Abstract: The backscatter data from multi-beam sonar contain acoustic information on the seafloor surface, which can be used to classify the seafloor. However, in practice, the cost of obtaining label information of a wide range of seafloor sediment types through physical sampling is too high, which restricts the performances of traditional supervised classification algorithms. In response to this problem, two semi-supervised learning classification algorithms based on auto-encoder pre-training and pseudo-label self-training are proposed, which can be used in the situation where there are only a large amount of unlabeled data and a small amount of labeled data in practical applications. The proposed algorithms are validated by using the multi-beam sonar backscatter data collected from two experiments in the same sea area in 2018 and 2019. Data processing results show that the proposed algorithms require less labeled data while ensuring classification accuracy, compared to the supervised classification algorithms only using labeled data. The classification accuracy of the semi-supervised learning classification method pre-trained by autoencoders is still above 75% when the number of labeled samples is very small.