Predicting Landslides: My Deep Dive into Slope Stability Models

Hey there! Let me tell you about something pretty important that keeps folks who live near steep hills and mountains up at night: shallow landslides. These aren’t your slow-moving, majestic glaciers; nope, these are the fast, furious, and frankly, terrifying ones. They can turn into debris flows in a blink, and when they hit populated areas, they’re seriously destructive. The tricky part? They often happen in clusters over wide areas, making them a real headache to predict.

So, my mission, should I choose to accept it (and I did!), was to figure out which kind of computer model is best suited to predict these nasty events, especially when we’re looking at a whole region, not just a single hillside. Think of it like trying to predict where popcorn will pop next in a whole bag – you need a good strategy!

Why Predicting Shallow Landslides is a Big Deal

Landslides, in general, are becoming more common. Our planet’s surface is literally changing because of them. Shallow landslides, the ones typically just a meter or two deep, are particularly sneaky. They come out of nowhere, often with no warning signs, and they’re heavily influenced by how much water is in the soil. And with climate change messing with weather patterns, bringing more extreme rain, these events are only getting more frequent and severe. Being able to predict where and when they might happen is absolutely critical for keeping people safe and managing risk.

Historically, we’ve had a couple of main ways to try and predict these things:

• Statistical Models: These look at past rainfall and landslide data to find patterns. They’re useful, but they don’t really get into the nitty-gritty of *why* a slope fails. They’re like saying, “When it rains this hard, landslides *often* happen here,” but not explaining the physics behind it. Plus, they need lots of historical data, which isn’t always available everywhere.
• Machine Learning: A newer, fancier version of the statistical approach. These algorithms are great at finding complex patterns in huge datasets. They can be more powerful, but they still rely heavily on data quality and quantity.
• Physically-Based Models (PBMs): Now we’re talking! These models dive deep into the mechanics. They use actual physics – soil strength, water flow, gravity, slope shape – to understand the failure process itself. This gives us a much better handle on how slopes will react to changing conditions.

Within PBMs, there are different flavors. Some use really complex math like Finite Element Method, but for shallow landslides, the old reliable Limit Equilibrium Method (LEM) is still super popular. It’s simpler but effective. LEM basically calculates a “Factor of Safety” (FS) – if the forces holding the slope up are bigger than the forces trying to pull it down, it’s stable (FS > 1). If the pull-down forces win, it’s unstable (FS elt; 1).

Many PBMs for shallow landslides use the LEM combined with the idea of an “infinite slope” – basically assuming the slope is long and uniform, and failure happens along a shallow, flat surface parallel to the ground. Models like SHALSTAB, TRIGRS, and HIRESSS use this approach.

But here’s a crucial point: how these models handle water is key! Some assume water flow is constant (steady-state), which is computationally easy but misses the drama of a heavy rainstorm. Others, like TRIGRS and HIRESSS, use transient hydrological models that track how soil saturation and water pressure change *over time* during rainfall. This dynamic view is crucial because it’s the rapid increase in pore water pressure that often triggers shallow slides.

Given the variety of models and the huge differences in landscapes and how water moves through them, picking the *right* model for a specific area is super important. That’s exactly what this study set out to do.

The Great Model Showdown: HIRESSS vs. SCOOPS3D

We decided to pit two physically-based models against each other: HIRESSS (HIgh REsolution Slope Stability Simulator) and SCOOPS3D. The testing ground? The stunning, but landslide-prone, Valle d’Aosta region in the Italian Alps. This area is perfect (or maybe terrible?) for this study because it has steep slopes, thin soils, and gets a decent amount of rain – basically, prime shallow landslide territory.

We focused on two significant rainfall events that actually caused shallow landslides: one in May 2013 and another in October 2020. The goal was to see if these models could predict the landslides that *did* happen.

Meet HIRESSS: This model is a bit of a powerhouse. It’s designed for large areas and can run simulations in near real-time, even down to hourly updates. It has two main parts: a hydrological module that tracks water pressure changes dynamically using a simplified version of the Richards equation (fancy math for water flow in soil), and a geotechnical module that calculates the Factor of Safety using the infinite slope method. It even considers things like root strength from plants and the uncertainty in soil properties using Monte Carlo simulations (basically, running the calculation many times with slightly different values to see the range of possible outcomes).

Meet SCOOPS3D: This one takes a different approach. It uses a 3D limit equilibrium method, looking at potential failure surfaces as spheres within columns of soil defined by the landscape grid. It can estimate the size and volume of potential slides. However, here’s its main difference from HIRESSS: it *doesn’t* have its own dynamic hydrological model. It needs you to tell it where the water table is.

To make the comparison fair, we did something clever. We ran HIRESSS first for the entire study area (a big chunk called Alert Area A). HIRESSS gave us detailed maps of soil saturation and water table depth over time. Then, we took the water table positions from three key moments in the HIRESSS simulation – *before* the rain (Pre-event), *during* the heaviest rain (Peak-event), and *after* the rain as things drained (Post-event) – and fed *those* static water table maps into SCOOPS3D. This way, both models were assessing stability under the *exact same* water conditions at those specific moments, despite their different internal workings.

Gathering the Ingredients (Data!)

