MARIDA_海洋废弃物档案

MARIne碎片档案（MARIDA）是Sentinel-2卫星图像上面向海洋碎片的数据集。它还包括共存的各种海洋特征。MARIDA主要专注于弱监督像素级语义分割任务。
引文：Kikaki K，Kakogeorgiou I，Mikeli P，Raitsos DE，Karantzalos K（2022）MARIDA：从Sentinel-2遥感数据中检测海洋碎片的基准。《公共科学图书馆·综合》17（1）：e0262247。https://doi.org/10.1371/journal.pone.0262247
如需快速入门指南，请访问marine-destrips.github.io
数据集包含：
i.1381个补丁（256 x 256），由唯一日期和S2块构成。每个补丁都与像素级注释类（*_cl）和置信度（*_conf）的相应掩码一起提供。补丁以GeoTiff格式给出。
ii。WGS'84/UTM投影中的Shapefiles数据，文件命名约定遵循以下方案：s2_dd-mm-yy_ttt，其中s2表示s2传感器，dd表示日，mm表示月，yy表示年，ttt表示s2图块。形状文件包括每个注释的类别以及置信度和海洋碎片报告描述。
iii.用于评估机器学习算法的训练、验证和测试分割。
iv.为每个补丁分配的多个标签（labels_mapping.txt）。
数字和类之间的映射是：
1：海洋废弃物
2：稠密的马尾藻
3：稀疏的马尾藻
4：天然有机材料
5：船
6：云
7：海水
8：含沙水
9：泡沫
10：浑浊的水
11：浅水
12：波浪
13：云影
14：醒来
15：混合水
数字和置信度之间的映射是：
1：高
2：中等
3：低
数字编号和海洋废弃物报告存在之间的映射是：
1：非常接近
2：远离
3：没有
最终的未压缩数据集需要4.38 GB的存储空间。。

Marine Debris Archive (MARIDA) is a dataset focused on marine debris derived from Sentinel-2 satellite imagery, which also encompasses various co-occurring marine features. MARIDA primarily targets weakly-supervised pixel-level semantic segmentation tasks.

Citation: Kikaki K, Kakogeorgiou I, Mikeli P, Raitsos DE, Karantzalos K (2022) MARIDA: A benchmark for marine debris detection from Sentinel-2 remote sensing data. PLOS ONE 17(1): e0262247. https://doi.org/10.1371/journal.pone.0262247

For a quick start guide, please visit marine-destrips.github.io

The dataset includes:
i. 1381 patches (256 × 256) constructed from unique dates and Sentinel-2 tiles. Each patch is paired with corresponding masks for pixel-level annotation classes (*_cl) and confidence scores (*_conf). All patches are provided in GeoTIFF format.
ii. Shapefile data in WGS'84/UTM projection, with a file naming convention following the schema: s2_dd-mm-yy_ttt, where s2 indicates the Sentinel-2 sensor, dd denotes the day, mm the month, yy the year, and ttt the Sentinel-2 tile identifier. The shapefiles contain the category of each annotation, along with confidence scores and descriptive reports of marine debris.
iii. Training, validation, and test splits for evaluating machine learning algorithms.
iv. Multiple labels assigned to each patch (stored in labels_mapping.txt).

The mapping between numerical IDs and object classes is as follows:
1: Marine debris
2: Dense sargassum
3: Sparse sargassum
4: Natural organic material
5: Vessel
6: Cloud
7: Seawater
8: Sediment-laden water
9: Foam
10: Turbid water
11: Shallow water
12: Wave
13: Cloud shadow
14: Wake
15: Mixed water

The mapping between numerical IDs and confidence levels is:
1: High
2: Medium
3: Low

The mapping between numerical IDs and the proximity of reported marine debris is:
1: Very close
2: Far away
3: None

The final uncompressed dataset requires 4.38 GB of storage space.