Sentinel 2 Ship_detection

## Overview

Sentinel 2 Ship_detection is a dataset for object detection tasks - it contains Ship annotations for 739 images.

## Getting Started

You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.

  ## License

  This dataset is available under the [BY-NC-SA 4.0 license](https://creativecommons.org/licenses/BY-NC-SA 4.0).

This database was generated by AGENIUM Space in the framework of the CORTEX project (https://esacortexproject.agenium-space.com/) funded by ESA.

The database was created using Sentinel-2 images distributed through the Copernicus open access hub (https://www.copernicus.eu/en, https://scihub.copernicus.eu/) and AIS (Automatic Identification System) data. Sentinel-2 images are all L1C products acquired in Danish sovereign waters in 2019. Danish government made available the AIS (Automatic Identification System) data around Denmark from 2009 until now ( https://www.dma.dk/SikkerhedTilSoes/Sejladsinformation/AIS/Sider/default.aspx ). More specifically, 14 tiles were selected, each of them with a cloud coverage below 10% according to the cloud mask products.

Three DBs are provided. Their description is given in S2-Ships-DB-description.pdf document attached to the DB.

This work is funded by a contract in the framework of the EO SCIENCE FOR SOCIETY PERMANENTLY OPEN CALL FOR PROPOSALS EOEP-5 BLOCK 4 issued by the European Space Agency.

This database comprises 26 Sentinel-2 images, totaling 258 ship exemplars. The images are generated using vessel detection reports provided by analysts from Collecte Localisation Satellites (CLS). Each Sentinel-2 image is accompanied by its land mask and a CSV file that includes the position of detected vessels along with their characteristics, such as length and heading.

xS2Wakes: A dataset for xAI of Wakes in S-2 (L2A).

Summary

The dataset is derived from Sentinel-2 Level-2A (L2A) satellite images and focuses on the marine domain over Danish fjords. It provides a comprehensive collection of ship wakes and background clutter (referred to as "no_wake_crop") for remote sensing applications. The dataset has undergone post-processing through the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm with a clip limit value of 0.12 and a tile size of 16x16. The dataset comprises four spectral bands: B2, B3, B4, and B8.

Importance and Relevance to Remote Sensing Community

Multifaceted Applications of Wake Detection

Ship wake detection serves as a cornerstone in a multitude of domains that are critical to both human and environmental well-being:

Navigational Safety: Understanding ship wakes can provide insights into water currents and traffic patterns. This is vital for ensuring the safe passage of marine vessels, particularly in narrow straits and busy ports.

Environmental Monitoring: The study of ship wakes can reveal the influence of vessels on aquatic ecosystems. For instance, excessive wake turbulence can lead to coastal erosion and can disrupt marine habitats.

Maritime Surveillance: Wake detection plays a crucial role in maintaining maritime security. Tracking the wakes of vessels can help in identifying illegal activities such as smuggling or unauthorized fishing.

Specifications

• Data Source: Sentinel-2 L2A
• Region of Interest: Danish fjords
• Classes: Wake, No-Wake
• Number of Samples:
  • Wake: 123
  • No-Wake: 150
• Spectral Bands: B2 (Blue), B3 (Green), B4 (Red), B8 (NIR)
• Post-Processing: CLAHE (Clip Limit = 0.12, Tile Size = 16x16)
• Average Wake Chip Size: 390x351 pixels
• Average No-Wake Chip Size: 380x390 pixels

Wake Detection and Analysis

Traditional Methods and Their Limitations

Traditionally, the process of ship wake detection has largely been a manual endeavor or employed simplistic statistical algorithms. Analysts would sift through satellite or aerial images to identify ship wakes, a process that is both time-consuming and prone to human error. Even automated statistical methods often lack the robustness needed to differentiate between true wakes and false positives, such as aquatic plants or natural water disturbances.

