A volcano that has erupted in the past is likely, on a long enough timeline, to erupt again, and the populations living on its flanks have a clear interest in knowing when. The scientific discipline of volcanic monitoring, which tries to convert that interest into concrete warnings, is only about a century old in its modern form. It combines geology, geochemistry, and seismology with a practical sense of how to communicate uncertainty to people whose lives may depend on the answer.

The earliest monitoring systems were purely visual. Observers stationed at the summit of Vesuvius from the 1840s onwards recorded the height of the lava column, the colour and density of the plume, and the arrival of ashfall. The notes are surprisingly informative: a modern volcanologist reading them can identify several eruption phases that would have required instruments to detect at the time. But visual observation misses everything that happens underground, and by the early twentieth century it was clear that the most useful warning signs appeared before any surface activity was visible.

The first of these signs is seismicity. Magma rising through the crust fractures the surrounding rock and generates small earthquakes, most of them too weak to be felt. A network of seismographs positioned around a volcano can not only detect these events but triangulate their locations, allowing analysts to track the upward progress of the magma over days or weeks. The Italian volcanologist Andrea Moretti has described how a sequence of thousands of small earthquakes beneath Mount Etna in 2001 clearly marked the path of an ascending dyke before the summit activity began.

Ground deformation provides the second major channel. As magma accumulates at shallow depth, the volcano inflates measurably. Traditional instruments called tiltmeters can detect angular changes of a few microradians in the slope of the ground, and modern satellite-based radar, known as InSAR, can map deformation across an entire mountain with centimetre precision. Moretti's group uses a combination of both, on the argument that tiltmeters offer minute-by-minute temporal resolution while the satellite system offers much richer spatial detail.

Gas emissions carry a third kind of information. Sulphur dioxide, carbon dioxide, and a number of trace species escape from magma at different depths, and the ratios between them change as the magma rises. A sharp increase in the ratio of sulphur dioxide to carbon dioxide is often a sign that magma is now close to the surface and has lost much of the deep-sourced carbon that was present earlier. Instruments on aircraft, drones, or the rim of the crater itself can measure these gases continuously, though operating them near an active vent remains dangerous and occasionally fatal.

The combination of seismic, deformation, and gas data forms what Moretti calls the three-channel picture. No single channel is reliable by itself: a seismic swarm without deformation may simply reflect tectonic stress adjustments, and ground inflation without seismicity is sometimes caused by shallow hydrothermal fluids rather than magma. Only when all three channels move together, and continue to move, is there a strong case for an eruption warning. Even then, the scientific judgement must be combined with the population's willingness to act, which depends heavily on the memory of past events.

Communicating uncertainty has proved to be as difficult as measuring it. In 1984, the authorities at Rabaul in Papua New Guinea ordered a major evacuation based on intense seismicity, only to cancel it when the activity declined; when the volcano eventually erupted ten years later, the inhabitants had lost confidence in the warning system and fled too late. In contrast, the successful evacuation of Mount Pinatubo in the Philippines in 1991, which saved perhaps twenty thousand lives, succeeded partly because a video explaining the likely appearance of a catastrophic eruption was distributed to local leaders in the weeks before the event. The video gave the public a concrete image to match against the unfolding situation, rather than an abstract probability.

Different volcanoes require different watching. The so-called shield volcanoes of Hawaii produce effusive eruptions of fluid lava, dangerous to property but generally not to life, and their monitoring concentrates on predicting the location of new vents. The stratovolcanoes of the Pacific Rim, which include Pinatubo, Mount St Helens, and Mount Fuji, are more explosive and more dangerous, and their monitoring must detect the deep build-up of a large volume of gas-rich magma. Caldera systems, of which Yellowstone is the best-known example, present their own challenge: any single parameter on a caldera may appear alarming when compared with a typical volcano but is routine for a caldera. A ground inflation of thirty centimetres over a decade is catastrophic at Etna and unremarkable at Yellowstone.

The future of volcanic monitoring lies in combining more channels of data with automated pattern recognition. Machine-learning systems trained on the records of past eruptions have shown modest but real improvements in short-term forecasting, particularly in the hours and days before an imminent event. Moretti is cautious about the outcome. The current record of eruptions is short, in geological terms, and the algorithms have not been tested against the range of possible behaviours. What the new tools can do, on his account, is help human forecasters sift through a larger volume of data; they cannot substitute for the experienced judgement that is needed when a population is about to be told to leave their homes.