Optimizing multi-classifier fusion for seabed sediment classification using machine learning

ABSTRACT

Seabed sediment mapping with acoustical data and ground-truth samples is a growing field in marine science. In recent years, multi-classifier ensemble models have gained prominence for classification problems by combining several base classifiers. However, traditional ensemble methods do not consider the confidence scores of base classifiers, leading to suboptimal fusion when there are conflicting predictions. The current study introduces a novel optimization strategy that enhances the ensemble’s accuracy by constructing an ideal ensemble predicted probability matrix based on the fusion of predicted probabilities of the base classifiers, to improve seabed sediment mapping. The proposed approach not only addresses the limitations of traditional ensemble methods but also significantly increases the ensemble’s performance. The proposed approach demonstrates significant accuracy improvements. On the under-sampled dataset, it achieves 73.5% improvement compared to individual classifiers (random forest, decision tree, support vector machine), surpassing their respective accuracies. On the standard dataset, the ensemble model attains an accuracy of 79.1%, surpassing individual classifiers. Employing over-sampling techniques further elevates accuracy to 94.9%, exceeding the individual classifier performances. The proposed method is evaluated on acoustical data obtained from the Irish Sea. The proposed method outperforms base classifiers in terms of accuracy, F1 score, and the Kappa coefficient.

KEYWORDS:

• Bayesian theory
• differential evolution optimization
• feature extraction
• machine learning
• seabed sediment classification

Acknowledgments

The authors want to thank the British Geological Survey (BGS) provides the multi-beam dataset and ground-truth samples.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are openly available in ‘public geographic science database belongs to the British Geological Survey (BGS)’ at http://mapapps2.bgs.ac.uk/geoindex_offshore/home.html.

Additional information

Funding

This research was supported by the National Natural Science Foundation of China under grant [number 52201400], and supported by Shandong Provincial Natural Science Foundation under grant [number ZR202111260306], and supported by Special Projects for Promoting High Quality Economic Development (Marine Economic Development) in Guangdong Province under grant (GDNRC[2023]42), and Funding of Key Laboratory of Submarine Geoscience, Ministry of Natural Resources, under grant number KLSG2203.