How Landslide Hazard Maps Are Made

In short
- A landslide susceptibility map is made by combining spatial factors such as slope, geology and rainfall.
- Statistical models and machine learning models work on different principles but share the same goal.
- The numbers on a hazard map are probabilities, not predictions, so they need careful interpretation.
After the rainy season ends, the news shows a caption: "This area is landslide risk grade 1." Where does that number come from? No one walked the mountain to rate the risk by hand. It is the result of gathering several kinds of spatial data, such as slope angle, geological properties, past rainfall records and land cover, feeding them into a model and calculating.
The process of making a landslide hazard map is called susceptibility modeling. Susceptibility (vulnerability) expresses, as a number, how easily a disaster occurs under given conditions. This post looks step by step at how such a model is built and what to watch out for when reading the map.
What data do you need to make a landslide hazard map?
The first step of susceptibility modeling is choosing the input factors. Input factors are the terrain, geology and weather information known to be related to landslides. The most basic is slope angle. The steeper the slope, the more easily soil slides in the direction of gravity. Slope direction (aspect) and curvature (whether the ground surface is concave or convex) are also used.
Geological information matters too. Even on the same slope, the speed at which water soaks in differs with the degree of rock weathering, clay content and permeability. Rainfall data takes into account both accumulated rainfall and hourly intensity. Land cover (forest, bare ground, farmland and so on) reflects the effect of roots holding the soil. Human factors, such as road cuts and distance to streams, can also be included. In practice, you start by aligning the digital elevation model (DEM), geological maps and weather observation data held by each agency to the same grid size in GIS.
What is the principle of calculating susceptibility with a statistical model?
A typical statistical model is logistic regression. The word "regression" may be unfamiliar, but the idea is simple. It compares points where landslides occurred in the past with points where they did not, and puts into numbers how strongly each combination of factors appeared together with landslides. The result is expressed as a value between 0 and 1, that is, a likelihood of occurrence.
A tool with a similar principle is the frequency ratio method. It gives a weight by calculating, for a given slope range, the ratio of actual landslides in that range relative to the share of the whole area that the range occupies. Its advantage is that the calculation is transparent, so the result is easy to explain. When a disaster safety officer explains a risk grade to a local government, this method can show which factor contributed how much, so it is widely used in practice.
How are machine learning models different from statistical models?
Machine learning models find nonlinear relationships among factors on their own. For example, a random forest builds hundreds of decision trees and produces a result by voting. A decision tree is a structure that branches on conditions like the branches of a tree, such as "dangerous if the slope is above 30 degrees and clay content is high." Combining the opinions of many trees reduces the problem of a single tree overfitting.
If you want to try a model yourself, you can start with the scikit-learn library for Python.
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
X_train, X_test, y_train, y_test = train_test_split(
features, labels, test_size=0.2, random_state=42
)
model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)
print("AUC:", roc_auc_score(y_test, model.predict_proba(X_test)[:,1]))
AUC (area under the receiver operating characteristic curve) is a measure of how well the model separates landslide points from non-landslide points. The closer to 1, the better it separates them. Before applying a model in practice, you must check this value on validation data.
How should you read the finished hazard map?
A susceptibility map shows grades in different colors by grid cell (pixel) or by small watershed. A zone marked "very high" is one with relatively high risk when the past occurrence pattern and the input factors are considered together. The point to remember is that this value is a probability, not a prediction. A landslide can occur even in a low-risk zone because of unexpected rainfall or terrain change, and a high-risk zone does not necessarily mean it will fail in this rainy season.
Some of this you only learn by trying. If you overlay the map the model produced with site photos or recent aerial images, you can check by eye how well actual landslide scars match the high-risk zones. In QGIS, load the raster layer, right-click it in the Layers panel > Properties > Symbology tab, set graduated colors, then place a satellite image layer underneath and adjust the transparency. This gives you an intuitive check of the map's reliability.
What limits should you know when using a hazard map in practice?
A model depends heavily on the quality of its training data. If the past landslide inventory is incomplete or concentrated in particular areas, the model inherits that bias as it is. Also, if the spatial resolution of the input factors is low (for example, when slope data is on a 30 m grid), it can miss fine risk zones such as narrow valleys and cut slopes.
