Datasets:
The dataset viewer is not available for this split.
Error code: RowsPostProcessingError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
SAER Mining Disturbance Research Bundle
Research dataset accompanying: Sparsity-Aware Evidence Reconstruction for Regional Mining Disturbance Mapping (manuscript submitted to GIScience & Remote Sensing; publication status not yet established).
- Dataset repository: https://huggingface.co/datasets/XiaonianD/SAER-Research-Bundle
- Citation: Use the version-specific DOI shown in the Hugging Face repository DOI section; see Citation.
- Study area: Beijing–Tianjin–Hebei region, China
- Annual mapping period: 2017–2025
- Repository format: Five downloadable ZIP archives and supporting text/manifest files
- Access: Public; no Hugging Face account or approval is required for normal file downloads
- License: Component-specific; original team-owned annotations, splits and derived mapping products are CC BY 4.0. See Licensing and the repository LICENSE file.
Overview
This repository preserves the research materials accompanying the SAER project: scene-separated Sentinel-2 image samples, segmentation labels, hard-negative samples, training and evaluation code, three Full SAER checkpoints, and annual mining-disturbance mapping products for the Beijing–Tianjin–Hebei region.
The five ZIP archives preserve the original project directory layout. Each archive can be extracted independently, but all five must be extracted into the same parent directory to preserve the expected relative paths. This repository is a downloadable research bundle rather than a Hugging Face load_dataset() package. The Dataset Viewer and automated model inference are not part of the supported workflow.
Repository contents
| File | Size (bytes) | Contents |
|---|---|---|
01_Code_and_Configuration.zip |
127,257,655 | Training, evaluation, inference, vectorization, supplementary experiments, figure scripts, and original dependency files |
02_Samples_and_Splits.zip |
4,659,121,067 | Sample GeoTIFFs, positive-sample PNG masks, hard-negative samples, normalization statistics, scene split records, and relative-path location index |
03_Models.zip |
3,323,574,233 | Full SAER checkpoints and configuration/log files for seeds 42, 43, and 44 |
04_Research_Results.zip |
294,628,740 | Annual raster/vector products, statistics, evaluation reports, manuscript figures, and supplementary materials |
05_Documentation.zip |
30,172,229 | Reproduction instructions, data notes, provenance/checksum records, and a copy of the associated manuscript |
Total archive size: 8,434,753,924 bytes (approximately 8.43 GB, decimal units). No multipart .partNNN merging is required. The archived project contents were preserved during packaging rather than rewritten into a new data structure.
Additional root-level files:
ORIGINAL_README.txt: Original Chinese bundle overview.DATA_NOTES.txt: Data description and encoding notes.REPRODUCTION.txt: Reproduction instructions.archive_manifest.json: Archive sizes, SHA-256 hashes, member counts, and unpacked sizes.SHA256SUMS.txt: Checksums recorded at the original packaging date. Archive hashes remain usable after documentation updates; historical checksum entries for edited documentation files may no longer match.
Data samples and splits
The packaged notes and archive inventory identify 2,659 image samples:
| Sample type | Training | Validation | Test | Total |
|---|---|---|---|---|
| Positive images with corresponding PNG masks | 862 | 183 | 187 | 1,232 |
| Hard-negative images | 968 | 242 | 217 | 1,427 |
| Total | 1,830 | 425 | 404 | 2,659 |
The original positive-sample split audit records 20 training scenes, 7 validation scenes, and 5 test scenes, with no scene overlap. The historical split_manifest.csv contains absolute paths from the original project environment. For portable paths and geospatial metadata, use 02_Samples_and_Splits/样本划分与位置索引.csv, which records package-relative locations, CRS, bounds, and affine transforms.
The historical manifest labels 459 validation/test hard-negative images as excluded. This value should not be interpreted as indicating that those samples are absent from the evaluation archives: their split membership is reflected in the packaged directories and relative-path index. Separate PNG mask files have not been generated for the hard-negative images.
Raster formats and labels
- Recorded Sentinel-2 spectral band order: B2, B3, B4, B8.
- Positive-sample PNG segmentation masks: 0 = background; 255 = foreground.
- Annual prediction GeoTIFFs: 0 = background; 1 = mining disturbance; 255 = NoData.
- Georeferencing for positive PNG masks should be obtained from the corresponding GeoTIFF grid.
Important: Do not treat the value 255 identically across training PNG masks and annual prediction GeoTIFFs. It represents foreground in the former and NoData in the latter.
The annual prediction rasters are per-scene GeoTIFFs, not nine complete region-wide mosaic rasters. Original annual vector files and region-clipped/merged vector products are supplied separately. Consult their metadata before spatial processing; the data notes record EPSG:4326 for annual prediction rasters and EPSG:6933 for equal-area statistics.
Download and integrity verification
The repository is public, and normal downloads do not require Hugging Face authentication.
Install the Hugging Face CLI and download the repository:
python -m pip install -U huggingface_hub
hf download XiaonianD/SAER-Research-Bundle --repo-type dataset --local-dir ./SAER_download
For a smaller access check, download DATA_NOTES.txt from the Files tab before downloading the full archives.
