LAnd Segment (LAS)

LAS is a large-scale, multi-source Earth observation dataset curated for flexible land use and land cover (LULC) semantic segmentation. It combines high-resolution imagery with accurately annotated masks and medium-to-low-resolution multispectral imagery with weak labels derived from existing LULC products. LAS was introduced in LandSegmenter: Towards a Flexible Foundation Model for Land Use and Land Cover Mapping.

LAS is designed to support the training of models that can handle different sensors, spatial resolutions, spectral bands, and class taxonomies without requiring a single, manually harmonized labeling scheme.

• Paper: ISPRS Journal of Photogrammetry and Remote Sensing (2026) · arXiv:2511.08156
• Code and documentation: zhu-xlab/LandSegmenter
• Model weights: vikki23/LandSegmenter
• Dataset: vikki23/LAS

Dataset overview

LAS contains approximately 150,000 geographically sampled locations, corresponding to 311,000 image patches and 200,000 semantic label masks, as reported in the paper and project repository. The curated training collection comprises eight source subsets, with an approximate 1:4 ratio of accurately labeled to weakly labeled data.

Source subset	Imagery	Approx. GSD	Image bands	Patch size	Label type
OpenEarthMap	High-resolution aerial/satellite RGB	0.25–0.5 m	3	320 × 320	Accurate annotations
DynamicEarthNet	Planet RGB-NIR	3–4 m	4	256 × 256	Accurate annotations
Iran	Sentinel-2	10 m	12–13	264 × 264	Weak annotations
GHSL	Sentinel-2	10 m	12–13	264 × 264	Weak annotations
WorldCover	Sentinel-2	10 m	12–13	264 × 264	Weak annotations
NLCD	Landsat-8/9	30 m	7–11	264 × 264	Weak annotations
USFS	Landsat-8/9	30 m	7–11	264 × 264	Weak annotations
SBTN	Landsat-8/9	30 m	7–11	264 × 264	Weak annotations

The reported spatial resolutions, band-count ranges, and patch sizes describe the curated subsets. Exact spectral channels and class vocabularies depend on the source dataset. The WorldCover subset is provided in separate v100 and v200 archives; these are two product versions, not additional source-subset categories.

Repository contents

The repository contains the LAS training data as well as separate processed datasets for the fine-tuning and evaluation experiments in LandSegmenter.

LAS/
├── train_weak_raw/        # Original packaged weakly supervised data
├── train_weak_lmdb/       # Processed weakly supervised training data
├── train_exact_lmdb/      # Processed accurately annotated training data
├── fine_tune_train_lmdb/  # Fine-tuning datasets for downstream benchmarks
└── test_lmdb/             # Evaluation datasets for downstream benchmarks

LAS training subsets

Weakly supervised training data (train_weak_lmdb/):

ghsl_lmdb.tar.gz
iran_lmdb.tar.gz
nlcd_lmdb.tar.gz
sbtn_lmdb.tar.gz
usfs_lmdb.tar.gz
worldcover_v100_lmdb.tar.gz
worldcover_v200_lmdb.tar.gz

Accurately annotated training data (train_exact_lmdb/):

dynamicearthnet_lmdb.tar.gz
openearthmap_lmdb.tar.gz

Raw weakly supervised data (train_weak_raw/): Source archives are supplied as .zip files, some divided into numbered .part000, .part001, etc. files. Download all numbered parts of a given archive before reassembling it.

Additional benchmark archives

The following datasets are provided for reproducing LandSegmenter fine-tuning and evaluation, rather than as additional LAS source subsets:

Dataset	Fine-tuning archive	Evaluation archive
Dynamic World (DW)	dw_train_lmdb.tar.gz	dw_test_lmdb.tar.gz
LoveDA	loveda_train_lmdb.tar.gz	loveda_val_lmdb.tar.gz
MultiSenGE	multisenge_train_lmdb.tar.gz	multisenge_test_lmdb.tar.gz
NYC	nyc_train_lmdb.tar.gz	nyc_test_lmdb.tar.gz
OSM	osm_train_lmdb.tar.gz	osm_test_lmdb.tar.gz
Potsdam	potsdam_train_lmdb.tar.gz	potsdam_val_lmdb.tar.gz

Download and use

The files are large and distributed as ZIP archives and compressed LMDB directories, rather than as a directly loadable Hugging Face datasets table. We recommend downloading only the subsets needed for an experiment.

