MIT license on GitHub. The full methodology and results are described in our new paper in Monthly Notices of the Royal Astronomical Society.

Ultimately, Capivara reflects a shift in how we approach galaxy analysis. Rather than fitting the data to our expectations, we allow the data to reveal its own structure—and, in doing so, uncover features that might otherwise remain hidden.

This story is part of , where researchers can report findings from their published research articles. for information about Science X Dialog and how to participate.

Rafael S. de Souza is a Senior Lecturer at the University of Hertfordshire, where his research focuses on developing advanced statistical methodologies for galactic and extragalactic astrophysics, cosmology, and nuclear astrophysics. He is the founder and chair of the Cosmostatistics Initiative (COIN)—an international, interdisciplinary collaboration that brings together astrophysicists, statisticians, computer scientists, philosophers, musicians, and biologists to foster novel approaches to scientific data analysis.

His book Bayesian Models for Astrophysical Data (Cambridge University Press, 2017) received the PROSE Award for Best Book in Cosmology and Astronomy. Rafael has also served three consecutive terms as Vice-President of the International Astrostatistics Association.