With the development of shipborne echo sounders in the early 20th century, bathymetric surveys saw massive leaps in both accuracy and convenience. However, even with modern echo sounders, there are still many hardships to overcome when conducting bathymetric surveys. These include high cost, unpredictable weather, high ship traffic, and potential geographic or diplomatic issues, to name a few.
To address these issues, scientists around the world have been developing satellite-derived bathymetry (SDB) techniques, which estimate water depth from multispectral satellite images. These methods can sometimes produce accurate results, especially for depths up to 20 meters. Unfortunately, most SDB models were developed using data from coastal regions with clear waters and a uniform distribution of seabed sediment. Since light reflects differently depending on water turbidity and the composition of the seabed, developing SBD models with consistent performance throughout different coastal environments has proven challenging.
Against this backdrop, a research team from Korea has been developing a new SDB model that leverages machine learning to shed light onto the various factors that can compromise accuracy, thus paving the way to potential solutions. Their latest study, which included Dr. Tae-ho Kim from Underwater Survey Technology 21 (UST21), was published in the Journal of Applied Remote Sensing on March 12, 2024.
One of the main goals of this study was to analyze how the model trained on different coastal regions would be affected by each region’s unique characteristics. To this end, they selected three areas around the Korean Peninsula: Samcheok, characterized by its clear waters, Cheonsuman, known for its turbid waters, and Hallim, where the seabed contains various types of sediments.
The team obtained multispectral satellite data of these regions from the Sentinel-2A/B missions, openly provided by the European Space Agency, and selected multiple images of these areas at different time points with clear skies. To train the SDB model on these data, they also acquired echo sounder-derived nautical charts from the Korea Hydrographic and Oceanographic Agency (KHOA); these charts were used as ground truth.
The SDB model itself was based on a well-established theoretical framework that links how light coming from the Sun is reflected by the atmosphere, the sea, and the seabed before reaching a satellite. As for the machine learning part of the model, the team employed a random forest algorithm because of its ability to adjust to multiple variables and parameters while handling large amounts of data.
Upon training and testing region-specific instances of the SDB model, the researchers found that accuracy was generally acceptable for Samcheok, with a root-mean-squared error of about 2.6 meters. In contrast, accuracy was markedly lower for both Cheonsuman and Hallim, with satellite-based depth predictions deviating significantly from KHOA measurements.
