Split-Conv: 1HD Depth Estimation Deep Learning Model

Bathymetry inversion is a common technique that is used to estimate the depth of the coastal area. That technique used dispersion relation to calculate the depth of the near coastal area from the characteristic of the waves. It can be approximately understood by first measuring the period and wavelength of the surface waves and inserting them to the dispersion relation. In this article we developed an interpretable deep learning model which explains the model to have the intended model. The model is expected to have physical interpretation like the conventional methods that use dispersion relation. Prior to our model, several methods that used deep learning models exclude the physical principle of dispersion relation. With some simple investigation to connect the model to the dispersion relation, and generate samples of possible time series dataset of wave profiles in 1HD (one-horizontal dimension), we present our solution in a model that is called Split-Conv. Our results show that the model follows the dispersion relation curves in the good agreement in the coastal depth between 5 and 35 meters. This is supported by testing the model to the test data that gives mean and standard deviation (−0.0046, 0.51) m respectively. Further exploration of the possibilities to scale the model into 2HD cases are discussed.