Install the Hugging Face Hub CLI:

pip install -U huggingface_hub

Download one processed weakly labeled subset, for example WorldCover v100:

hf download vikki23/LAS \
  --repo-type dataset \
  --include "train_weak_lmdb/worldcover_v100_lmdb.tar.gz" \
  --local-dir ./LAS

mkdir -p ./data/weak
tar -xzf ./LAS/train_weak_lmdb/worldcover_v100_lmdb.tar.gz -C ./data/weak

To download another subset, replace the path supplied to --include with the corresponding archive path shown above. For a raw ZIP split into numbered parts, concatenate its parts in numerical order before extracting, for example:

cat ./LAS/train_weak_raw/ghsl.zip.part* > ghsl.zip
unzip ghsl.zip -d ./data/ghsl

Important: The extraction layout and LMDB contents are specific to the LandSegmenter preprocessing pipeline. Refer to the official implementation for dataset loading, class definitions, band handling, and training configuration. In particular, set data_dir in dataset/datasets_settings.py to the directory containing the extracted LMDB datasets. Standard load_dataset("vikki23/LAS") is not an LMDB reader and is not the intended loading method for these archives.

Data creation and annotations

LAS was curated to reduce dependence on extensive manual pixel-level labeling in Earth observation. Its weakly supervised component uses existing LULC products as semantic supervision for satellite images selected across geographically diverse regions. Its accurately labeled component incorporates OpenEarthMap and DynamicEarthNet.

The data span multiple observation scales, sensor modalities, and taxonomies. Class labels are source-specific: a class ID in one subset should not be assumed to have the same meaning as the same ID in another subset. Consult the source taxonomies and the LandSegmenter dataset configuration before combining subsets or interpreting label values.

Intended uses

LAS is intended for research on LULC semantic segmentation, weakly supervised learning, multi-source and multispectral model training, and transfer across spatial resolutions or semantic taxonomies. The accompanying fine-tuning and evaluation archives enable reproduction of the experiments reported in LandSegmenter.

Limitations and responsible use

• Weak-label uncertainty: Existing LULC products can contain errors, omissions, inconsistent boundaries, and temporal mismatches with the paired imagery. Weak labels should not be treated as manually verified ground truth.
• Heterogeneous taxonomies: Semantic classes, label IDs, available bands, and preprocessing choices vary by source. Combining subsets may require source-aware handling.
• Sampling and transfer: Geographic sampling increases spatial diversity but does not guarantee uniform coverage of all regions, environments, classes, or acquisition conditions. Performance on underrepresented domains may differ.
• Evaluation separation: The fine_tune_train_lmdb/ and test_lmdb/ folders are separate benchmark resources. Users should avoid mixing evaluation material into training when reporting comparable results.
• High-stakes applications: Model outputs trained on LAS should be independently validated before use in consequential mapping or policy decisions.

Licensing and attribution

The repository currently declares Apache-2.0 in its Hugging Face metadata. LAS incorporates imagery and labels originating from third-party datasets and geospatial products, which may carry their own licenses, attribution obligations, and restrictions. The repository-level license should not be interpreted as overriding the terms attached to upstream material. Before redistribution or commercial use, review the terms of the specific source subsets and cite the relevant original datasets/products in addition to the LAS paper.

Citation

If you use LAS or the provided processed benchmark archives, please cite:

@article{liu2026landsegmenter,
  title   = {LandSegmenter: Towards a flexible foundation model for Land Use and Land Cover mapping},
  author  = {Chenying Liu and Wei Huang and Xiao Xiang Zhu},
  journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
  volume  = {238},
  pages   = {409--421},
  year    = {2026},
  doi     = {10.1016/j.isprsjprs.2026.04.056},
  url     = {https://doi.org/10.1016/j.isprsjprs.2026.04.056}
}

Contact

For questions about LAS and LandSegmenter, please use the GitHub repository or contact Chenying Liu at chenying.liu@tum.de/chenying.liu023@gmail.com.