In this thesis, we investigate a semi-supervised clustering approach for the use of benthic habitat mapping. Specifically, we introduce a model that extends existing deep clustering methods by utilizing labeled as well as unlabeled data. Our model builds upon deep-learning-based dimensionality reduction and integrates a clustering loss, so that feature extraction and class seperation can jointly be optimized. By utilizing the full range of available backscatter data alongside a limited number of labeled grab samples, our method ultimately aims to reduce the dependency on extensive ground-truthing, making habitat mapping more efficient. We evaluate our model by using benchmark datasets as well as real-world backscatter data from two study sites in the Sylt Outer Reef.