Global Datasets
Chinese Academy of Sciences - Landslide Dataset
Description: • integrating satellite and unmanned aerial vehicle data from nine regions (specifies region)
Number of Landslide Records: 20,865 RGB images
Dataset Details: Dataset is subdivided into img, label, and mask
Landslide Inventory Type: polygon
Models Used: - DeepLabV3+ (using ResNet50 as backbone) - U-Net (using ResNet50 as backbone) - MFFENet (using ResNet50 as backbone) - FCN (Fully Convolutional Network) (using VGG16 as backbone)
Input Model: - Remote sensing images - Ground truth segmentation masks (binary masks where landslide-affected areas are labeled as foreground and non-landslide areas as background)
Data Resolution: 512 × 512 pixels, Ground Resolution is provided in m and varies
Output Model: - segmentation mask separated where each pixel is classified as: Foreground (landslide) or Background (non-landslide) - satelitle images performed better than Unmanned Aerial Vehicles
Other Paper Information: NaN
Other Dataset Information: Integrates satellite and unmanned aerial vehicle (UAV) data for 9 regions [which are separated]. The sensors used in each dataset is specified in the README
HR-GLDD: Global Dataset
Description: • rainfall triggered and five earthquake-triggered multiple landslide events
Number of Landslide Records: 10
Dataset Details: Dataset is divided to what seems like 60% training, 20% testing, 20% validation. Running code on the testX.npy dataset the shape is: (355, 128, 128, 4) they do not say what the 4th channel is
Landslide Inventory Type: polygon
Models Used: NaN
Input Model: NaN
Data Resolution: 128 x 128 pixels (spatial resolution of up to 3 m)
Output Model: NaN
Other Paper Information: NaN
Other Dataset Information: They come from 10 different physiographical regions globally but the datset is not separated by region.
Landslide4Sense: Reference Benchmark Data and Deep Learning Models for Landslide Detection
Description: • multi-source satellite remote sensing imagery
Number of Landslide Records: 4,844 image patches
Dataset Details: training/validation/test, consisting of 3799, 245, and 800 image patches 14 bands consisting of: Multispectral data from Sentinel-2: B1, B2, B3, B4, B5, B6, B7, B8, B9, B10, B11, B12. Slope data from ALOS PALSAR: B13. Digital elevation model (DEM) from ALOS PALSAR: B14. All bands are resized to the resolution of ~10m per pixel
Landslide Inventory Type: polygon
Models Used: NaN
Input Model: NaN
Data Resolution: 128 x 128 pixels
Output Model: NaN
Other Paper Information: NaN
Other Dataset Information: NaN
GDCLD: Global Dataset of Coesismic Landslide Mapping
Description: • coseismic landslide mapping via multi-source high-resolution remote sensing images
Number of Landslide Records: 466
Dataset Details: About 66% for training, 27% for validation, and 6% for testing
Landslide Inventory Type: polygon
Models Used: - U-Net - ResU-Net - DeepLabV3 - HRNet - UPerNet - SwinU - SegFormer (performed the best)
Input Model: - multi-source remote sensing images (which they found performed better than single-sourced)
Data Resolution: 1024 pixels x 1024 pixels Resolutions varies and there is a table in the paper with this information
Output Model: - semantic segmented image - Show that transformer architecture better than the conventional CNN architecture in the task of landslide identification using multi-source remote sensing imagery
Other Paper Information: NaN
Other Dataset Information: Training Set - Train Data: 11,162 TIFF matrices of shape (1024, 1024, 3) - Train Labels: 11,162 TIFF matrices of shape (1024, 1024, 1) Validation Set - Validation Data: 4,459 TIFF matrices of shape (1024, 1024, 3) - Validation Labels: 4,459 TIFF matrices of shape (1024, 1024, 1) Training and Validation Sets contain UAV, PlanetScope, Gaofen-6 and Map World images of the 5 earthquake regions of Luding, Nippes, Hokkaido, Jiuzhaigou and Mainling. There is no overlapping area in each TIFF. Test Set (Lushan, Sumatra, Mesetas, and Palu Earthquake Regions) - Contains: Remote sensing images from UAV, Map World, and PlaneScope sources cross-regional validation was conducted by training and validating models on seismic events from five regions (Luding, Jiuzhaigou, Hokkaido, Mainling, and Nippes) and testing on four independent regions (Lushan, Mesetas, Sumatra, and Palu)
NASA Cooperative Open Online Landslide Repository
Description: NaN
Number of Landslide Records: 39634
Dataset Details: LHASA 2.0: Model trained on 2015-2018 landslide events, tested on 2019-2020 data.
Landslide Inventory Type: polygon, point
Models Used: LHASA 2.0: XGBoost (Machine Learning Model) LHASA 1.1: Decision Tree Model. LHASA 2 provides a nearly real-time view of global landslide hazard for a variety of stakeholders
Input Model: LHASA Version 1 – Input Data: - Satellite precipitation (7-day rainfall) - Static susceptibility map (based on slope, land cover, geology) LHASA Version 2 – Input Data: - Daily precipitation - Antecedent rainfall - Soil moisture - Snow cover/snowmelt - Explanatory variables (e.g., slope, land cover, faults) - Landslide inventories (GLC + others for training)
Data Resolution: LHASA v1 Output Resolution: ~0.1°, every 3 hours LHASA v2 Output Resolution: ~0.1°, daily
Output Model: LHASA v1 - Output: Categorical (High, Moderate, Low hazard) LHASA v2 - Output: Probabilistic landslide nowcast (0–1 score)
Other Paper Information: COOLR (Cooperative Open Online Landslide Repository) - An open-access global landslide catalog. - Aggregates citizen-reported landslides and event-based inventories. - Used for validating landslide hazard models, including LHASA. LHASA (Landslide Hazard Assessment for Situational Awareness) - LHASA 2.0: A global nowcasting model combining satellite-based rainfall data with landslide susceptibility factors. - LHASA 1.1: A heuristic-based global susceptibility mapping model.
Other Dataset Information: Model is databased not trained on landslides but landslides are used for cooboration of results