In this paper, we propose a neural network approach for automatic classification of underwater objects on sonar images. A major problem with sonar imagery applications is the difficulty to obtain large databases for training. Real sonar devices are costly and staged experiments where objects are well known and manually placed are rare because of cost. We show that the simultaneous use of parameters extraction and neural network pruning can significantly help to obtain good generalization rates (despite the lack of large training databases) and to reduce the complexity of the classifier. Furthermore, a method derived from neural network pruning is proposed to evaluate the significance of each parameter.