The input data used for this competition is composed of raw and preprocessed subtiles extracted from the European Space Agency's Sentinel-2 L1C satellite products. Bathymetry survey data were retrieved from The French Naval Hydrographic and Oceanographic Service (SHOM) and The European Marine Observation and Data Network (EMODnet).

We use four different locations to create our datasets. French Guiana and Saint Louis in Senegal are used to create the training set. Two different locations on the Portuguese west coast are used to create the test set. The corresponding Sentinel-2 tiles are 22NCL, 28PCC, 29TNE, 29SMD for French Guiana, Saint Louis and the two locations in Portugal, respectively.
File descriptions
The following CSV files are provided:
mixed_train.csv
- the training set containing data from French Guiana (22NCL) and Saint Louis (28PCC).
mixed_validation.csv
- the validation set containing data from French Guiana (22NCL) and Saint Louis (28PCC).
mixed_test.csv
- the test set containing data from the Portuguese west coast (29TNE, 29SMD).
Data fields
All input CSV files have the following format:
image_file_name
- input image file identifier
z
- the depth corresponding to the input image
Input image files
Each input image is provided in two formats:
Raw
- The raw image extracted from a full-sized Sentinel-2 image(filename_RAW.npy).
Filtered
- The preprocessed image extract (filename.npy).
Input image preprocessing
To preprocess the input subtiles, raw subtiles are passed through a pass-band filter to remove all signals that do not resemble ocean waves, the normalized cross correlation (normxcorr) is calculated based on the remaining signals. The preprocessed images provided in these datasets are the results of the normxcorr function.