Global ML-ready dataset for mining areas in satellite images

This dataset is a global resource for machine learning applications in mining area detection and semantic segmentation on satellite imagery. It contains Sentinel-2 satellite images and corresponding mining area masks + bounding boxes for 1,210 sites worldwide. Ground-truth masks are derived from Maus et al. (2022) and Tang et al. (2023), and validated through manual verification to ensure accurate alignment with Sentinel-2 imagery from specific timestamps. 

The dataset includes three mask variants:

Masks exclusively from Maus et al. (n=1,090)

Masks exclusively from Tang et al. (n=817)

A preferred mask selected from either Maus or Tang based on alignment quality determined during manual review (n=1,210).

Each tile corresponds to a 2048x2048 pixel Sentinel-2 image, with metadata on mine type (surface, placer, underground, brine & evaporation) and scale (artisanal, industrial). For convenience, the preferred mask dataset is already split into training (75%), validation (15%), and test (10%) sets. 

Furthermore, dataset quality was validated by re-validating test set tiles manually and correcting any mismatches between mining polygons and visually observed true mining area in the images, resulting in the following estimated quality metrics: 

Combined
Maus
Tang

Accuracy
99.78
99.74
99.83

Precision
99.22
99.20
99.24

Recall
95.71
96.34
95.10

Note that the dataset does not contain the Sentinel-2 images themselves but contains a reference to specific Sentinel-2 images. Thus, for any ML applications, the images must be persisted first. For example, Sentinel-2 imagery is available from Microsoft's Planetary Computer and filterable via STAC API: https://planetarycomputer.microsoft.com/dataset/sentinel-2-l2a. Additionally, the temporal specificity of the data allows integration with other imagery sources from the indicated timestamp, such as Landsat or other high-resolution imagery.

Source code used to generate this dataset and to use it for ML model training is available at https://github.com/SimonJasansky/mine-segmentation. It includes useful Python scripts, e.g. to download Sentinel-2 images via STAC API, or to divide tile images (2048x2048px) into smaller chips (e.g. 512x512px). 

A database schema, a schematic depiction of the dataset generation process, and a map of the global distribution of tiles are provided in the accompanying images.

本数据集为卫星影像矿区检测与语义分割领域的机器学习应用提供全球尺度的支撑资源。数据集包含全球范围内1210个点位的哨兵二号（Sentinel-2）卫星影像，以及对应的矿区掩膜与边界框。其真值掩膜源自Maus等人（2022）与Tang等人（2023）的研究成果，并经过人工核验，确保与特定时相的哨兵二号卫星影像精准匹配。

本数据集包含三类掩膜变体：
1. 仅源自Maus等人研究的掩膜（样本量n=1090）；
2. 仅源自Tang等人研究的掩膜（样本量n=817）；
3. 经人工审核后基于匹配质量从上述两类掩膜中优选出的掩膜（样本量n=1210）。

每个影像瓦片对应一张2048×2048像素的哨兵二号影像，附带矿山类型（露天矿、砂矿、地下矿、盐卤与蒸发矿）与开采规模（手工采矿、工业化采矿）的元数据。为便于使用，优选掩膜数据集已划分为训练集（75%）、验证集（15%）与测试集（10%）。

此外，本数据集通过人工重新核验测试集影像瓦片，修正矿区多边形与影像中目视解译的真实矿区之间的不匹配项，完成了数据集质量验证，最终得到如下评估质量指标：
| 评估指标 | 综合 | Maus组 | Tang组 |
|----------|------|--------|--------|
| 准确率 | 99.78 | 99.74 | 99.83 |
| 精确率 | 99.22 | 99.20 | 99.24 |
| 召回率 | 95.71 | 96.34 | 95.10 |

需注意，本数据集未直接存储哨兵二号影像本身，仅提供对应特定影像的索引信息。因此，在开展任何机器学习应用前，需先自行获取对应影像。例如，哨兵二号影像可从微软（Microsoft）行星计算机（Planetary Computer）获取，并可通过STAC API进行筛选：https://planetarycomputer.microsoft.com/dataset/sentinel-2-l2a。此外，本数据集具备时间特异性，可与对应时相的其他影像源（如陆地卫星（Landsat）影像或其他高分辨率影像）进行融合使用。

本数据集的生成代码及用于机器学习模型训练的配套代码已开源至https://github.com/SimonJasansky/mine-segmentation，其中包含实用的Python脚本，例如通过STAC API下载哨兵二号影像的脚本，以及将2048×2048像素的影像瓦片分割为更小的切片（如512×512像素）的脚本。

配套附件影像中包含了本数据集的数据库架构、数据集生成流程示意图以及影像瓦片全球分布地图。