INTERNAL WAVES SERVICE

In partnership with

1. Overview

Oceanic internal waves (IWs) propagate within the ocean when it is stratified and disturbed by physical mechanisms. These waves, typically nonlinear (Internal Solitary Waves – ISWs), can reach amplitudes of over 100 meters. They generate the highest vertical velocities in the ocean and strong horizontal shear currents, often causing underwater navigation accidents and damaging sea platforms (Osborne and Burch 1980). Furthermore, ISWs resuspend sediments over the continental shelf (Quaresma et al. 2007) and can produce intense mixing in the deep ocean. IWs can be detected by remote sensing satellites through variations in sea surface roughness, observable with optical and radar sensors (Magalhaes and da Silva 2017; Zhang et al. 2020; Santos-Ferreira et al. 2018, 2019, 2022, 2023). However, the most effective method for observing IWs across the global ocean is using SAR imaging, as the Sentinel-1 images used in this service (due e.g. to the cloud contamination in case of the optical images). 

The “Internal Waves Service” will provide near real-time updates by mapping detected events on an interactive platform, highlighting locations where internal waves are observed. Leveraging a pre-trained model on similar datasets, the service will analyse thousands of images from the Sentinel-1 satellite (WV mode vignettes). This innovative service will generate daily/weekly/monthly/yearly global maps, offering valuable insights into the distribution of internal waves worldwide, except in regions where such imagery is unavailable.

2. Internal Waves Service

This service sources all Copernicus Sentinel 1 WV mode vignettes SAR (Synthetic Aperture RADAR) images (imagettes) and classifies them according to wether or not they depict an internal wave. It proceeds to store the positive images on S3 and their metadata on SQL a database for efficient retrieval.

  • Data Acquisition and Processing:
    – Continuously ingests Sentinel-1 WV mode vignettes from global acquisitions.
    – The images were sourced from: IFREMER: X-Waves
  • Machine Learning Classification:
    – Employs a state-of-the-art machine learning model to classify each vignette, determining the presence or absence of internal waves. The model is trained on a diverse dataset of imagettes of different oceans featuring various oceanic conditions
  • Data Management:
    – Positive detections: Vignettes classified as containing internal waves are stored in S3 object storage for efficient retrieval and further analysis.
    – Comprehensive metadata for all processed vignettes, including classification results, acquisition details, and geolocation information, is stored in a SQL database.
  • Key Features:
    – Global Coverage: Analyzes Sentinel-1 WV mode data from all ocean basins worldwide.
    – Near Real-Time Processing: Rapidly ingests and processes new acquisitions to maintain an up-to-date dataset.
    – Efficient Retrieval: Optimized database schema and object storage integration allow for quick access to classified imagery and associated metadata.
    – Scalability: Designed to handle the high volume of Sentinel-1 WV mode acquisitions, which can reach up to 75 minutes per orbit. Other satellites might be added in the future.

This service provides researchers and oceanographers with a powerful tool for studying internal wave phenomena on a global scale, leveraging the unique capabilities of Sentinel-1’s Wave mode acquisitions.

3. Dataset used to develop the model

A competition was launched in Kaggle platform titled “Automatic Identification of Internal Waves”. In that competition (available here), the dataset comprises 4864 images, resembling photographs of the ocean, provided in .PNG format. These images are WV mode vignettes (also known as imagettes) from the Sentinel-1 satellite. Captured in a single polarization, each image covers a 20 km by 20 km area, with acquisitions made every 100 km along the satellite’s orbit and 5 m by 5 m of spatial resolution. A team of experts with 7 to 30 years of experience in internal wave research manually annotated (labeled and classified) the images. The dataset includes samples from the Atlantic, Indian, and Pacific Oceans, collected at various dates and times over the past four years. All images were acquired and processed using a consistent methodology. The data has been organized according to the guidelines provided by the Kaggle platform as follows:

  • Dataset Balance: During the preparation of the dataset, it was made sure that both the training (70%) and test (30%) datasets contain a balanced number (50%) of images with the target (feature that we aim to detect) internal waves.
  • Data Augmentation: It was applied a basic data augmentation task based on image rotation, whereby each original was rotated by 90, 180 and 270 degrees. These images are included in the dataset, which is distributed by the Train and Test folders.
  • Folders: There are two folders with the images: Train folder, that contains images for training models, in .PNG format. The name of each file corresponds to the id of the image in the folders (e.g. image 5.png corresponds to id 5 on the tables); Test folder, that contains images for testing models, in .PNG format.