On Windows PowerShell, verify SHA-256 checksums of the five ZIP archives before extraction. This deliberately skips checksums for README and other metadata files that may have been revised after the original packaging:
Set-Location ./SAER_download
Get-Content ./SHA256SUMS.txt | ForEach-Object {
$expected, $name = $_ -split ' ', 2
if ($name -notlike "*.zip") { return }
$actual = (Get-FileHash -LiteralPath $name -Algorithm SHA256).Hash.ToLowerInvariant()
if ($actual -ne $expected) { throw "Checksum mismatch: $name" }
}
Extract each archive into the same directory:
python -m zipfile -e 01_Code_and_Configuration.zip ./SAER
python -m zipfile -e 02_Samples_and_Splits.zip ./SAER
python -m zipfile -e 03_Models.zip ./SAER
python -m zipfile -e 04_Research_Results.zip ./SAER
python -m zipfile -e 05_Documentation.zip ./SAER
Copy-Item ORIGINAL_README.txt ./SAER/README.txt
Use a new extraction destination with sufficient free disk space. The unpacked archives require approximately 10.25 GB in addition to the downloaded archives. Keep all five numbered directories as siblings. In particular, 01_Code_and_Configuration/000000TGRS and 01_Code_and_Configuration/SAER_HG_v1_run must remain siblings.
Read REPRODUCTION.txt, DATA_NOTES.txt, and the extracted project documentation before running the code. Paths saved from the original computer environment must be adapted to the new installation.
Model checkpoints and reproduction scope
The archived annual mapping record identifies the following configuration:
- Model: Full SAER
- Mapping seed: 42
- Checkpoint epoch: 43
- Pixel threshold: 0.35
- Pretrained backbone:
nvidia/segformer-b5-finetuned-ade-640-640
Full SAER checkpoints and associated run information are included for seeds 42, 43, and 44. The upstream pretrained-backbone download cache is not included; the relevant model must be obtained separately according to the upstream provider's terms.
Known reproduction limitations
- Full-region raw Sentinel-2 source scenes are not bundled; the sample GeoTIFFs and derived mapping outputs are included as documented.
- Baseline, hard-negative variant, and all ablation checkpoints have not been fully assembled into this archive. Reproducing every comparison may require additional materials or rerunning the corresponding experiments.
- Dependency files reflect the original project environment rather than a newly tested, fully pinned software environment.
- Training and inference were not rerun during packaging. The presence of scripts and checkpoints must not be interpreted as an independent verification that every published numerical result can be reproduced from this archive alone.
These limitations are disclosed to help users distinguish archived materials from claims of end-to-end reproducibility.
Data provenance and source materials
Satellite imagery obtained through Google Earth Engine
All 2,659 packaged multispectral satellite image samples are derived from Copernicus Sentinel-2 imagery accessed through Google Earth Engine (GEE). The corresponding mining-disturbance segmentation labels were created by the SAER research team. GEE was used as the imagery access/processing platform; it is not the underlying satellite-data rights holder. The underlying imagery remains governed by the Copernicus Sentinel Data Legal Notice, not by the SAER team's CC BY 4.0 grant.
- GEE Sentinel-2 data catalog: https://developers.google.com/earth-engine/datasets/catalog/sentinel-2
- Copernicus Data Space Ecosystem: https://dataspace.copernicus.eu/
- Copernicus Sentinel Data Legal Notice: https://cds.climate.copernicus.eu/licences/ec-sentinel
- Recommended attribution: Contains modified Copernicus Sentinel data (2017–2025).
The specific GEE collection identifier (for example, Level-1C versus Level-2A and harmonized versus non-harmonized) and any preprocessing settings must be obtained from the original project scripts/processing records. They are not inferred here. Full-region source scenes are not redistributed in this bundle; the processed samples and annual mapping outputs are provided as described above.
Other reference imagery, where applicable
The statement above describes the distributed sample imagery; it does not assert that any additional imagery was or was not used in manuscript figures or visual validation. If there are other third-party images, those remain subject to their providers' terms. If an unshared third-party input is necessary to verify an analysis, the paper's Data Availability Statement must specify its actual source, redistribution restriction, and access procedure. No ethical, institutional-review-board or security restriction is claimed here without supporting facts.
Access and licensing
Public accessibility
The dataset repository is publicly listed and gated access has been disabled. Readers can visit the Files tab and download the deposited files without requesting repository approval. For reproducibility, users should verify downloaded files using SHA256SUMS.txt.