To feed these hungry models, we needed lots of data. We used a high-resolution Digital Elevation Model (DEM) to get slope angles and terrain shape. We also did some serious fieldwork! We went out to 12 different sites representing the main soil types in the area, dug holes, and performed tests right there in the ground (like Borehole Shear Tests to measure soil strength and Amoozemeter tests for how fast water drains). We also collected soil samples to analyze back in the lab (particle size, density, etc.). This allowed us to assign specific geotechnical properties (like cohesion, friction angle, how dense the soil is) to different soil types across the region.

We also needed to account for the helpful work of plant roots, which add extra strength (cohesion) to the soil. We used land cover maps and existing research to estimate this “root cohesion” for different vegetation types.

For the dynamic part, we needed rainfall data. We collected hourly rainfall data from 26 weather stations scattered across the area for the two events and the two weeks leading up to them (antecedent conditions are important!). We then used a technique called Thiessen polygonation to turn those point measurements into rainfall maps covering the whole area for every hour of the simulation.

Putting the Models to Work: What Happened?

HIRESSS ran its simulations for the entire Alert Area A, producing hourly maps of Factor of Safety (FS) and, importantly, the probability of failure (basically, how likely it is for the FS to drop below a critical threshold, which we set slightly conservatively at 1.2 instead of the theoretical 1, to account for uncertainties).

SCOOPS3D, being more computationally intensive for large areas, was run on four smaller sub-basins within Alert Area A where landslides were known to have occurred during the events. We gave it the three different water table maps (Pre-event, Peak-event, Post-event) derived from HIRESSS and asked it to calculate the minimum FS for every point in those sub-basins.

The Results Are In!

First, let’s look at HIRESSS’s dynamic output. We saw a really clear temporal correlation between the peaks in rainfall and the increase in the number of pixels showing a high probability of failure (egt;75%). When the rain poured, the model showed more areas becoming unstable, and when the rain stopped, the probability dropped. This matched up nicely with the actual dates the shallow landslides were reported in the regional database.

We also did a spatial check. For each reported landslide location, we looked at a small “influence zone” around it and counted how many pixels within that zone showed a high failure probability (egt;75%) at the peak of the event. For most landslides, this number increased significantly during the rainfall, suggesting HIRESSS was highlighting the right areas. There were a couple of cases where this didn’t happen, which might be due to inaccuracies in the reported landslide locations or the resolution of our input data (10m might not capture very small landslides perfectly).

Now, let’s compare the FS maps from both models for the three saturation scenarios in the sub-basins. Interestingly, both models tended to identify *similar areas* as generally more prone to instability (lower FS values). So, in terms of relative susceptibility, they weren’t miles apart.

However, the *absolute* FS values were often quite different between the two models. More importantly, when we looked at how the FS values changed with the water table depth across the different scenarios, we saw a big difference in sensitivity.

The Sensitivity Story

As you’d expect, for both models, the shallower the water table (meaning the soil is more saturated), the lower the FS value (less stable). But HIRESSS showed a much more pronounced response. During the Peak-event scenario, when the water table was highest, HIRESSS showed a clear concentration of points with FS values near or below the 1.2 threshold. It was really sensitive to that increase in water, clearly showing the destabilizing effect.

SCOOPS3D, on the other hand, showed a less distinct pattern. The distribution of FS values didn’t change nearly as much between the Pre-event, Peak-event, and Post-event scenarios, even though we gave it the same changing water table data. It seemed less sensitive to the rapid pore pressure changes that trigger shallow landslides in this kind of environment.

Why the difference? It boils down to their fundamental approaches. HIRESSS’s infinite slope model and its dynamic hydrological module are really well-suited for the conditions in Valle d’Aosta: thin, permeable soils on steep slopes where rapid increases in pore water pressure cause failure along shallow, planar surfaces. It effectively captures that critical triggering mechanism.

SCOOPS3D’s 3D limit equilibrium approach, while powerful for other types of landslides or geological settings, didn’t seem to capture the specific dynamics of these rainfall-induced shallow slides as effectively in this context. It seemed to be identifying potential deeper failures, which wasn’t the focus of this study.

So, What’s the Takeaway?

This study really hammered home that choosing the right model isn’t just about picking the fanciest one; it’s about picking the one that best fits the specific landscape and the type of landslide you’re trying to predict. In the case of rainfall-induced shallow landslides in steep, thin-soiled environments like Valle d’Aosta, HIRESSS, with its dynamic hydrology and infinite slope approach, proved to be the more effective tool.

These findings are super valuable for folks involved in risk management and land-use planning. Knowing which areas are most susceptible and how that susceptibility changes during a rain event helps us develop better early warning systems and target mitigation efforts more effectively.

Of course, models are never perfect. The resolution of the data matters (a 10m DEM is good, but higher resolution might capture tiny features better, though it’s a lot more work to get!). There are always uncertainties in soil parameters and how we represent complex natural processes.

Looking ahead, it would be great to incorporate even higher-resolution data, explore the specific impacts of future climate change scenarios on these landslides, and move towards quantitative risk analysis that considers not just *where* a landslide might start, but also *how* it might travel and what it might impact.

Ultimately, it’s about using the best tools we have, tailored to the specific challenge, to help keep communities safe from the unpredictable power of nature.

Source: Springer