In recent years the focus of anti-submarine warfare (ASW) has changed to more shallow operation areas. The mostpromising sensor in littoral areas is the low frequency active sonar (LFAS). but the combination of an effective sensor in a complicated environment has resulted in the need to cope with huge numbers of false-alarms. One way to reduce the falsealarm rate is by extracting characteristic features from sonar echoes and applying (single-ping) classification techniques. In this paper, pattern recognition algorithms are described to extract features from echo clusters in a standard sonar display (range versus bearing). Cluster analysis and investigation of the physical context yields a list of suitable features, which proved to be robust against environmental changes and produce optimal class-separation between target and non-target echoes. It was found on five training sets in different environments that the feature distributions can best be exploited by using a non-linear (quadratic) classification algorithm. The final classifier is implemented in such a way that a scaled (possub) number is generated, according to the submarine-likeness of each cluster. By applying a threshold to the possub number false alarm rates, were reduced by 60% in several test sets other than the training sets, while probability of detection remained nearly unaffected.