Lifelong Clustering for Seafloor Images Interpretation in AUV Surveys

Online image interpretation to show habitat distribution patterns is essential for enhancing the environmental perception of Autonomous Underwater Vehicles (AUVs). Classifi-cation methods are challenging to perform because the collected images may fall outside predefined classification boundaries. Clustering, however, offers an alternative approach without relying on predefined boundaries. To enable online image interpre-tation, we propose a LifeLong Clustering (LLC) method, which incrementally clusters streaming data. Specifically, a memory bank is employed to store a sparse subset of previously collected images to reduce the computational load. Cluster merging and splitting techniques are utilized to maintain the model components adaptively, ensuring it accommodates new data while pre-serving previously learned knowledge. The clustering results on a field survey dataset demonstrate that the proposed LLC method achieves an average Normalized Mutual Information (NMI) score of 0.65 when compared to the Non-LLC method, and 0.62 when compared to other LLC results obtained from different survey paths. It highlights the robustness and reliability of the proposed method, showcasing its ability to deliver consistent clustering results across varying conditions. This consistency underscores the potential of the method to enhance AUV decision-making in real-time, enabling more informed and efficient operations.