Natural disasters, avoidable or unpredictable?

We’ve all seen the images. And they are incredible. Like many natural disasters.

On 25 January, Niscemi, a small Sicilian town, split in two, after the torrential rains associated with Storm Harry, leaving a scar several kilometres long. We saw roads collapsing, cars vanishing, houses left hanging on the edge of a void. More than 1,600 people were evacuated. In the Guardian article, we are reminded that this was not the first time. We are told that the very same area had already slid in the 18th century, and again in 1997, and yet construction continued, especially from the 1950s and 1960s onwards.

Could we have known? Did we already know? And if we did, why does it still happen? Almost inevitably, after every natural disaster, I find that these questions come back too, again and again… I thought I could write a short post to recall that, on the one hand, there are striking regularities in the way some disasters unfold. But on the other hand, “predictable” does not mean “avoidable”, because disaster is also a social fact, a story of vulnerability, exposure, and public choices (at the risk of repeating myself, I know).

Niscemi, or a disaster that had a past

The piece by Lorenzo Tondo and the photographs by Alessio Mamo in the Guardian tell the story of a town, Niscemi, perched on an unstable edge, with its red and black zones, its escorted visits to retrieve a few belongings. And the fear that the next section could give way at any moment. But the article takes a wider view, telling a story that clearly goes beyond Niscemi. There is the rain, there is the local geology, there are fragile clayey soils, but there is also, and above all, urban planning policies, housing built where it should never have been encouraged

… along riverbeds, on unstable slopes, near cliffs, and in zones exposed to landslides, floods and seismic risk. In short, places where people should never have been encouraged to live.

And that is the very root of the problem. We can debate endlessly the violence of the meteorological event, but disaster appears when that violence meets a town, people, buildings, communities, vulnerability. Phil O’Keefe, Ken Westgate and Ben Wisner already wrote, in 1976, in Taking the naturalness out of natural disasters, that a disaster marks the encounter, the interface, between an extreme physical phenomenon and a vulnerable population.

Without people there is no disaster

Without people, there is no disaster. In other words, even if the hazard were identical, the disaster depends first and foremost on exposure and vulnerability. That is why the “predictability” of disasters becomes complex, or ambiguous. We can be very good at forecasting the hazard, and very bad once we have to take into account building rules that tolerate illegality, delayed stabilisation work, or the absence of an operational warning system. And I am writing about this today because I am regularly in meetings where we are told that climate hazard forecasting is getting finer and finer. But it is not so simple to go from climate modelling to detecting events that could be the source of a disaster. For instance, if we want to understand how much rain, and for how long, is enough to trigger a landslide (to keep my guiding thread, the Niscemi disaster).

Rainfall thresholds as a rule of thumb

When we look at rainfall-triggered landslides, we see episodes, often short, when suddenly, the slope moves. The two key factors are a (mean) rainfall intensity, and a duration. Of course, we can refine things, add antecedent soil moisture, soil type, land use, but let us start simply. There are many empirical studies that look for “thresholds” for an (intensity, duration) pair. The goal is not to explain each case, but to draw an operational boundary between rains that often trigger landslides and rains that rarely do. In central and southern Europe, Fausto Guzzetti, Silvia Peruccacci, Mauro Rossi and Colin Stark reviewed a large number of published thresholds (some twenty years ago) in The rainfall intensity–duration control of shallow landslides and debris flows: an update. This allowed them to build a database of 853 rainfall events, some followed by landslides, others not. They also list 124 published empirical thresholds, and propose new thresholds, estimated with a Bayesian approach, to ultimately recover a rule of thumb. It simply says that the longer it rains, the less intense the rain needs to be to trigger a landslide. With a kind of table linking duration and intensity, and making it possible to quantify risk. But, unfortunately, it is hard to transpose. When they compare their thresholds to older ones, they observe that their curves can be lower than some local and regional thresholds, precisely because they aggregate data from different places. They note that there is no universal physical law, only a “statistical boundary”, built from heterogeneous data, sometimes pessimistic, but also sometimes too optimistic. As surprising as it may seem, this kind of tool is particularly appreciated because, even if it is crude, it turns warning into a machine for deciding under uncertainty. It helps choose where to set the cursor between two errors: triggering the alert too often and tiring everyone out, or triggering too late and missing the event. And it also helps transfer a share of responsibility, but that is another story.

