Ensemble Learning and Pruning for Image Classification in Maritime Management Systems
Abstract
In this thesis, the exploration of how maritime management systems can be optimized is considered, and it also revisits cutting-edge deep learning architectures based on ensemble learning techniques. Within the analysis and application of deep learning approaches, the study aims to fill gaps in maritime system methodologies. Those are: 1) Customizing more deep learning models for different maritime operations; and 2) Innovative ensemble learning can improve model performance and accuracy. A novel framework combining ensemble learning with pruning techniques is presented in this study. The goal of the framework is to improve efficient model selection while minimizing computational resource needs. Following diverse testing environments, for methodological assessment and considering various deep learning architectures within marine and maritime datasets, Kaggle’s public datasets platform is used. In this study, the consideration is given to enhancing ensemble models through strategic pruning. To create a more efficient predictive system, this involves carefully selecting and combining individual models within the ensemble. This approach proves effective across various marine and maritime situations, demonstrating its adaptability to different requirements. By offering innovative AI-based solutions for decision support and operational optimization, this study also contributes to maritime management systems, while the findings and methodologies proposed open avenues for future research and development in this area.
Several contributions offered by this thesis. First, covering topics from machine learning to deep learning, focusing on convolutional neural networks, and progressing towards ensemble learning and pruning techniques as a literature review. Secondly, the “Maritime mAnaGement eNsemble leArning sysTem (MAGNAT)” and the “Maritime mAnaGement eNsemble prUrning sysTem (MAGNUT) are proposed as two novel frameworks and strategies for enhancing model selection and performance optimization in maritime management systems. MAGNAT uses an ensemble aggregation strategy by assigning weighted factors based on their historical performance. MAGNUT is also an intelligent aggregation strategy that combines ensemble and pruning techniques to identify the minimal subset of nonredundant models within the ensemble. Thirdly, a diverse range of maritime datasets, including jellyfish datasets relevant to marine science and ship datasets related to maritime technology, is utilized for practical evaluation and calculation of the efficiency results presented in the thesis, along with hands-on Python programming.