Role of xAI (Explainable AI) in Wake Identification

The introduction of explainable AI (xAI) techniques brings another layer of sophistication to wake analysis. While traditional machine learning models may offer high performance, they often act as "black boxes," making it difficult to understand how they arrive at a certain conclusion. In a critical domain like navigational safety or maritime surveillance, the ability to interpret and understand model decisions is indispensable. xAI methods can make these machine learning models more transparent, providing insights into their decision-making processes, which in turn can aid in fine-tuning or fully trusting the models.

Spectral Bands Selected

The inclusion of four key spectral bands—B2, B3, B4, and B8—offers the scope for multi-spectral analysis. Different bands can capture varying features of water and wake textures, thereby offering a richer feature set for machine learning models. We use these spectral bands as referred to in [Liu, Yingfei, Jun Zhao, and Yan Qin. "A novel technique for ship wake detection from optical images." Remote Sensing of Environment 258 (2021): 112375.]

Understanding Optical vs. SAR Imaging Modalities

It is important to note the fundamental differences between wakes captured in Synthetic Aperture Radar (SAR) images and those in optical imagery. In SAR images, narrow-V wakes often arise due to Bragg scattering, a phenomenon that does not exist at optical wavelengths. In optical images, bright lines close to turbulent wakes are actually foams generated by the interaction between the surface horizontal flow of turbulent wakes and the surrounding background waves (Ermakov et al., 2014; Milgram et al., 1993; Peltzer et al., 1992). This can make the detection of wakes in optical images more challenging as there are usually no bright lines near turbulent wakes, and Kelvin arms may also show dark contrast. Methods that solely rely on searching for a trough and peak pair, taking the trough as the turbulent wake, would miss many actual wakes and could also result in the identification of false wakes.

Contrast Enhancement

The application of the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm to this dataset allows for enhanced local contrast, enabling subtle features to become more pronounced. This significantly aids machine learning algorithms in feature extraction, thereby improving their ability to distinguish between complex patterns.

Environment and Clutter Assessment

In addition to wakes, the dataset contains samples labeled as "No-Wake," which include environmental clutter and clouds. These samples are crucial for training robust models that can differentiate wakes from similar-looking natural phenomena.

This dataset contains annotated marine vessels from 15 different Sentinel-2 product, used for training object detection models for marine vessel detection. The vessels are annotated as bounding boxes, covering also some amount of the wake if present.

Source data

from geo2ml.scripts.data import create_coco_dataset
raster_path = '

The dataset used in the study consists of imagery capturing ship wake patterns. It is a manually curated dataset specifically created for the purpose of training and evaluating the wake component detection model. The dataset contains a collection of image chips, each focusing on a specific ship wake instance.

The imagery in the dataset is acquired from satellite sensors, specifically on Sentinel-2 satellite imagery. Sentinel-2 provides multispectral data with high spatial resolution, allowing for detailed analysis of ship wake patterns. The dataset includes images captured on B8 spectral band, enabling the exploration of the wake detection model's performance under various spectral conditions. These images have been pre-processed (by scaling+CLAHE) to highlight ocean surface features.

Each image chip in the dataset is annotated with keypoint locations representing specific wake components, such as the ship wake vertex, the ending of the turbulent wake, and the ending of Kelvin arms. These annotations serve as ground truth labels for training and evaluating the wake component detection model.

Additionally, the dataset includes samples with variations in environmental conditions, such as different sea states, lighting conditions, and wake complexities. This variability allows for a comprehensive evaluation of the model's generalization capability and robustness across diverse scenarios.

###the following abstract is also provided as a file using proper formatting###

Abstract: Ship wakes observed by Synthetic Aperture Radar augmented by manually retraced wake components

Satellite-based Synthetic Aperture Radar (SAR) sensors offer the opportunity to observe the maritime domain, even during nighttime and under foggy or cloudy weather conditions. Depending on the nature of oceanographic observations, gaining information of position and movement of maritime objects is an essential element. Radar signatures of man-made maritime objects typically have extents of up to a few hundred meters. However, the transit of a moving ship can affect the ocean surface up to hundreds of kilometers creating large scale artefacts in SAR images, the so-called wake signatures. The published data is focused on the observation of moving ships by exploiting those wake signatures imaged by the SAR sensors.