From dashboards to the physics of infiltration

A rainfall-triggered landslide is, first and foremost, rain infiltrating, changing pore-water pressures, and reducing shear strength. In engineering terms, rain lowers the “factor of safety”. In Landslide triggering by rain infiltration, Richard Iverson starts from a very concrete observation. Water does not merely “fill” the soil, it also transmits a change in pressure, a signal that can propagate downward faster than one might expect. To organise this idea, he relies on the Richards equation, which governs the flow of variably saturated water in a porous medium under rainfall. He then simplifies it in a reasoned way, by separating time scales. On the one hand, a long time scale, which lets the system install background pressures, linked to topography, geology, and climate, and which influence the propensity to fail without triggering rupture. On the other hand, a shorter time scale, that of storms and rainfall episodes, where the transient transmission of pore-water pressure becomes the true trigger. Between the two, Richard Iverson stresses a useful distinction. We can infiltrate water as a flux, but pressure can also diffuse like a wave, sometimes with very little actual “movement” of water. This reading then makes it possible to combine a simplified hydrological solution with a slope-stability model, to link rainfall intensity and duration to a timeline, a failure depth, and even an acceleration rate. Yes, that is how physical models are built when we try to “predict” disasters. And often, the general idea is simple to state, even if the details are technical. There is a background component, linked to the mean hydrologic state and the water table, which makes a slope more or less ready to fail, and a transient component, linked to rainfall episodes, which actually triggers rupture.

If we want to run this physics on a map, cell by cell, and produce something operational, we quickly end up with tools like the one developed by the USGS in 2008. Rex Baum, William Savage, and Jonathan Godt, in A Fortran Program for Transient Rainfall Infiltration and Grid-Based Regional Slope-Stability Analysis (TRIGRS), model shallow rainfall-induced landslides. The model computes transient changes in pore-water pressure and the associated variations in factor of safety, representing infiltration with analytical solutions of one-dimensional (vertical) differential equations, for saturated or unsaturated conditions. It uses an infinite-slope model to obtain a factor of safety “cell by cell”, and it interfaces with a GIS (Geographic Information System) to prepare the input grids and visualise the outputs. On paper, this is exactly what we want, linking rainfall, topography, and instability with an explicit mechanism. But the TRIGRS manual also makes clear where things get stuck when we try to transpose it to the real world. The model is very sensitive to initial conditions, in particular the initial depth of the water table and the background flow field. If that depth is poorly estimated, the results can become questionable. Forecasting a landslide requires knowing the rainfall, sometimes with quantitative forecasts a few hours ahead. But it also requires knowing the moisture state, the depth of the water table, the hydraulic parameters, and knowing them in a spatially distributed way. In practice, rain gauges are unevenly distributed, radars have biases, soils are heterogeneous, slopes are modified by roads and networks, and urbanisation changes runoff. In other words, we can have a good model, and still be limited by what we can observe. Hence the frequent return, in practice, to empirical thresholds. Not because we give up on physics, but because we are looking for decision robustness.

