In the young Universe, about ten billion years ago, galaxies resembled endless fields after a thunderstorm: everywhere, giant — thousands of light-years across — clumps of star formation were boiling up, where young stars were born with unbridled energy. Today, these patches are an almost vanished species: in nearby galaxies, raucous hotbeds are rare. To unravel the reasons for this fading, we need statistics on a truly cosmic scale — and here comes an automated combine harvester onto the scene.

The driver‑mechanic here is a neural network Faster R-CNN, layered on the foundation of Zoobot — a model that absorbed millions of meticulous galaxy classifications from volunteers of the Galaxy Zoo project. This trained “brain” scans images in six filters ugrizy (from ultraviolet to near-infrared), as if peering at each field through a set of sensitive eyepieces. To train the detector, synthetic clumps born from numerical simulations of stellar populations were injected into real frames, and the aim was to flawlessly separate “sprouts” from background galaxies, interloper stars, and digital artifacts.

The result exceeded expectations: for objects brighter than 26th magnitude in the r-band, detection completeness topped 0.9, and purity exceeded 0.8. After combing through 710,000 galaxies out to a redshift of 0.5, the harvester tallied almost one and a half million candidates. Efficiency dips slightly near galactic centers and for low-contrast structures, but remains stable over a wide range of masses and star formation rates of the parent systems. The most intriguing part: multi-channel vision lets the harvester sniff out clumps even where the human eye would see only a fuzzy smudge — thanks to a clever normalization of brightness via the hyperbolic arcsine (see formula).

Now, with a map of the “crop yield” of the nearby cosmos, astronomers can for the first time directly compare it with the distant Universe of the cosmic noon epoch. This is a reliable compass on the path to understanding how galaxies built up mass and why their furious star formation gradually petered out. Next steps include mass fitting of spectral energy distributions to estimate ages, metallicities, and masses; with the arrival of Euclid and LSST surveys — resolving substructures and incorporating spectroscopy to measure gas dynamics. This work sets a gold standard for automated morphological classification, turning the study of stellar evolution and interstellar medium enrichment into a precise statistical science.