It is also common for different models to give different results. If you analyze the same area with logistic regression and with a random forest, the location and area of the high-risk zones can differ. So rather than trusting a single model's result absolutely, you compare the results of several models or use an ensemble (combining the results of several models). In practice, keep the organization that made the hazard map, the reference year of the data used and the model validation metrics as metadata, so the next person in charge can interpret the map correctly.
Step by step on screen
The figure below is a diagram that summarizes the order of the steps. Menu positions are the same in the English interface.

- Download a DEM
From the public DEM portal of the National Geographic Information Institute (Korea), download the digital elevation model file for your study area. - Load it in QGIS
In the QGIS menu, choose Layer > Add Layer > Add Raster Layer, select the downloaded DEM file and click Add. - Calculate slope
Choose Raster > Analysis > Slope, set the DEM as the input layer and click Run (menu positions may differ by QGIS version; if you cannot find it, type 'Slope' in the search box at the top). - Set the symbology
Right-click the slope layer > Properties > Symbology tab, choose Singleband pseudocolor, set the number of classes to 5 and click Classify. - Overlay a satellite image
Under XYZ Tiles, drag the Google Satellite layer below the others in the Layers panel, and set the slope layer's transparency to 50% to view the two layers together. - Check the result
Load past landslide occurrence points (data provided by a provincial or city disaster safety department) as a vector layer and check by eye whether they coincide with steep-slope zones.
Menu names may differ slightly between program versions.
Prompts you can copy
Use this to organize landslide susceptibility results into material for residents or an internal report
You are a specialist assistant who helps write reports in the disaster safety field. Write a summary of landslide susceptibility analysis results that meets the conditions below. - Study area: [name of city, county or district] - Main factors used in the analysis: [list of the factors actually used, such as slope, geology and rainfall] - Model type: [logistic regression / random forest / other] - Share of high-risk zones: [% of total area] - Model validation metric (AUC and so on): [value] Conditions: 1. Explain technical terms briefly in parentheses so that residents and non-specialists can understand. 2. Always include a sentence stating that the hazard map is a probabilistic result and not a prediction. 3. Do not exaggerate the results or use wording saying a specific spot will certainly fail. 4. Keep it within half an A4 page (about 400 characters). 5. After writing, list separately three items that a person must check.
Replace the bracketed parts for your situation. Check the facts in any answer before you use it.
Try it yourself
- In the Vworld map service in Korea, turn on the "landslide hazard map" layer and check for yourself the grade of the area where you live.
- In QGIS, download public DEM data and run a slope analysis, then overlay it with a satellite image and compare whether the resulting colors match the terrain as you see it.
- Run logistic regression and random forest on the same area and compare how the resulting maps differ, to feel the effect of model choice for yourself.
Common questions
Q. If a landslide hazard map shows "grade 1", is the place necessarily dangerous?
A. Grade 1 means the susceptibility is relatively high among areas with similar conditions. It is not a prediction that a landslide will certainly occur. Grades can change with the data and criteria the model used, so it is important to check the metadata of the organization that made the map (data used, reference year, validation metrics).
Q. Are machine learning models always more accurate than statistical models?
A. Which model is more accurate depends on data quality, regional characteristics and the number of training samples. Machine learning models capture complex patterns well but their results are hard to explain, while statistical models are easy to interpret but can miss nonlinear relationships. It is advisable to understand the strengths and weaknesses of both approaches and choose according to your purpose.
Q. Can I make a hazard map from slope and geology alone, without rainfall data?
A. It is possible, but the accuracy may fall. Rainfall often acts as the trigger for landslides, so without a rainfall factor the model cannot reflect the timing of occurrence even on the same terrain. Combine factors according to the data you have, and always check the reliability of the result with model validation metrics.
What comes next
A landslide hazard map is the result of compressing decades of terrain and geological information and weather records into numbers. The people who read the map need to understand the difference between a probability and a prediction for it to lead to sound evacuation plans and budget decisions. Making the best judgment while knowing that the model can be wrong is the core of disaster safety practice. A later post in this series looks at how to connect hazard maps made this way to evacuation route planning and residents' evacuation decisions, along with the basics of spatial network analysis.
This is an English adaptation of a post in the Korean GeoAI basics series. The original is here, and more learning materials are on GeoAI for Disaster Safety.
This post was drafted with the help of AI tools and reviewed by the author. Tool features and data portals change over time, so check the official documentation when you practice.
Comments
Post a Comment