Identifying the optimal k-mean clustering for ship noise spectrograms

Unsupervised learning techniques have the potential to be useful for ocean acoustic data. For example, the k-means clustering algorithm has been trained on synthetic spectrograms from surface ships and then applied to measured data to try to identify probabilities of seabed type and details such as the speed and location of the ships. Although the algorithm is effective in grouping data into clusters, the variability of these clusters is heavily dependent on the number of clusters, “k.” This paper aims to investigate the range of variability in clusters when different values of k are chosen for a given ocean acoustics dataset. To assess the variability of the results, different methods will be used to measure distances from the centroid of each cluster. These methods include measuring the average, maximum, and sum of squared distances from each data sample to its assigned centroid. The results when using the optimal number of clusters are shown and a discussion is provided about how to interpret the results.