SARLAND
收藏资源简介:
SARLAND是一个多模态遥感分析数据集,结合了高分辨率合成孔径雷达(SAR)图像与精确对齐的土地覆盖信息。它复用ONERA/SARLO-80的SAR样本(包括约80厘米分辨率的复杂UMBRA SAR裁剪图、SICD元数据和描述),并为每个样本新增了来自哥白尼“10米分辨率土地覆盖图 - 年度V1版”(LCM-10)产品的10米土地覆盖层,该层经过重投影以精确叠加在SAR裁剪图上。数据以WebDataset格式组织,包含训练集分块和分片,每个样本包括复数SAR数据(.sar.npy)、SAR幅度图像(.sar.png)、SICD元数据(.sicd.xml)、元数据文件(.meta.json)以及两种形式的土地覆盖数据:一个包含标签和掩码的uint8数组(.landcover.npy),其中标签是每个像素的LCM-10类别代码,掩码指示像素有效性;以及一个彩色化的RGBA图像(.landcover.png),其Alpha通道为有效性掩码。土地覆盖数据基于FAO土地覆盖分类系统,包含11个主要类别,如树木覆盖、灌木地、草地、耕地、建筑区、裸地/稀疏植被、冰雪、永久性水体、草本湿地、红树林以及苔藓和地衣。该数据集适用于图像分割(特别是土地覆盖分类)和图像分类等计算机视觉任务,旨在促进SAR光学数据融合、土地覆盖监测和变化检测等遥感应用的研究与开发,规模在10K到100K样本之间,并遵循复合许可证(土地覆盖数据使用CC BY 4.0,SAR数据遵循UMBRA使用条款及ONERA/SARLO-80规定)。
SARLAND is a multimodal dataset for remote sensing analysis, combining high-resolution synthetic aperture radar (SAR) images with precisely aligned land cover information. It reuses SAR samples from ONERA/SARLO-80 (including complex UMBRA SAR crops at approximately 80 cm resolution, SICD metadata, and descriptions) and adds a 10-meter land cover layer from the Copernicus 10m Resolution Land Cover Map - Annual V1 (LCM-10) product for each sample. This land cover layer is reprojected using an optical-to-SAR transformation derived from SICD, enabling precise overlay on SAR crops. The data is organized in WebDataset format, with training set chunks and shards. Each sample (identified by an 8-digit index) includes the following files: original complex SAR data (.sar.npy), SAR magnitude image (.sar.png), SICD metadata (.sicd.xml), metadata file with crop, projection, description, and incidence angle information (.meta.json), and the newly added land cover data aligned to the SAR frame. Land cover data is provided in two forms: a uint8 array containing labels and masks (.landcover.npy), where labels are LCM-10 category codes per pixel and masks indicate pixel validity; and a colorized RGBA image (.landcover.png) with an Alpha channel as the validity mask. Land cover data is resampled to the SAR grid using nearest-neighbor affine transformation to maintain sharp category boundaries. The land cover data originates from the Copernicus CLMS LCM-10 product (2020 baseline year), following the FAO land cover classification system with 11 main categories, such as tree cover, shrubland, grassland, cropland, built-up areas, bare/sparse vegetation, snow/ice, permanent water bodies, herbaceous wetlands, mangroves, and moss/lichens. This dataset is suitable for computer vision tasks like image segmentation (particularly land cover classification) and image classification, aiming to advance research and development in remote sensing applications such as SAR-optical data fusion, land cover monitoring, and change detection. The dataset scale falls into the 10K<n<100K category. Usage of this dataset requires compliance with its composite license: land cover data follows the CC BY 4.0 license, while SAR data must adhere to UMBRA terms of use and the regulations of the source dataset ONERA/SARLO-80.
