Published November 2026 | Version 1.0.0
Dataset Embargoed

A Systematic Benchmark of Deep Learning Architectures for Landslide Inventory Mapping with Bi-Temporal Sentinel-2 and ALOS AW3D30 in Northern Vietnam

  • 1. Vrije Universiteit Brussel
  • 2. ROR icon Hanoi University of Civil Engineering

Description

Northern Vietnam Landslide Patch Dataset (v1.0.0)
Prepared 20-band raster patches and binary labels for landslide segmentation in Northern Vietnam.
Associated code (v0.4.0): dl-landslide-vietnam

1. Dataset Overview and Inventory

This dataset accompanies a study of landslide inventory mapping using bi-temporal Sentinel-2 imagery and the JAXA ALOS World 3D–30 m (AW3D30). The release contains prepared input features and corresponding binary labels. It retains the supplied raster files, band order, and existing training, validation, and test assignments.

Partition Patch pairs TIFF files TIFF size (bytes)
TRAIN_256 3,600 7,200 22,899,763,684
VAL_256 450 900 2,863,351,906
TEST_256 450 900 2,861,845,814
Total 4,500 9,000 28,624,961,404

Each pair contains one STACK TIFF and one LABEL TIFF. The table reports raster-file sizes, excluding archive overhead and supporting files. The total raster size is approximately 28.63 GB, where 1 GB equals 1,000,000,000 bytes.

2. Data Specification and Scope

  • STACK: 20 Float32 feature bands, each measuring 256 × 256 pixels. Band names and their stored order are documented in metadata/dataset_schema.json.
  • LABEL: one UInt8 band measuring 256 × 256 pixels; 0 denotes background and 1 denotes landslide foreground.
  • Input values: the associated code expects finite feature values within [-1, 1], with an endpoint checking tolerance of 10−5. NoData or invalid pixels are not supported by this input contract.

For use with the supplied v0.4.0 configurations, read all 20 bands in their stored order without applying upstream 49-band selection indices or additional normalization. Retain the supplied label encoding and partition assignments for this workflow.

The patch TIFFs have no assigned coordinate reference system and do not provide geographic georeferencing. The associated evaluation protocol uses a nominal scale of 10 m per pixel.

3. Extraction and Verification

The release contains four ZIP64 archives:

  • dataset-v1.0.0-train.zip
  • dataset-v1.0.0-val.zip
  • dataset-v1.0.0-test.zip
  • dataset-v1.0.0-metadata.zip

Extract all four archives into the same parent directory. Each archive already contains the dl-landslide-vietnam/ root folder. In Windows File Explorer, select the same extraction destination for all four archives rather than the automatically suggested archive-specific folders.

dl-landslide-vietnam/
├── data/
│   └── SAMPLE_DATASET_256x256/
│       ├── TRAIN_256/
│       │   ├── STACK/
│       │   └── LABEL/
│       ├── VAL_256/
│       │   ├── STACK/
│       │   └── LABEL/
│       └── TEST_256/
│           ├── STACK/
│           └── LABEL/
├── metadata/
├── documentation/
└── scripts/

The following Windows PowerShell examples require Python 3.12. They assume that the four ZIP archives, SHA256SUMS.txt, and the standalone verify_dataset.py are saved in D:\LandslideDownload.

Before extraction, verify the downloaded archives:

py -3.12 "D:\LandslideDownload\verify_dataset.py" --archives "D:\LandslideDownload"

Extract all four archives to D:\LandslideWork, then verify the extracted TIFF files:

py -3.12 "D:\LandslideWork\dl-landslide-vietnam\scripts\verify_dataset.py"

ZIP archives use STORE mode and retain the original lossless LZW-compressed TIFF contents.

4. License, Data Sources and Citation

The prepared STACK/LABEL dataset is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) , to the extent of rights held by the licensor. This license does not replace applicable third-party source terms or alter rights independently granted by data providers. See documentation/LICENSE-DATA.md.

Helper scripts are licensed separately under the MIT License; see scripts/LICENSE.txt.

  • Copernicus Sentinel-2: Contains modified Copernicus Sentinel data 2024. See the Sentinel Data Legal Notice .
  • JAXA ALOS AW3D30: Terrain-derived features originate from the ALOS World 3D–30 m digital surface model provided by the Japan Aerospace Exploration Agency (JAXA). The reported source collection is JAXA/ALOS/AW3D30/V4_1. This collection includes data from versions 3.1, 4.0 and 4.1; individual source-tile versions have not been independently verified for this release. See the collection description and provider terms .

Files

Embargoed

The files will be made publicly available on January 1, 2027.

Additional details

Related works