Reading the Restless Earth
Volcanic eruptions are among the few natural disasters that almost always announce themselves in advance. Long before molten rock reaches the surface, it forces its way upward through the crust, cracking rock, lifting the ground and releasing gases, and each of these processes leaves traces that instruments can detect. The difficulty has never been an absence of warning signs but rather the problem of interpreting them correctly. Over the past half-century, the science of eruption forecasting has advanced considerably further than the public generally recognises, yet it remains far from the precision that people tend to expect of it.
The oldest and still the most widely used monitoring tool is the seismometer. As magma rises, it fractures the surrounding rock, producing swarms of small earthquakes that are usually too weak to be felt but are readily recorded. Volcanologists also listen for a distinctive kind of low-frequency vibration, generated not by breaking rock but by fluids and gases moving through cracks, which often intensifies in the days before an eruption. Unfortunately, earthquake swarms are not a dependable guide on their own. Many volcanoes experience intense seismic unrest that eventually dies away without any eruption at all, a phenomenon researchers refer to, somewhat ruefully, as a 'failed eruption'.
A second line of evidence comes from the shape of the volcano itself. When magma accumulates in a chamber beneath the surface, the ground above swells, sometimes by several centimetres in a matter of weeks. Tiltmeters, which measure minute changes in the angle of the ground, and networks of satellite-linked GPS stations can detect this inflation with remarkable sensitivity. Since the 1990s, these instruments have been joined by satellite radar, which compares images taken on successive orbits to produce maps of ground movement across an entire volcano. Its great advantage is reach: it allows scientists to observe volcanoes in remote regions where no ground equipment has ever been installed. Its limitations are equally real, however, since a satellite may pass overhead only every few days, and dense vegetation or fresh snow can make the images impossible to interpret.
Gases offer a third window into what is happening underground. As magma approaches the surface and the pressure on it falls, the gases dissolved in it escape in a broadly predictable sequence, with carbon dioxide tending to be released earlier than sulphur dioxide. A change in the ratio of one to the other can therefore indicate that magma is ascending. Until recently, measuring these gases meant that scientists had to climb to the edge of an active crater, often wearing gas masks, to collect samples by hand. Small drones fitted with sensors now fly directly through volcanic plumes, gathering data from places that would be far too dangerous for a human to approach.
Yet even with this array of tools, forecasting remains as much an art as a science, because volcanoes behave idiosyncratically. A pattern of earthquakes and swelling that preceded an eruption at one volcano may mean nothing at another, and even the same volcano may behave differently from one century to the next. Reliable interpretation therefore depends on long records of 'normal' behaviour against which unusual activity can be judged. Many of the world's potentially active volcanoes, however, have never been monitored at all, and some have not erupted within recorded history. In my view, this lack of baseline data, rather than any shortcoming in the instruments, is the most serious weakness in the entire field.
Even a perfect forecast would be of little value if it were not acted upon. Most observatories communicate risk through a system of alert levels, ranging from normal activity to imminent eruption. Raising the level can trigger evacuations that close schools, empty hotels and halt local businesses for weeks, and if no eruption follows, residents may become reluctant to leave the next time they are told to. Scientists are acutely aware of this danger, and some have been publicly criticised when their warnings were judged to have been either too cautious or not cautious enough. Such criticism seems to me largely misplaced: scientists can describe probabilities, but the decision to evacuate is a political one and should rest with the civil authorities, who must balance the competing risks.
Some researchers now hope that machine learning will transform the field by detecting subtle patterns in monitoring data that human analysts might overlook. There are good reasons for caution. Such systems learn from examples, and large eruptions are, fortunately, rare events, so the data from which an algorithm could learn are scarce. Where machine learning has already proved its worth is in more mundane tasks, such as classifying the thousands of small earthquakes recorded each day, which frees scientists to concentrate on interpretation. The realistic goal, then, is not a machine that names the day of an eruption, but forecasts expressed in terms of probability, much as weather forecasters describe the chance of rain.
