OpenSAR Insight

SAR imagery

This dataset uses Synthetic Aperture Radar (SAR) images from the European Space Agency's (ESA) Copernicus Sentinel-1 mission, which come from two satellites (Sentinel-1A and Sentinel-1B) in polar orbit that operate in all weather conditions, both day and night. The selected Level 1 SAR products are Ground Range Detection (GRD) in wide interferometric mode (IW) and Single Look Complex (SLC) burst products.

The two available polarisation channels (VV and VH) have been exploited from these data, taking into account that VH (vertical-horizontal) polarisation is usually more conducive to ship detection. This is because it provides greater contrast between vessels and marine clutter. In contrast, VV (vertical-vertical) polarisation provides more information about the characteristics of the sea surface.

An important new feature of this dataset is the availability of Level 0 data (L0, RAW data) extracted from the Level 1 data (L1) used.

SAR images location

The SAR data composing the database have the following geographical coverage:

SAR Data & labeling

The dataset is structured as a series of SAR scenes. In this case study, a total of 15 different SAR scenes were selected. For each Level 1 SAR product (GRD-SLC), the corresponding Level 2 Ocean products (OCN) were retrieved, allowing additional information on wind conditions over the sea surface to be included. This information has been used to complement that provided by the AIS (Automatic Identification System) data available in the xView3 reference database.

All SAR L1B products employed in this dataset were preprocessed to enable block-wise partitioning. The partitioning algorithm operates by segmenting the complex image (SLC) into non-overlapping blocks of 512 × 512 pixels, thereby producing square patches suitable for subsequent training and validation tasks. Corresponding patches were also extracted from the GRD products, ensuring spatial co-registration and alignment with the SLC-derived patches over the same geographic area of interest. Consequently, a total of four co-registered patches were obtained for each selected block region.

The labelling strategy is based on XML (Extensible Markup Language) format. In the next table are detailed the considered tags, levels, and descendants for all Dark Vessel Detection files (including metadata, data, and labels).

Level	Tag Name	Descendants	Tag Contents
0	Scene_Info	Scene_ID; Date; Version; CaseStudy	-
1	Scene_ID	N/A	Naming of the SAR scene into the database
1	Date	N/A	YYYYMMDD
1	Version	N/A	Version of the SAR scene
1	CaseStudy	N/A	DarkVesselDetection
0	SARData	SAR_Mission; SAR product; SLCSwath; Time_interval	-
1	SAR_Mission	N/A	S1A-S1B
1	SARProduct	N/A	Name of the SLC product SAR
1	SLCSwath	N/A	Number of swath
1	Time_Interval	Start; Stop	-
2	Start	N/A	YYYYMMDD_HHMMSS
2	Stop	N/A	YYYYMMDD_HHMMSS
0	ProcessingData	Corner_Coord; WindData; StatisticsReport; List_of_ships	-
1	Corner_Coord	SARData_Sample; SARData_Line; Scene_Sample; Scene_Line; Latitude; Longitude	-
2	SARData_Sample	N/A	Number of columns of 512 pixels in the geocoded image
2	SARData_Line	N/A	Number of rows of 512 pixels in the geocoded image
2	Scene_Sample	N/A
2	Scene_Line	N/A
2	Latitude	N/A	Latitudes of corners
2	Longitude	N/A	Longitudes of corners
1	WindData	Direction; Speed	-
2	Direction	N/A	Wind direction over the patch
2	Speed	N/A	Speed direction over the patch
1	StatisticsReport	Number_of_ships	-
2	Number_of_ships	N/A	Number of detected ships in the SAR image
1	List_of_ships	Ship	-
2	Ship	Name; Centroid_Position; Size; BoundingBox	-
3	Name	N/A	String with the ID of the ship
3	Centroid_Position	Latitude; Longitude; SARData_Sample; SARData_Line; Scene_Sample; Scene_Line	-
4	Latitude	N/A	Latitude coordinate of the ship centroid
4	Longitude	N/A	Longitude coordinate of the ship centroid
4	SARData_Sample	N/A	Column coordinate of the ship centroid (patch)
4	SARData_Line	N/A	Row coordinate of the ship centroid (patch)
4	Scene_Sample	N/A	Column coordinate of the ship centroid (scene)
4	Scene_Line	N/A	Row coordinate of the ship centroid (scene)
3	Size	N/A	Ship length size in m
3	BoundingBox	Top; Left; Bottom; Right	-
4	Top	N/A	Top row of the detected pixels
4	Left	N/A	Left column of the detected pixels
4	Bottom	N/A	Bottom row of the detected pixels
4	Right	N/A	Right column of the detected pixels

Data partitions

The data relating to this case study are organised as follows:

Vessel use-case	XML labels	L0 patches	SLC patches	GRD patches
Train set	2671	5342	5342	5342
Val set	311	622	622	622
Test set	1065	2130	2130	2130
Total patches	4047	8094	8094	8094

Note: XMLs denote the number of unique patches. L0, SLC, and GRD are doubled as they include both VV and VH polarisations.