Multibeam echosounders are widely used for 3D bathymetric mapping, and increasingly for water column studies. However, they rapidly collect huge volumes of data, which poses a challenge for water column data processing that is often still manual and time-consuming, or affected by low efficiency and high false detection rates if automated. This research describes a comprehensive and reproducible workflow that improves efficiency and reliability of target detection and classification, by calculating metrics for target cross-sections using a commercial software before feeding into a feature-based semi-supervised machine learning framework. The method is tested with data collected from an uncalibrated multibeam echosounder around an offshore gas platform in the Adriatic Sea. It resulted in more-efficient target detection, and, although uncertainties regarding user labelled training data need to be underlined, an accuracy of 98% in target classification was reached by using a final pre-trained stacking ensemble model.