MADOS - Marine Debris and Oil Spill

Marine Debris and Oil Spill (MADOS) is a marine pollution dataset based on Sentinel-2 remote sensing data, focusing on marine litter and oil spills. Other sea surface features that coexist with or have been suggested to be spectrally similar to them have also been considered. MADOS formulates a challenging semantic segmentation task using sparse annotations. Citation: Kikaki K., Kakogeorgiou I., Hoteit I., Karantzalos K. Detecting Marine Pollutants and Sea Surface Features with Deep Learning in Sentinel-2 Imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 2024. For the implementation code, pre-trained models and exploratory analysis visit our project page https://marine-pollution.github.io/ MADOS Overview & Structure MADOS is structured in 174 scene folders, named Scene_0 through Scene_173 each corresponding to a unique Sentinel-2 (S2) scene. Each of these folders contains multiple image crops of the specific scene, indicated by the `_CROP` identifier. The total number of patch crops (240x240) is 2803. The next level of hierarchy seperates imagery by spatial resolution into three distinct folders: 10, 20 and 60 denoting 10m, 20m, and 60m resolution data, respectively. This allows to keep each S2 band at the initial resolution, supporting a wide range of applications (e.g., pansharpening). The 10 folder contains crops of S2 10 m resolution bands (`492` nm, `560` nm, `665` nm, `833` nm) with Rayleigh corrected reflectance values along with the corresponding masks of pixel-level annotations (_cl), confidence levels (_conf) and reports (_rep). RGB images (_rgb) and water turbidity outputs extracted by ACOLITE (_TUR_Dogliotti, _TUR_Nechad2016) are also provided. The 20 and 60 folders follow a similar pattern, containing cropped images of S2 bands relevant to their respective resolutions. Note that for the 20m resolution, we also provide aggregated mask annotations (_cl), confidence levels (_conf) and reports (_rep). At the root folder of MADOS, there is a `splits` folder containing three text files: train_X.txt, val_X.txt, and test_X.txt. These files describe the division of the dataset into training, validation and testing sets, respectively, crucial for structuring machine learning experiments to evaluate model performance. Folder structure └── MADOS ├── Scene_0 │ ├── 10 │ │ ├── Scene_0_L2R_rhorc_492_CROP.tif │ │ ├── Scene_0_L2R_rhorc_560_CROP.tif │ │ ├── Scene_0_L2R_rhorc_665_CROP.tif │ │ ├── Scene_0_L2R_rhorc_833_CROP.tif │ │ ├── Scene_0_L2R_cl_CROP.tif │ │ ├── Scene_0_L2R_conf_CROP.tif │ │ ├── Scene_0_L2R_rep_CROP.tif │ │ ├── Scene_0_L2R_rgb_CROP.png │ │ ├── Scene_0_L2W_TUR_Dogliotti_CROP.tif │ │ ├── Scene_0_L2W_TUR_Nechad2016_665_CROP.tif │ │ └── ... │ ├── 20 │ │ ├── Scene_0_L2R_rhorc_704_CROP.tif │ │ ├── Scene_0_L2R_rhorc_783_CROP.tif │ │ ├── Scene_0_L2R_rhorc_865_CROP.tif │ │ ├── Scene_0_L2R_rhorc_1614_CROP.tif │ │ ├── Scene_0_L2R_rhorc_2202_CROP.tif │ │ ├── Scene_0_L2R_cl_CROP.tif │ │ ├── Scene_0_L2R_conf_CROP.tif │ │ ├── Scene_0_L2R_rep_CROP.tif │ │ └── ... │ └── 60 │ ├── Scene_0_L2R_rhorc_443_CROP.tif │ └── ... ├── ... ├── Scene_173 │ ├── 10 │ │ └── ... │ ├── 20 │ │ └── ... │ └── 60 │ └── ... └── splits ├── test_X.txt ├── train_X.txt └── val_X.txt Mapping The mapping between Digital Numbers and Classes in _cl files is: 0. Non-annotated 1. Marine Debris 2. Dense Sargassum 3. Sparse Floating Algae 4. Natural Organic Material 5. Ship 6. Oil Spill 7. Marine Water 8. Sediment-Laden Water 9. Foam 10. Turbid Water 11. Shallow Water 12. Waves & Wakes 13. Oil Platform 14. Jellyfish 15. Sea snot The mapping between Digital Numbers and Confidence level in _conf files is: 1: High2: Moderate3: Low The mapping between Digital Numbers and marine debris Report existence in _rep files is: 1: Very close2: Away3: No The final uncompressed dataset requires 5.35 GB of storage.海洋垃圾与溢油（Marine Debris and Oil Spill, MADOS）是一款基于哨兵-2（Sentinel-2）遥感数据构建的海洋污染数据集，聚焦于海洋垃圾与溢油目标。研究同时纳入了与目标物共存或光谱特征疑似相似的其他海面特征，并基于稀疏标注构建了具有挑战性的语义分割任务。 引用文献：Kikaki K.、Kakogeorgiou I.、Hoteit I.、Karantzalos K.，《利用深度学习在哨兵-2影像中检测海洋污染物与海面特征》，ISPRS Journal of Photogrammetry and Remote Sensing，2024年。 若需获取实现代码、预训练模型与探索性分析内容，请访问项目主页：https://marine-pollution.github.io/ ### MADOS 概述与架构 MADOS 包含174个场景文件夹，命名为Scene_0至Scene_173，每个文件夹对应唯一的哨兵-2（S2）影像场景。每个文件夹下包含若干该场景的图像裁剪块，以`_CROP`标识符标注。所有尺寸为240×240像素的图像裁剪块总计2803份。 数据集架构按空间分辨率划分为三个独立文件夹：分别为10、20、60，对应10米、20米、60米分辨率的遥感数据，以此保留哨兵-2各波段的原始分辨率，可支撑包括全色锐化（pansharpening）在内的多种应用场景。 10米分辨率文件夹包含哨兵-2 10米分辨率波段的裁剪影像，对应波段为492 nm、560 nm、665 nm、833 nm，数据为经过瑞利校正的反射率值；同时提供对应的像素级标注掩码（`_cl`后缀）、置信度等级文件（`_conf`后缀）与报告文件（`_rep`后缀）。此外还包含RGB合成影像（`_rgb`后缀），以及通过ACOLITE工具提取的水体浊度输出结果（`_TUR_Dogliotti`、`_TUR_Nechad2016`后缀）。 20米与60米分辨率文件夹遵循类似架构，分别存放对应分辨率的哨兵-2波段裁剪影像。需注意，针对20米分辨率数据，我们额外提供了聚合后的掩码标注（`_cl`后缀）、置信度等级文件（`_conf`后缀）与报告文件（`_rep`后缀）。 在MADOS数据集的根目录下，存在一个`splits`文件夹，内含三个文本文件：train_X.txt、val_X.txt与test_X.txt，分别用于划分训练集、验证集与测试集，为构建机器学习实验以评估模型性能提供了关键支撑。 ### 文件夹架构 └── MADOS ├── Scene_0 │ ├── 10 │ │ ├── Scene_0_L2R_rhorc_492_CROP.tif │ │ ├── Scene_0_L2R_rhorc_560_CROP.tif │ │ ├── Scene_0_L2R_rhorc_665_CROP.tif │ │ ├── Scene_0_L2R_rhorc_833_CROP.tif │ │ ├── Scene_0_L2R_cl_CROP.tif │ │ ├── Scene_0_L2R_conf_CROP.tif │ │ ├── Scene_0_L2R_rep_CROP.tif │ │ ├── Scene_0_L2R_rgb_CROP.png │ │ ├── Scene_0_L2W_TUR_Dogliotti_CROP.tif │ │ ├── Scene_0_L2W_TUR_Nechad2016_665_CROP.tif │ │ └── ... │ ├── 20 │ │ ├── Scene_0_L2R_rhorc_704_CROP.tif │ │ ├── Scene_0_L2R_rhorc_783_CROP.tif │ │ ├── Scene_0_L2R_rhorc_865_CROP.tif │ │ ├── Scene_0_L2R_rhorc_1614_CROP.tif │ │ ├── Scene_0_L2R_rhorc_2202_CROP.tif │ │ ├── Scene_0_L2R_cl_CROP.tif │ │ ├── Scene_0_L2R_conf_CROP.tif │ │ ├── Scene_0_L2R_rep_CROP.tif │ │ └── ... │ └── 60 │ ├── Scene_0_L2R_rhorc_443_CROP.tif │ └── ... ├── ... ├── Scene_173 │ ├── 10 │ │ └── ... │ ├── 20 │ │ └── ... │ └── 60 │ └── ... └── splits ├── test_X.txt ├── train_X.txt └── val_X.txt ### 类别与数值映射 `_cl`文件中数字值与类别的映射关系如下： 0. 未标注区域 1. 海洋垃圾 2. 密集马尾藻 3. 稀疏漂浮藻类 4. 天然有机物质 5. 船舶 6. 溢油 7. 正常海水 8. 含沙水体 9. 泡沫 10. 浑浊水体 11. 浅水区 12. 波浪与尾迹 13. 石油平台 14. 水母 15. 海洋黏液（Sea snot） `_conf`文件中数字值与置信度等级的映射关系如下： 1: 高置信度 2: 中置信度 3: 低置信度 `_rep`文件中数字值与海洋垃圾存在性报告的映射关系如下： 1: 距离极近 2: 存在距离 3: 无相关报告 该数据集未压缩的最终版本总占用存储空间为5.35 GB。