Predictable does not mean avoidable

Back to Niscemi. In the Guardian article, we read that the area had already slid, that churches might go next, that a library with rare books is inaccessible, and that many residents think this could have been prevented. There is also talk of investigations and responsibility. Even if we accept that the rainfall episode was exceptional, the question “could it have been avoided” points to decades, not hours. The first saves buildings, the second saves lives. To go back to what we said at the beginning, disaster is not the hazard. It is the hazard plus exposure plus vulnerability. Put differently, we have predictability at two levels. At the first level, there is physical predictability. Shallow rainfall-triggered landslides obey mechanisms and regularities. Intensity–duration thresholds raise red flags. Infiltration and stability models produce dynamic susceptibility maps. Often, we can say that an episode is risky, and sometimes where risk is highest. At the second level, there is social predictability. Long before the event, we know where vulnerability is high. Buildings on the edge of an unstable slope, networks laid on fragile terrain, lack of drainage, lack of maintenance, construction in areas known to be problematic. And that knowledge exists even without sophisticated modelling, through institutional memory, archives, mapping, and feedback.

For rainfall-triggered landslides, operational warning systems remain rare and highly localised, to the point that the literature argues they cover only a tiny fraction of the Earth’s land surface. In their paper Rainfall thresholds for the initiation of landslides in central and southern Europe (on central and southern Europe), Fausto Guzzetti, Silvia Peruccacci, Mauro Rossi and Colin Stark note that local or regional thresholds have been defined only in a few places, and that for most of the study area thresholds are not available, due to lack of data (or because data are costly to obtain). This limited coverage is a political fact as much as a technical one, because it depends on budgets, responsibilities, governance, and priorities. The world knows how to build thresholds and models. The world is less good at scaling up warning systems, maintaining them, explaining them, and connecting them to evacuation plans. And then there is a point we talk about less, because it is complicated, not to say uncomfortable. Evacuation is socially costly. We disturb everyone, interrupt the local economy, displace fragile people, and we may trigger anger if the event does not occur. I discussed this in Public authorities faced with risk, from trust to doubt while mentioning the TV series Chernobyl, and continued in “Everything is poison, nothing is poison, it is the dose that makes the poison”, which explained the difficulty and importance of choosing the threshold (yes, this is Statistics 101, or, put differently, the mathematics of decision-making). A threshold is never an oracle. Even when we cross it, nothing may happen where we are looking. A serious warning system (fewer deaths, less damage, more reaction time) must explicitly acknowledge its errors. Type I and type II errors, evacuating a city for nothing, or not evacuating when we should have.

What an honest forecast can promise

What science most often offers is a range, an alert level, a probability map, a ranking of areas. I have the feeling that an honest forecast should mention these four points, one way or another

  1. We do not predict a landslide, we predict a danger regime. Intensity–duration thresholds, like those discussed by Fausto Guzzetti, Silvia Peruccacci, Mauro Rossi and Colin Stark, are useful statistical boundaries, especially when local thresholds are missing.
  2. Physics helps us understand when a threshold makes sense. Richard Iverson shows why characteristic time scales and the transmission of pore-water pressure matter, and why the average background rainfall is not the trigger.
  3. Operational models pay dearly for poor initial conditions. The USGS TRIGRS model is interesting, but it is also very sensitive to the initial water table and background flow. This does not condemn models, it only says that predictability is a chain, and it often breaks at its weakest links.
  4. Finally, disaster is a matter of public policy. Without people there is no disaster, and if exposure increases, we can have more disasters without the Earth becoming “meaner”. Social predictability, the one that says we know this area is fragile, is sometimes the cruelest, because it points to avoidable choices.

And perhaps that is where I should end this post (because one has to end at some point, before the next one). The question “could we have predicted it” is not only a question of scientific accuracy. It is also, as always, a question of responsibility. What do we do with what we already know, even imperfectly? In Niscemi, the Guardian reports a prosecutor’s investigation and a local feeling that this could have been prevented. In that context, talking about rainfall thresholds and infiltration models is not a technical detour. It seems to me it is прежде all a way to clarify what “predictable” really means, and what society chooses to do with it.


OpenEdition suggests that you cite this post as follows:
Arthur Charpentier (February 12, 2026). Natural disasters, avoidable or unpredictable? Freakonometrics. Retrieved March 5, 2026 from https://doi.org/10.58079/15o9k


Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.