The appearance of ship wakes in SAR imagery has been investigated for decades. Radar signatures of ship wakes are complex structures consisting of multiple wake components. Those wake components appear with different shapes and extents in SAR acquisitions, depending on various influencing parameters describing the present situation during the observation. Those influencing parameters are categorized into three types: ship properties, environmental conditions and image acquisition parameters.

Recently, the characteristic effect of the influencing parameters on the detectability of ship wakes has been modelled and systematically analyzed for the first time on the basis of this dataset, now available to the public. The results are published in the following journal publications [1, 2, 3, 4, 5, 6, 7] and all-encompassing in the following dissertation’s monography [8]. The published dataset has also been applied to develop the first Deep-Learning-based detector for individual wake components in SAR imagery [9].

This published dataset offers the following unique features:

1. This dataset contains extracted image patches of SAR acquisitions from the SAR missions TerraSAR X, CosmoSkymed, Sentinel 1 and RADARSAT 2 in tiff file format. The X-band and C-band radar frequencies of the SAR sensors operated by those four missions are an ideal choice for indirect detection of ships on the ocean surface. The acquisitions were taken in the years 2013 to 2018 over North Sea, Baltic Sea and Mediterranean Sea.
2. Each image patch contains the position of a moving ship, i.e. a candidate wake sample, with 5.1 km x 5.1 km extent and pixel spacing of 1.5 m. Each image patch is complemented with metadata information and information on influencing parameters in ods file format:
   1. ship properties, derived CFAR detection algorithm [10] and from data of the Automatic Identification System (AIS) [11],
   2. environmental conditions, estimated using SAR-SeaStaR’s empirical model functions for wind and sea state parameter retrieval [12, 13, 14, 15] as well as Weather Research and Forecasting Model (WRF) [16], and
   3. image acquisitions parameters, extracted from the SAR product’s metadata.
3. All candidate wake samples have been manually inspected by two different experts in the field of SAR oceanography. All look-a-likes of wake signatures have been filtered out. Position of individual wake components have been retraced, in case of wake components with curved characteristics, i.e. near-hull turbulence, turbulent wakes, Kelvin wake arms, V-narrow wake arms and ship-generated internal waves, or flagged, in case of wake components with oscillating characteristics, i.e. transverse waves and divergent waves. Retracing information is made available in csv file format and python source code for interpretation of all files is also delivered in the package.

The publication of this dataset shall enable users to,

• reproduce the wake detection methods or modelling and systematical analysis of wake detectability developed and published by the authors [1, 2, 3, 4, 5, 6, 7, 8], and
• develop their own methods for recognition of ship wakes in SAR imagery.

Acknowledgments

• Data provided by the European Space Agency.
• Includes material from COSMO-SkyMed satellite image © ASI (2018 & 2019), provided by e-GEOS, all rights reserved. Please note: the extracts of CosmoSkymed (CSK) images exceed the maximum dimensions allowed by e-GEOS for data publication by 24 pixels in width and height dimension, respectively (i.e. 1024x1024 pixels instead of 1000x1000 pixels), as restricted in ESA's TPM terms and conditions. E-GEOS has given their written consent to the publishing authors that the extracts from CSK images can be published in their current form.
• RADARSAT is an official mark of the Canadian Space Agency. RADARSAT-2 Data and Products @ MDA Geospatial Services Inc. (2013 to 2019) — All Rights Reserved
• Contains modified Copernicus Sentinel data 2015.
• TerraSAR-X/TanDEM-Y data © DLR <2013 to 2017>

