Structure-from-Motion (SfM) photogrammetry is a time and cost-effective method for high-resolution 3D mapping of cold-water corals (CWC) reefs and deep-water environments. The accurate classification and analysis of marine habitats in 3D provide valuable information for the development of management strategies for large areas at various spatial and temporal scales. Given the amount of data derived from SfM data sources such as Remotely-Operated Vehicles (ROV), there is an increasing need to advance towards automatic and semiautomatic classification approaches. However, the lack of training data, benchmark datasets for CWC environments and processing resources are a bottleneck for the development of classification frameworks. In this study, machine learning (ML) methods and SfM-derived 3D data were combined to develop a novel multiclass classification workflow for CWC reefs in deep-water environments. The Piddington Mound area, southwest of Ireland, was selected for 3D reconstruction from high-definition video data acquired with an ROV. Six ML algorithms, namely: Support Vector Machines, Random Forests, Gradient Boosting Trees, k-Nearest Neighbours, Logistic Regression and Multilayer Perceptron, were trained in two datasets of different sizes (1,000 samples and 10,000 samples) in order to evaluate accuracy variation between approaches in relation to the number of samples. The Piddington Mound was classified into four classes: live coral framework, dead coral framework, coral rubble and sediment and dropstones. Parameter optimisation was performed with grid search and cross-validation. Run times were measured to evaluate the trade-off between processing time and accuracy. In total, eighteen variations of ML algorithms were created and tested. The results show that four algorithms yielded f1-scores > 90% and were able to discern between the four classes, especially those with usually similar characteristics, e.g., coral rubble and dead coral. The accuracy variation among them was 3.6% which suggests that they can be used interchangeably depending on the classification task. Furthermore, results on sample size variations show that certain algorithms benefit more from larger datasets whilst others showed discrete accuracy variations (< 5%) when trained in datasets of different sizes.