Depth-aided prediction of vessel latitude based on a neural network

A method for determining vessel latitude based on depth using a neural network is proposed. The network takes as input a sequence of depth values measured by a single-beam echo sounder and predicts the vessel's latitude at the moment of the latest depth measurement. The network has two layers. The first layer contains neurons with hyperbolic tangent activation functions. The second layer consists of a single neuron with an identical activation function. The training dataset consists of training and validation sets. The training set is formed based on a depth layer contained within an electronic navigational chart (ENC). The validation set is formed by pseudorandom variations of input samples from the training set. Each of these variations corresponds to a constant sea level variation due to measurement errors and/or fluctuations of wind and/or tidal nature. The network is trained using the Adamax optimization algorithm. The maximum absolute value of latitude prediction error for the validation set is used as a criterion for training efficiency. After training, the network is tested using test samples obtained in the same manner as for the validation set. The simulation is conducted using the Python programming language. The TensorFlow library is used for training and operating the neural network. The simulation is conducted for several network configurations, each differing in the number of hidden neurons. As a result, it has been found that the networks show a tendency to learn how to predict vessel latitude using depth values as input data. This fact allows them to be considered as promising tools for bathymetric navigation.