County Landslide Risk Mapping with Remote Sensing

ISEF Category: Earth and Environmental Sciences

Subcategory: Geosciences  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

A hillside can look calm and still fail in seconds. That makes landslides hard to predict, but not impossible to study. With open satellite data and mapping tools, you can turn a county into a risk map and test what terrain features matter most.

What Is It?

Landslide-susceptibility mapping asks a simple question: where is the ground more likely to fail? You are not predicting the exact day or storm. You are ranking places by risk, like marking parts of a road map as safer or more dangerous.

Your model uses clues from the land itself. Elevation models show slope, roughness, and shape of the terrain. Land cover data shows whether the surface has forest, crops, bare soil, or buildings. Historical landslide points tell the model where failures already happened. A U-Net is a neural network often used for image-style mapping, so it can learn patterns across a grid instead of just in a table.

Think of it like training a weather app, but for hills. The input layers are the ingredients, and the output is a risk score for each pixel or cell in the county. Your job is to see which ingredients improve the map and how well the map matches known landslide records.

Why This Is a Good Topic

This is a strong science fair topic because it has real stakes, clear data, and measurable outputs. Landslides threaten homes, roads, and emergency planning, so your project connects directly to public safety. You can test whether terrain, land cover, or model design changes accuracy, and you can judge success with metrics like precision, recall, and ROC-AUC.

Research Questions

• How does adding land cover data change landslide prediction accuracy compared with elevation data alone?
• What is the effect of DEM source, SRTM versus Copernicus, on the final susceptibility map?
• Does a U-Net outperform a simpler machine learning model when trained on the same landslide catalog points?
• To what extent does class imbalance handling improve detection of known landslide zones?
• Which terrain derivatives, slope, aspect, curvature, or roughness, contribute most to model performance?
• How does changing the size of the mapping grid affect hotspot detection and false positives?
• What is the effect of excluding urban areas or water bodies on model calibration?

Basic Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Stable internet access for downloading open geospatial data.
• QGIS for viewing and checking raster and vector layers.
• Python installed with geospatial and machine learning libraries.
• Open SRTM or Copernicus DEM data for the target county.
• MODIS land cover data for the target county.
• Historical NASA landslide catalog points or another public landslide inventory.
• External drive or cloud storage for large raster files.
• Spreadsheet software for organizing metadata and results.

Advanced Materials

• High-performance laptop or workstation with a dedicated GPU.
• Python environment with TensorFlow or PyTorch, rasterio, geopandas, numpy, scikit-learn, and xarray.
• Jupyter Notebook for model development and analysis.
• ArcGIS Pro if your school has access, for comparison with open-source mapping workflows.
• High-resolution local elevation data, if available from a state or county GIS portal.
• Additional validation data from field reports, road closure records, or published landslide inventories.
• GIS-ready watershed or county boundary shapefiles.
• ImageJ or similar image analysis software for visual inspection of map outputs.

Software & Tools

• QGIS: Reviews terrain layers, clips rasters to your county, and checks whether data line up correctly.
• Python: Runs data preparation, model training, and accuracy testing.
• Jupyter Notebook: Keeps your workflow organized and makes it easier to document each step.
• Google Earth Engine: Helps you explore large land cover and remote sensing datasets without downloading everything first.
• scikit-learn: Calculates baseline models and evaluation metrics like precision, recall, and ROC-AUC.

Experiment Steps

1. Define the county boundary, the target landslide inventory, and the exact prediction unit you will map.
2. Choose your input layers, then decide which terrain derivatives and land cover classes belong in the first model.
3. Build a clean training table or image stack, and plan how you will separate training, validation, and test areas.
4. Set up a baseline model so you can compare the U-Net against a simpler method.
5. Plan the evaluation metrics that match the problem, including how you will handle class imbalance and spatial autocorrelation.
6. Design a map comparison strategy that checks whether predicted hotspots overlap known landslide points better than chance.

Common Pitfalls

• Using landslide point locations as both training and test data, which inflates accuracy.
• Mixing rasters with different resolutions or projections, which shifts features out of alignment.
• Treating all nearby pixels as independent samples, which makes the model look better than it really is.
• Ignoring class imbalance, which can produce a map that predicts almost everything as stable ground.
• Trusting the model output without checking whether roads, bare slopes, or cloud gaps are creating false hotspots.

What Makes This Competitive

A competitive project goes beyond making a pretty risk map. You need careful validation, a clear baseline, and a reasoned choice of model inputs. Strong entries compare multiple DEM sources, test several feature sets, and use spatially separated test regions so the score reflects real generalization. You can also add a novel twist, like checking whether the model transfers to a second county or whether certain terrain variables dominate in different geology zones.

Project Variations

• Try the same workflow on a second county with a different geology and compare how well the model transfers.
• Swap the U-Net for random forest or XGBoost and test whether a simpler model performs nearly as well.
• Focus on post-fire or deforested slopes and see whether land cover change shifts susceptibility patterns.

Learn More

• NASA Landslide Catalog: Search the catalog for historical landslide points and event details used for validation.
• USGS Landslide Hazards Program: Find background on landslide causes, mapping methods, and public safety context on the USGS site.
• NOAA National Centers for Environmental Information: Use climate and precipitation data to test whether rainfall patterns line up with mapped risk.
• Copernicus DEM Documentation: Read product notes and resolution details on the Copernicus Land Monitoring Service site.
• MIT OpenCourseWare Remote Sensing and GIS courses: Review free lecture material on raster analysis, map projections, and spatial modeling.
• PubMed: Search review articles on landslide susceptibility mapping, machine learning, and terrain analysis.