A machine-learning algorithm originally built to spot impact craters on Mars has been retrained on ocean-floor data, and the results have roughly doubled the known count of submarine calderas, the collapsed volcanic structures that sit hidden beneath miles of seawater. The work, authored by Christopher Lee and James Hogan, adapts a crater detection algorithm that proved competitive with expert-human performance on planetary surfaces and applies it to global bathymetric maps of Earth’s seafloor. The discovery lands at a moment when the scientific community is still reckoning with the destructive reach of the 2022 Hunga Tonga eruption, which sent volcaniclastic flows across vast stretches of ocean floor and damaged undersea infrastructure.
Why a flood of new caldera detections changes the risk picture
For decades, seafloor mapping relied on sparse ship-track sonar passes that left enormous gaps between survey lines. That approach was good enough to catalog the largest volcanic features but systematically missed smaller calderas, particularly along slow-spreading mid-ocean ridges where tectonic stretching produces narrow, irregular depressions rather than the broad shields found at hotspot volcanoes. The semi-automated framework described in a recent environmental study changes the equation by scanning compiled global bathymetry at a scale and speed no human team could match.
The practical consequence is straightforward: if the global inventory of submarine calderas was missing roughly half its members, then hazard assessments built on that inventory were incomplete. Submarine eruptions can generate tsunamis, disrupt undersea cables, and release massive volumes of volcanic sediment. The 2022 Hunga Tonga event demonstrated all three risks simultaneously, with long-runout density currents traveling far beyond the eruption site and severing communication cables. Every newly detected caldera represents a site where similar activity could occur, even if the probability at any single location remains low.
A testable hypothesis follows from the detection pattern. If higher-resolution regional bathymetry surveys confirm that the new caldera detections cluster along slow-spreading ridges, the algorithm is likely surfacing a distinct population of smaller, tectonically controlled features that earlier compilations missed entirely. That would mean the apparent doubling is not an artifact of looser detection thresholds but a genuine expansion of the known volcanic inventory into a size class that ship-track data could not resolve.
This shift matters for risk models that inform coastal planning and subsea infrastructure design. Cable operators, for instance, have traditionally routed lines to avoid known seamount chains and active arcs; a denser field of collapse structures may warrant revisiting those routes. Likewise, tsunami hazard assessments that rely on historical catalogs of submarine eruptions may underestimate the number of potential source regions, especially in basins where slow-spreading ridges dominate the tectonic fabric.
From Martian craters to ocean-floor calderas
The algorithmic lineage traces back to Christopher Lee’s deep-learning crater detection system trained on the HRSC/MOLA blended digital elevation model of Mars. That earlier work, published in Icarus, combined convolutional neural networks with geometric circle-matching to identify impact craters across the Martian surface. By automating what had long been a painstaking manual task, the system showed that machine vision could match or exceed human mappers while processing global datasets in hours rather than months.
Lee and Hogan subsequently generalized this approach into a more flexible crater-finding pipeline, described in a later preprint that emphasized scalability and robustness across planetary bodies. That work framed crater detection as a two-step problem: first, a neural network highlights candidate depressions; then, geometric filters refine those candidates based on expected rim and floor morphology. The key insight was that topographic patterns, not just image textures, could anchor reliable detections over vast, heterogeneous terrains.
Transferring the algorithm from Mars to Earth’s ocean floor required retraining on a fundamentally different type of depression. Impact craters are roughly circular and formed by collisions; calderas are formed by volcanic collapse and tend to be more irregular. Yet both share a common geometric signature: a roughly enclosed depression with raised rims. That structural overlap allowed the core detection architecture to carry over, even though the bathymetric input data differ from planetary elevation models in resolution, noise characteristics, and coverage.
To adapt the system, Lee and Hogan assembled training sets from previously mapped submarine calderas and surrounding non-volcanic terrain. The neural network learned to distinguish collapse structures from other concave features such as fault-bounded basins and erosional scars. While the specific code changes and hyperparameters are not fully detailed in the available literature, the conceptual shift is clear: instead of searching for nearly perfect circles, the algorithm now tolerates elongated and segmented rims, as long as the overall pattern matches a caldera-like depression.
The caldera study sits within a broader wave of machine-learning applications to seafloor mapping. A separate effort cataloged tens of thousands of previously unknown seamounts using global bathymetric compilations, expanding the total seamount inventory and revealing how much of the ocean floor remains poorly characterized. Seamounts are underwater mountains, many of volcanic origin, and their sheer number suggests that the caldera count was similarly underestimated. Together, these discoveries paint a picture of an ocean floor far more volcanically active, at least historically, than existing catalogs reflected.
Gaps in the data and what to watch next
Several questions remain open. The primary study does not publish exact pre-study and post-study caldera totals that would let independent researchers verify the doubling claim with precision. Without those raw numbers, the magnitude of the increase rests partly on the authors’ characterization rather than a transparent before-and-after ledger. Releasing the full detection catalog, including confidence scores and geographic coordinates for each new caldera, would allow the oceanographic community to cross-check the results against regional surveys already in hand.
The algorithm’s false-positive rate also needs scrutiny. Bathymetric data contain many circular or semi-circular depressions that are not volcanic: sediment-filled pockmarks, iceberg scours, and tectonic grabens can all mimic caldera geometry at coarse resolution. If even a modest fraction of the new detections turn out to be non-volcanic features, the effective doubling shrinks. Ground-truthing with multibeam sonar or remotely operated vehicles will be essential to confirm which candidates are true collapse structures and which are look-alikes produced by other processes.
Data quality is another limiting factor. Global bathymetric grids blend satellite altimetry with sparse ship-based soundings, producing maps that smooth over fine-scale relief. In regions where only satellite-derived depths are available, the algorithm is working with blurred topography that may hide small calderas or distort their shapes. Follow-up surveys with higher-resolution multibeam systems could both validate individual detections and reveal an even larger population of small collapse features below the current detection threshold.
There is also a temporal blind spot. Bathymetry captures the present-day shape of the seafloor but says little about when a caldera formed or whether the underlying magma system remains active. Some of the newly cataloged structures may be ancient and long extinct; others could sit atop magma reservoirs capable of future eruptions. Integrating the caldera inventory with seismicity patterns, heat-flow measurements, and geochemical data from hydrothermal vents will be crucial for separating dormant hazards from purely historical relics.
Despite these caveats, the broader implication is hard to ignore: machine learning is rapidly reshaping how scientists see the deep ocean. By repurposing tools first honed on Martian landscapes, researchers are extracting new geological insight from datasets that have existed for years but were too vast for manual inspection. As more algorithms are trained on seafloor imagery, gravity anomalies, and in-situ measurements, the catalog of undersea volcanoes, faults, and landslides is likely to grow even further.
For now, the newly detected calderas serve as a map of questions as much as a map of hazards. Each depression marked by the algorithm is an invitation for closer study-an opportunity to refine models of submarine volcanism, reassess regional risks, and better understand how Earth’s interior reshapes the planet beneath the waves. The next phase will depend less on clever code and more on coordinated fieldwork, as ships and robots trace the algorithm’s digital outlines back to the real seafloor and test how many of these hidden volcanoes are still alive.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