References

[1] B. Tings and D. Velotto, "Comparison of ship wake detectability on C-band and X-band SAR," International Journal of Remote Sensing, vol. 39, no. 13, pp. 1-18, 2018, doi: 10.1080/01431161.2018.1425568.
[2] B. Tings, C. Bentes, D. Velotto and S. Voinov, "Modelling Ship Detectability Depending On TerraSAR-X-derived Metocean Parameters," CEAS Space Journal, vol. 11, p. 81–94, 2018, doi: 10.1007/s12567-018-0222-8.
[3] B. Tings, A. Pleskachevsky, D. Velotto and S. Jacobsen, "Extension of Ship Wake Detectability Model for Non-Linear Influences of Parameters Using Satellite Based X-Band Synthetic Aperture Radar," Remote Sensing, vol. 11, no. 5, pp. 1-20, 2019, doi: 10.3390/rs11050563.
[4] B. Tings, S. Jacobsen, S. Wiehle, E. Schwarz and H. Daedelow, "X-Band/C-Band-Comparison of Ship Wake Detectability," in EUSAR-Preprints 2020, Leipzig, 2020, doi: 10.20944/preprints202012.0480.v1.
[5] B. Tings, S. Wiehle and S. Jacobsen, "Ship wake component detectability on synthetic aperture radar (SAR)," in IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, Waikoloa, 2020, doi: 10.1109/IGARSS39084.2020.9323097.
[6] B. Tings, "Non-Linear Modeling of Detectability of Ship Wake Components in Dependency to Influencing Parameters Using Spaceborne X-Band SAR," Remote Sensing, vol. 13, no. 2, p. 165, 2021, doi: 10.3390/rs13020165.
[7] B. Tings, A. Pleskachevsky and S. Wiehle, "Comparison of detectability of ship wake components between C-Band and X-Band synthetic aperture radar sensors operating under different slant ranges," ISPRS Journal of Photogrammetry and Remote Sensing, vol. 196, pp. 306-324, 2023, doi: 10.1016/j.isprsjprs.2022.12.008 (corrigendum 10.1016/j.isprsjprs.2025.01.026).
[8] B. Tings, „Dissertation: Erkennung der Bug- und Heckwellen von Schiffen durch satellitenbasierte C-Band- und X-Band-Radarsensoren mit synthetischer Apertur,“ Helmut-Schmidt-Universität, Hamburg, 2024.
[9] B. Tings, Y.-J. Yang, C. Schnupfhagn and S. Jacobsen, "Tuning Detection of Ship Wakes by Detectability Modelling," 4th European Workshop on Maritime Systems, Resilience and Security 2024 (MARESEC 24), Bremerhaven, 2024, doi: 10.5281/zenodo.14524265.
[10] B. Tings, C. Bentes and S. Lehner, "Dynamically adapted ship parameter estimation using TerraSAR-X images," International Journal of Remote Sensing, pp. 1990-2015, 2016, doi: 10.1080/01431161.2015.1071898.
[11] B. J. Tetreault, "Use of the Automatic Identification System (AIS) for maritime domain awareness (MDA)," Proceedings of OCEANS 2005 MTS/IEEE, vol. 2, pp. 1590-1594, 2005, doi: 10.1109/OCEANS.2005.1639983.
[12] A. Pleskachevsky, B. Tings, S. Jacobsen, S. Wiehle, E. Schwarz and D. Krause, "A System for Near Real Time Monitoring of the Sea State using SAR Satellites," IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-18, 2024, doi: 10.1109/TGRS.2024.3419582.
[13] X.-M. Li and S. Lehner, "Algorithm for Sea Surface Wind Retrieval From TerraSAR-X and TanDEM-X Data," IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 5, pp. 2928-2939, 2014, doi: 10.1109/TGRS.2013.2267780.
[14] S. Jacobsen, X. Li, S. Lehner, J. Hieronimus and J. Schneemann, "Joint Offshore Wind Field Monitoring with Spaceborne SAR and Platform-Based Doppler LiDAR Measurements," International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. 40, p. 959–966, 2015, doi: 10.5194/isprsarchives-XL-7-W3-959-2015.
[15] F. Monaldo, C. Jackson, X. Li and W. G. Pichel, "Preliminary Evaluation of Sentinel-1A Wind Speed Retrievals," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 9, no. 6, pp. 2638-2642, 2016, doi: 10.1109/JSTARS.2015.2504324.
[16] W. C. Skamarock, J. B. Klemp, J. Dudhia, D. O. Gill, D. M. Barker, M. G. Duda, X.-Y. Huang, W. Wang and J. G. Powers, "A Description of the Advanced Research WRF Version 3," NCAR Technical Notes, Boulder, 2008, doi: 10.5065/D68S4MVH.

