On-board Anomaly Detection for Marine Environmental Monitoring — PhiSat-2 Database 2025

This folder contains data (images, metadata, and segmentation masks) used to train and test the marine anomaly detection application developed by the IRMA team at IRT Saint Exupéry for the PhiSat-2 OrbitalAI challenge.

Description

This dataset contains images generated from Sentinel-2 satellite data using the ESA simulator, designed to replicate the characteristics of the onboard PhiSat-2 sensor.

It was created to train and evaluate the AI pipeline for marine anomaly detection developed by the IRMA team at IRT Saint Exupéry. This pipeline is designed to detect anomalies (any event deviating from defined marine normality) in satellite images of the ocean, and to segment some of these anomalies (algae blooms, sediment floods, oil spills). Its training and evaluation required two types of data:

Images representing normal marine conditions;
Marine images containing various types of anomalies (algae, sediments, oil, plastic).

The dataset includes 26 images of normality and 24 images showing anomalies, along with their corresponding segmentation masks, manually annotated by IRMA.

Mask Label/Class Correspondence

The correspondence between labels and classes in the segmentation masks is as follows:

0 : NO DATA
1 : CLEAR WATER
2 : TURBID WATER
3 : LAND
4 : PLASTIC
5 : OIL
6 : ALGAE
7 : SEDIMENTS
8 : CLOUD

Metadata

Each image is associated with metadata. The following metadata fields are provided:

cloud_cover
crs
datatake
geolocation
sensor
sensor_resolution
spectral_bands_ordered
task

For more details about metadata, refer to this ReadMe.

Data Split

The 50 images in the dataset are split into two subsets: training + validation (TrainVal) and testing (Test).

Anomaly images are distributed as follows:

Algae: 4 images total — 3 in TrainVal, 1 in Test
Sediments: 6 images total — 5 in TrainVal, 1 in Test
Oil: 4 images total — 3 in TrainVal, 1 in Test
Plastic: 10 images total — 8 in TrainVal, 2 in Test

The PLASTIC, OIL, ALGAE, and SEDIMENTS classes are thus represented in both subsets.

Normality images were randomly split between TrainVal and Test using a 70% / 30% ratio. This split ensures that the CLEAR WATER, LAND, and CLOUD classes are represented in both subsets.

In total, TrainVal contains 36 samples and Test contains 14.

Usage Notes

The NO DATA class is used to indicate unannotated pixels; it is excluded from both training and evaluation.

The TURBID WATER class is used to characterize ambiguous ocean areas (uncertain whether clear water or anomaly). This class is excluded from training but may be considered during testing: a prediction will be considered correct if it predicts either clear water or an anomaly. We recommend using this class for visual inspection purposes only.

Note: For training and evaluating its AI pipeline, IRMA did not use any samples containing the PLASTIC class.

Structure

The database is organized as follows:

marine_area_dataset.zarr/
├── trainval/
│   ├── <sample_id>/
│   │   ├── img/            # Image array
│   │   ├── label/          # Label array or map
│   │   ├── .zattrs         # Metadata attributes
│   │   └── .zgroup         
│   ├── …
│   └── .zgroup 
├── test/
│   ├── <sample_id>/
│   │   ├── img/            # Image array
│   │   ├── label/          # Label array or map
│   │   ├── .zattrs         # Metadata attributes
│   │   └── .zgroup         
│   ├── …
│   └── .zgroup 
├── .zgroup                 
├── CHANGELOG.md
├── LICENSE    
└── README.md

The dataset uses the .zarr format, which explains the presence of .zgroup files. These indicate that the directories are valid Zarr groups and contain metadata about those groups.

Training data is stored in the trainval/ folder, and test data in the test/ folder. Each of these contains one subfolder per image; folders are numbered from 0000000/ to 0000035/ in trainval/, and from 0000000/ to 0000013/ in test/. Each subfolder contains the following:

img/ holds the image with shape (C, H, W) = (8, 4096, 4096) and ground resolution of 4.75 m;
label/ contains the segmentation mask with shape (1, H, W);
.zattrs contains metadata related to the image.

License

This dataset is licensed under the Creative Commons Attribution 3.0 IGO License (CC BY 3.0 IGO).

To view a copy of this license, visit: https://creativecommons.org/licenses/by/3.0/igo/legalcode.en

Datasets:

ESA-PhiLab-Edge
/

marine_area_dataset

You need to agree to share your contact information to access this dataset