Attention-Enhanced Temporal LSTM Network as a Sonar Position Predictor for Side-Scanning Sonar Waterfall Image

Analysis of the river bottom using side-scan sonar allows for obtaining the so-called ‘‘waterfall’’ image - a scrolling visualization of sonar reflections from the bottom and underwater objects. This type of image allows for a detailed characterization of the bottom structure and detection of various objects. This image is created based on the measurement of the return time of the sound wave reflected from various surfaces and the position of the sonar. Unfortunately, quite often, objects (such as bridges) contribute to the loss of GPS signals and, consequently, the generation of distortions in this area. In this paper, we present a recurrent neural network model with attention modules to predict the lost GPS signal during measurements. The proposed predictor architecture combines local processing with a deep tower of alternating unidirectional and bidirectional LSTM (Long Short-Term Memory) layers with residual connections. Furthermore, we implemented multihead self-attention to capture global temporal dependencies and RevIN (Reversible Instance Normalization), which allows for adaptation to changing input data distributions. The evaluation of the methodology was tested on data collected in real conditions on the Odra River in Szczecin, Poland. The results indicate a very good adaptation of the predictor to GPS signal correction during its loss under bridges.