kinit/landslides

!!! Important note: Dataset creation is still in progress !!!

Introduction

This dataset was created through a collaboration between the State Geological Institute of Dionýz Štúr and the Kempelen Institute of Intelligent Technologies. It covers the Biele Karpaty mountain range, located in the western part of Slovakia.

The dataset is organized as follows:

/landslides
├── layers/
│ ├── processed/
│ │ ├── test/
│ │ ├── train/
│ │ └── validation/
│ └── raw/
│ │ ├── test/
│ │ ├── train/
│ └ └── validation/
├── raw/
└── README.md

The /landslides/raw/ folder contains a 1x1 meter resolution Digital Elevation Model (DEM) of Biele Karpaty (lidar.tif) and expert-annotated landslide data from the State Geological Institute of Dionýz Štúr (landslides.tif).

General information

Dataset split	Train	Validation	Test	Total
Area (km2)	265.075141	169.684267	171.510959	606.270367
Minimal elevation (meters above mean sea level)	227.262	201.377	190.486	160.486
Maximal elevation (meters above mean sea level)	916.335	776.459	905.524	916.335
Number of landslides	1155	564	391	2110
Largest landslide area (m2)	1,904,451	1,005,255	933,910	1,904,451
Smallest landslide area (m2)	136	305	430	136
Average landslide area (m2)	54,509.82	40,887.01	60,343.41	51,949.47

Layers

The /landslides/layers/processed/ and /landslides/layers/raw/ directories each contain subfolders for the test, train, and validation sets. These include various geospatial layers derived from the raw data.

• The raw folder contains unnormalized data.
• The processed folder includes enhanced, normalized versions of these layer

The following layers are available:

• LiDAR
• Landslides
• Curvature
• Hillshade
• NDVI
• Roughness
• Slope

A detailed description and visualization of each layer are provided below.

LiDAR

The LiDAR layer was derived from a Digital Elevation Model (DEM) with a spatial resolution of 1×1 meter.

• The raw folder contains the original, unprocessed elevation data.
• The processed folder contains normalized elevation data. Normalization was performed using a minimum value of 0.0 and a maximum value equal to the highest point in the Biele Karpaty region: 916.335 meters.

Landslides

The Landslides layer represents ground truth data, manually annotated by experts from the State Geological Institute of Dionýz Štúr. A pixel value of 1 indicates the presence of a landslide. The data in both the raw and processed folders are identical, as no additional processing was applied to this layer.

Aspect

TODO

Curvature

The Curvature layer was calculated from the DEM after applying a smoothing filter to reduce noise. A 3×3 kernel was used to smooth the DEM before curvature computation. The curvature was computed using the following formula (see accompanying image for the legend):

• A = (Z1 + Z9) / 2 - Z5
• B = (Z7 + Z9) / 2 - Z5
• C = (Z2 + Z8) / 2 - Z5
• D = (Z4 + Z6) / 2 - Z5
• Curvature = A + B + C + D

Image source: https://help.arcgis.com/en/arcgisdesktop/10.0/help/index.html#//00q90000000t000000

A fixed distance L = 10 meters was used to ensure the output captures broader terrain features and appears smoother.

• The raw folder contains curvature values directly computed from the formula above
  • Edge artifacts were cropped
  • Positive values indicate convex surfaces
  • Negative values indicate concave surfaces
• The processed folder contains a clipped and normalized version:
  • Values were clipped to the range [-2, 2]
  • The clipped values were then normalized to the [0, 1] range using:
    • Minimum = -2 → maps to 0.0
    • Maximum = 2 → maps to 1.0
  • In the normalized data:
    • Values below 0.5 represent concave surfaces
    • Values above 0.5 represent convex surfaces

Hillshade

The Hillshade layer was generated from the DEM using the multidirectional hillshade function in QGIS, with a 45° vertical angle to simulate natural illumination.

• The raw folder contains an RGB image with pixel values ranging from 0 to 196
• The processed folder includes a single selected band extracted from the RGB image. The following enhancements were applied:
  • CLAHE (Contrast Limited Adaptive Histogram Equalization) filter from the OpenCV library:
    • Clip limit: 20
    • Tile grid size: 8×8
  • The output was then normalized to the [0, 1] interval

Normalized Difference Vegetation Index (NDVI)

The NDVI layer was computed from the annual average satellite data covering the Biele Karpaty region (Sentinel 2 surface reflectance imagery). It is used to determine the health and density of vegetation and is computed as a normallized difference of bands B8 (Near-Infrared) and B4 (Visible Red). The data in both the raw and processed folders are identical, as no further processing was applied to this layer.

Negative openness

TODO

Positive openness

TODO

Possitive and negative openness difference

TODO

Roughness

The Roughness layer was derived from the DEM using the roughness function in QGIS, which measures the variability of elevation within a local neighborhood.

• The raw folder a single-band raster with the original roughness values, unprocessed
• The processed folder contains a normalized version of the data. The following processing steps were applied:
  • Clipping: Values above 20 were clipped to 20 to reduce the influence of extreme outliers
  • Logarithmic transformation: The clipped values were transformed using: ln(roughness + 1.0)
  • Normalization: minimum value set to 0, maximum value set to ln(21), resulting values were scaled to the range [0, 1]

Slope

The Slope layer was derived from the DEM using the slope function in QGIS, which calculates the steepness of the terrain in degrees.

• The raw folder contains a single-band raster with the original slope values expressed in degrees of inclination
• The processed folder contains a normalized version of the data: minimum value set to 0°, maximum value set to 90°, values were scaled linearly to the range [0, 1]