Abstract: Aiming to address the localization problem of surface or underwater targets in a shallow sea environment, this paper investigates the localization of an underwater acoustic target using its radiated noise. A method for localizing underwater acoustic targets based on a densely connected neural network with multi-task learning and a convolutional block attention mechanism is proposed in this paper, which can simultaneously estimate the range and depth of the target. The input feature for network training is obtained by calculating the normalized sample covariance matrix of broadband data received by the vertical line array using the normal mode propagation model. Additionally, vectorization compression is applied to process the input features, effectively reducing the required number of parameters. Furthermore, through parameter sensitivity analysis, we discuss how mismatches between marine environmental parameters and sound speed profiles affect target localization performance. Simulation results demonstrate that compared to traditional matched field processing methods, our proposed approach exhibits higher localization accuracy and generalization ability when there is a mismatch in marine environments.