MARIDA (Marine Debris Archive) is the first dataset based on the multispectral Sentinel-2 (S2) satellite data, which distinguishes Marine Debris from various marine features that co-exist, including Sargassum macroalgae, Ships, Natural Organic Material, Waves, Wakes, Foam, dissimilar water types (i.e., Clear, Turbid Water, Sediment-Laden Water, Shallow Water), and Clouds. MARIDA is an open-access dataset which enables the research community to explore the spectral behaviour of certain floating materials, sea state features and water types, to develop and evaluate Marine Debris detection solutions based on artificial intelligence and deep learning architectures, as well as satellite pre-processing pipelines. Although it is designed to be beneficial for several machine learning tasks, it primarily aims to benchmark weakly supervised pixel-level semantic segmentation learning methods.

MARIDA can be downloaded from the repository Zenodo (https://doi.org/10.5281/zenodo.5151941). A quick start guide for all ML benchmarks and the detailed overview of the dataset are available at https://marine-debris.github.io/.

The Multi-Spectral Ship Wake Dataset (MSSWD) is a dataset designed for ship wake detection in multi-spectral satellite imagery. It is structured as follows:

• Source: 661 image chips derived from 50 Sentinel-2 images, captured by the Multi-Spectral Instrument (MSI) at 10-meter resolution across the visible, near-infrared (VNIR), and short-wave infrared (SWIR) spectral bands. The chips come already pre-processed to highlight sea surface features by using a Contrast Limited Adaptive Histogram Equalization (CLAHE) technique. - Content: The dataset includes 1059 ship wakes, with various configurations such as: - Single ship wakes - Multiple ship wakes - False wakes (e.g., airplane wakes, sea crests) - Sea clutter with no visible wakes

• Wake Characteristics: Diverse patterns of ship wakes are captured, including: - Vertical, horizontal, and tilted wakes - Cluttered sea scenes - Partial occlusions due to cloud cover

• Data Quality: Focused on quality over quantity, MSSWD reflects real-world complexity by collecting data in congested, crowded maritime environments.

• Data Labelling: Manually annotated using polygonal annotations to delineate wake contours, which allows: - Instance segmentation - Enhanced refinement during data augmentation

MARIne Debris Archive (MARIDA) is a marine debris-oriented dataset on Sentinel-2 satellite images. It also includes various sea features that co-exist. MARIDA is primarily focused on the weakly supervised pixel-level semantic segmentation task.

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 the quick start guide visit marine-debris.github.io

The dataset contains:

i. 1381 patches (256 x 256) structured by Unique Dates and S2 Tiles. Each patch is provided along with the corresponding masks of pixel-level annotated classes (*_cl) and confidence levels (*_conf). Patches are given in GeoTiff format.

ii. Shapefiles data in WGS’84/ UTM projection, with file naming convention following the scheme: s2_dd-mm-yy_ttt, where s2 denotes the S2 sensor, dd denotes the day, mm the month, yy the year and ttt denotes the S2 tile. Shapefiles include the class of each annotation along with the confidence level and the marine debris report description.

iii. Train, Validation and Test split for evaluating machine learning algorithms.

iv. The assigned multi-labels for each patch (labels_mapping.txt).

The mapping between Digital Numbers and Classes is:

1: Marine Debris
2: Dense Sargassum
3: Sparse Sargassum
4: Natural Organic Material
5: Ship
6: Clouds
7: Marine Water
8: Sediment-Laden Water
9: Foam
10: Turbid Water
11: Shallow Water
12: Waves
13: Cloud Shadows
14: Wakes
15: Mixed Water

The mapping between Digital Numbers and Confidence level is:

1: High
2: Moderate
3: Low

The mapping between Digital Numbers and marine debris Report existence is:

1: Very close
2: Away
3: No

The final uncompressed dataset requires 4.38 GB of storage.