Application of mathematical transformation and deep learning methods as a new approach to classification of underwater objects

Sound Navigation and Ranging (Sonar) is a system used to detect size, distance, direction and other information about objects using sound waves. It is widely used in submarine oil exploration, seafloor mapping, tracking fish shoals, detecting shipwrecks and debris, and most importantly detecting mines and other objects such as rocks that are very similar in shape and structure to mines. Feature extraction, which should be used for the identification and classification of sonar signals, the selection of the most appropriate algorithms, and the hyperparameter optimization of these algorithms are scientific problems that have been studied for many years. In this study, instead of classical machine learning algorithms and feature extraction processes, with an innovative approach; The representation of one-dimensional sonar data in image format is proposed by using three different mathematical transformation methods: Gramian Angular Field (GAA, GATA, GAFA), Markov Transition Field (MGA) and Repetition Graph (TG). Thanks to the representation of the data in image format, it is possible to use Convolutional Neural Network (ESA) Algorithms instead of classical machine learning algorithms. Within the scope of the proposed innovative method, the performance of four different models created by using the new data set created by 2D images obtained by using mathematical transformations, performance enhancement techniques and different convolutional neural network architectures were tested. In the data for performance increase; It has been ensured that the most remarkable features of the data are found with Single Value Decomposition (TDA), and new data with more features are obtained by transferring the images obtained from three different transformation methods with Image Merging methods to three channels (RGB) of a single image. The performance results of the models produced, other studies with the same data set examined in the literature, and the results obtained with the classical algorithms studied before changing the representation of the data in this study were compared using well-known metrics. In the light of these results, it was determined that the proposed innovative approach provided the best results among the studies conducted with the same data set examined in the literature.