SARLAND 数据集概述
基本信息
- 数据集名称: SARLAND (SARLO-80 SAR crops + ESA WorldCover 10 m land cover, aligned to the SAR frame)
- 许可证: other(包含多个组件的独立许可)
- 语言: 英语
- 任务类别: 图像分割、图像分类
- 标签: SAR、合成孔径雷达、遥感、土地覆盖、Umbra、Copernicus、ESA WorldCover、WebDataset
- 数据集规模: 10K 到 100K 样本之间
数据集构成
SARLAND 数据集复用了 ONERA/SARLO-80 的 SAR 样本(复杂 UMBRA SAR 裁剪图,~80 cm 分辨率,含 SICD 元数据和描述),并为每个样本额外添加了一层 10 m 分辨率的土地覆盖层,该层通过光学到 SAR 的变换(基于 SICD 推导)重新投影,使其与 SAR 裁剪图逐像素对齐。
土地覆盖类别
使用 11 个 ESA WorldCover 类别代码:
| 代码 | 类别 | RGB 颜色 |
|---|---|---|
| 10 | 树木覆盖 | (0, 100, 0) |
| 20 | 灌木地 | (255, 187, 34) |
| 30 | 草地 | (255, 255, 76) |
| 40 | 农田 | (240, 150, 255) |
| 50 | 建成区 | (250, 0, 0) |
| 60 | 裸地/稀疏植被 | (180, 180, 180) |
| 70 | 冰雪覆盖 | (240, 240, 240) |
| 80 | 永久水体 | (0, 100, 200) |
| 90 | 草本湿地 | (0, 150, 160) |
| 95 | 红树林 | (0, 207, 117) |
| 100 | 苔藓和地衣 | (250, 230, 160) |
数据集结构
采用 WebDataset 格式,按块(chunk)和分片(shard)组织:
train/ chunk_000/ shard-00000.tar shard-00001.tar ... chunk_001/ ...
每个样本包含的文件
每个样本共享键 {id}(8 位数字索引),包含以下成员:
| 成员 | 内容 |
|---|---|
{id}.sar.npy |
复杂 SAR(SLC),与 SARLO-80 相同 |
{id}.sar.png |
SAR 幅度(斜距),uint8 格式 |
{id}.sicd.xml |
SICD 元数据(采集几何信息) |
{id}.meta.json |
元数据(裁剪、投影、描述、入射角等) |
{id}.landcover.npy |
新增 (2, H, W) uint8 = [label, mask],与 SAR 帧对齐 |
{id}.landcover.png |
新增 彩色土地覆盖图,RGBA 格式(alpha = 有效性掩码) |
label: 逐像素 ESA WorldCover 类别代码,位于 SAR 裁剪图像素网格上(~80 cm)mask: 1 = 有效土地覆盖像素,0 = 无效/超出覆盖范围- 土地覆盖通过仿射变换(
INTER_NEAREST)重采样到 SAR 网格,保持清晰的类别边界
加载示例
python import io, json import numpy as np from PIL import Image from huggingface_hub import hf_hub_download import webdataset as wds
local_tar = hf_hub_download( repo_id="SoleneDEBUYSERE/SARLAND", repo_type="dataset", filename="train/chunk_000/shard-00000.tar", )
ds = wds.WebDataset(local_tar, shardshuffle=False) sample = next(iter(ds))
sar_complex = np.load(io.BytesIO(sample["sar.npy"]), allow_pickle=False) sar_amp = np.asarray(Image.open(io.BytesIO(sample["sar.png"])).convert("L")) meta = json.loads(sample["meta.json"].decode("utf-8"))
Land cover aligned to the SAR frame
lc = np.load(io.BytesIO(sample["landcover.npy"]), allow_pickle=False) # (2, H, W) label, mask = lc[0], lc[1] lc_rgba = np.asarray(Image.open(io.BytesIO(sample["landcover.png"])).convert("RGBA"))
print("SAR amplitude:", sar_amp.shape) print("Land cover :", label.shape, "classes:", np.unique(label[mask > 0]))
许可与引用
数据集各组件具有不同的许可:
- 土地覆盖(ESA WorldCover 10 m, 2020 v100):© ESA WorldCover 项目,由 VITO 牵头的联盟制作。采用 CC BY 4.0 许可。请引用 ESA WorldCover 产品。
- SAR:图像源自 UMBRA 数据,请参考 UMBRA 使用条款及源数据集
ONERA/SARLO-80。
使用本数据集时,请同时引用 ESA WorldCover、UMBRA 数据源、源数据集 ONERA/SARLO-80 以及本数据集。在 SARLAND 相关论文发布前,请引用 ONERA/SARLO-80 作为 SAR 图像的参考数据集。
致谢
作者感谢 DEMR-ONERA 为本数据集生成提供计算资源,特别感谢 Nicolas Trouvé 及其团队的支持。




