Abstract: A lightweight robust underwater radiated noise target recognition network construction and optimization method based on VGG-GAP and redundant feature pruning is proposed. This method optimizes VGGNet by combining global average pooling (GAP) to create a lightweight VGG-GAP network. The redundant convolutional kernels in VGG-GAP are removed using a network pruning algorithm relying on feature map correlation to obtain the optimal network structure. Experimental results on the ShipsEar dataset and DeepShip dataset show that this method can achieve nearly the same recognition accuracy with over 94% reduction in parameters and over 30% reduction in computational cost without retraining or fine-tuning the parameters. In the test on datasets with decreasing sample numbers and datasets with mismatched hydroacoustic channels, the networks improved by the proposed optimization method are more robust in small-sample datasets and mismatched underwater acoustic environments.