  • Files: There are four files .csv: train.csv, that contains a table listing the images of the training set (contents of the Train folder). The table contains the columns ‘id’ and ‘ground_truth’. 70% with guaranteed 50% positive; test.csv, that contains a table listing the images of the testing set (contents of the Test folder). The table contains the column ‘id’. 30% with guaranteed 50% positive; solution.csv – A table listing the images of the testing set (contents of the Test folder). The table contains the columns ‘id’ and ‘ground_truth’; sample_submission.csv – a sample submission file in the expected format.

  • Columns: The files can contain one or two columns: id, that is the unique identifier of an image in the whole dataset. In other words, id’s do not repeat on the test and trains datasets. The id of an image is the file name plus the extension.png (e.g. id 5 on tables refers to image 5.png); ground_truth that is the confirmation by experts wether the image contains (1) or not (0) the feature of interest (internal wave).
Table 1 – Examples of 16 images containing internal waves can be found in the Train folder. In these cases, the ground truth is labeled as 1. The id of these images are 70, 83, 87, 128, 1283, 1628, 1736, 1898, 2383, 3252, 3305, 3341, 3368, 3420, 3494, and 3549, respectively.
Table 2 –  Examples of 16 images that don’t contain internal waves can be found in the Train folder. In these cases, the ground truth is labeled as 0. The id of these images are 770, 877, 1023, 1099, 1461, 1468, 2448, 2510, 2579, 2704, 2786, 2891, 3127, 3613, 4069, and 4089, respectively.

The images were sourced from: https://xwaves.ifremer.fr/#/quicklook
The use of the images is regulated by: http://en.data.ifremer.fr/All-about-data/Data-access-conditions

4. References

Magalhaes, J. M., and J. C. B. da Silva, 2017: Satellite altimetry observations of large-scale internal solitary waves. IEEE Trans. Geosci. Remote Sens., 14, 534-538, https://doi.org/10.1109/LGRS.2017.2655621.

Osborne, A. R., and T. L. Burch, 1980. Internal solitons in the Andaman Sea. Science, 208, 451–460, https://doi.org/10.1126/science.208.4443.451.

Pinelo, J., A. M. Santos-Ferreira, C. Capinha,  and J. C. B. da Silva, 2024: Automatic Identification of Internal Waves. Kaggle, https://kaggle.com/competitions/internal-waves.

Quaresma, L. S., J. Vitorino, A. Oliveira, and J. C. B. da Silva, 2007: Evidence of sediment resuspension by nonlinear internal waves on the western Portuguese mid-shelf. Mar. Geol., 246(2-4), 123-143, https://doi.org/10.1016/j.margeo.2007.04.019.

Santos-Ferreira, A. M., J. C. B. da Silva, and J. M. Magalhaes, 2018: SAR Mode altimetry observations of internal solitary waves in the Tropical Ocean Part 1: Case studies. Remote Sens., 10, 644, https://doi.org/10.3390/rs10040644.

Santos-Ferreira, A. M., J. C. B. da Silva, J. M. Magalhaes, and M. Srokosz, 2019: SAR mode altimetry observations of internal solitary waves in the Tropical Ocean Part 2: A method of detection. Remote Sens., 11(11), 1339, https://doi.org/10.3390/rs11111339.

Santos-Ferreira, A. M., J. C. B. da Silva, J. M. Magalhaes, S. Amraoui, T. Moreau, C. Maraldi, F. Borde, N. Picot, and F. Borde, 2022: Effects of Surface Wave Breaking Caused by Internal Solitary Waves in SAR Altimeter: Sentinel-3 Copernicus Products and Advanced New Products. Remote Sens., 14, 587, https://doi.org/10.3390/rs14030587.

Santos-Ferreira, A. M., J. C. B. Silva, B. St-Denis, D. Bourgault, and L. R. M. Maas, 2023: Internal Solitary Waves within the Cold Tongue of the Equatorial Pacific generated by buoyant gravity currents. J. Phys. Oceanogr., 53(10), 2419-2434, https://doi.org/10.1175/JPO-D-22-0165.1.