Licensing and third-party materials
The repository is publicly downloadable and contains materials governed by different licenses. Please consult the repository LICENSE file for the component-level scope, exclusions, and upstream notices. Its Hugging Face metadata therefore uses license: other, which does not mean that the original research data lack an open license.
| Component | Terms of reuse |
|---|---|
| Original team-owned mining-disturbance PNG annotations, original dataset split records and accompanying indices, and original derived annual mapping outputs and statistics | Creative Commons Attribution 4.0 International (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/ |
| Sentinel-2 image content in the 2,659 multispectral samples accessed through GEE | Copernicus Sentinel Data Legal Notice; Google Earth Engine is the access/processing platform, and the SAER team does not relicense the underlying imagery |
| Original SAER team-authored code | MIT License, only for code whose copyright holders have authorized release under those terms; third-party code keeps its upstream licenses |
| SAER checkpoint weights and inherited pretrained-model components | Applicable upstream model-license terms, including the NVIDIA SegFormer license where relevant; these files are excluded from the CC BY data grant |
| High-resolution third-party reference imagery (if contained in accompanying figures or documentation) | Original provider terms; excluded from the CC BY and MIT grants |
Associated manuscript included in 05_Documentation.zip |
Separate manuscript copyright terms; excluded from the research-data and software grants |
The research team's CC BY 4.0 grant covers only original research-generated material for which it holds the necessary licensing rights, not the copyright of underlying Copernicus imagery or external material. CC BY permits reuse, adaptation, redistribution, and commercial use with appropriate attribution and identification of changes. The original research data remain reusable under this open license even though some other archive components carry distinct terms.
Copernicus acknowledgement: Contains modified Copernicus Sentinel data (2017–2025).
Upstream model information: https://huggingface.co/nvidia/segformer-b5-finetuned-ade-640-640 and https://github.com/NVlabs/SegFormer/blob/master/LICENSE. The upstream SegFormer license contains a non-commercial research/evaluation use restriction and requires inclusion of its complete license upon redistribution of covered works. The SAER dataset and code licenses do not remove those conditions. Check any accompanying upstream notices in the archives before further redistribution.
Where an actual third-party input is necessary to verify a reported result but cannot legally be redistributed, the manuscript Data Availability Statement should identify its provider, the specific legal restriction, and how readers or reviewers can apply for access. Merely stating that data are available on request is not a substitute for these details.
Licensing boundaries and responsibilities
The SAER team does not assert ownership of third-party satellite image copyrights, upstream model weights, or incorporated third-party code. The README and the component-level LICENSE file should be interpreted together. Distributions of code or models must preserve the upstream copyright and license notices applicable to their contents.
The manuscript and related figures may include third-party imagery under conditions narrower than those for research-generated labels; check original source terms before reusing such images. This section is intended to describe the current release accurately, not to claim that every experiment is reproducible or every file carries an unrestricted license.
Citation
For academic referencing, use the version-specific DOI displayed by Hugging Face for the dataset revision you used. DOI registration and revision history are managed by Hugging Face, so no particular DOI number is embedded in this README or in the repository's LICENSE file. This avoids introducing an outdated DOI when these documentation files are edited.
- Permanent repository URL: https://huggingface.co/datasets/XiaonianD/SAER-Research-Bundle
- Dataset title: SAER Mining Disturbance Research Bundle
- Citation metadata (creators, year and revision DOI): Use the official DOI record accessible from the dataset page.
- Associated manuscript: Sparsity-Aware Evidence Reconstruction for Regional Mining Disturbance Mapping (submitted to GIScience & Remote Sensing; not yet cited as a published article).
In a journal article, include the actual dataset DOI and the verified creator information in both the Data Availability Statement and the reference list. For reproducibility, you may additionally report the exact Hugging Face commit hash of the files used.
中文说明
本仓库为论文 Sparsity-Aware Evidence Reconstruction for Regional Mining Disturbance Mapping 的研究资料归档,包含京津冀地区通过 Google Earth Engine(GEE)获取和处理的 Copernicus Sentinel-2 遥感样本、团队自行制作的语义分割标签、困难负样本、训练与评估代码、三组 Full SAER 模型权重,以及 2017—2025 年矿业扰动制图成果。GEE 是影像访问与处理平台,影像使用遵循 Copernicus Sentinel 数据条款。
- 数据集链接:https://huggingface.co/datasets/XiaonianD/SAER-Research-Bundle
- 数据引用:请使用 Hugging Face 数据集页面显示的、与所用版本对应的 DOI;仓库说明文件不固定写入 DOI 编号。
- 当前仓库公开,可在无需登录或访问审批的情况下下载普通仓库文件。
- 五个 ZIP 文件各自完整,无需合并分片;应解压至同一父目录,并保留五个编号目录之间的相对关系。
- 正样本 PNG 标签中
255表示前景;年度预测 GeoTIFF 中255表示 NoData,不能混淆。 - 本归档未配置为可通过
load_dataset()直接读取的标准数据集,也不提供自动模型推理服务。 - 团队拥有授权权利的原创矿区分割标签、数据划分与年度矿区制图成果采用 CC BY 4.0 开放许可。Sentinel-2 影像遵循 Copernicus 条款并注明来源;团队有权授权的原创代码采用 MIT;NVIDIA SegFormer 相关模型权重遵循适用的上游许可(涉及非商业研究/评测及再分发条件);论文手稿及其他第三方材料不包含在上述许可范围内。详见仓库根目录的
LICENSE文件。 - 模型实验的具体复现范围与尚未收集的材料见上文及
REPRODUCTION.txt。
Documentation
Original packaging date: 2026-10-21.
README revision: 2026-10-27.
License notice: See LICENSE.
- Downloads last month
- 134