Forward scattering of the target could cause the amplitude and phase aberration of the received sound field, which received attentions in harbor monitoring and anti-submarine. However, the localization under forward scattering configuration is a challenging task due to the strong direct blast. The method based on sensitive kernel function which exploit the aberration of the received signals is sensitive to the environment mismatch and a localization method based on transfer learning framework is developed. The envelopes of aberrations caused by the forward scattering of intruder are firstly extracted by applying pulse compression technique on the received signals, and then normalized by comparing with the case of intruder absent. The data set near the first arrivals on the normalized aberrations are selected as the learning physical parameters. A convolution neural network is trained with these data generated by the forward scattering model to establish a mapping relationship between intruder’s localization and the aberrations of received signal, thus the localization problem is transformed into classification. In the second step, the parameters of the convolutional pooling layer in the pre-trained model are frozen in the transfer learning procedure, and the parameters of the fully connected layer in the pre-trained model are updated using a small amount of data under the fluctuated environment. Simulation of the localization of ellipsoidal targets with a signal-to-noise ratio of 0 dB under a shallow water environment is performed for a scenario to explore the robustness of the method. The results show that the accurate target localization could be achieved in the case of sound velocity profile mismatch. Also, the method is not significantly sensitive to the target scattering function, sound properties of sediment and deployment of transceivers. The sensitivities to the waveguide amplitude and phase fluctuations are further modeled. The results show that good localization accuracy can be obtained in a relatively stable environment, and results are distinguished between the presence and absence of the target. Since the proposed method is derived by the model and real data, the accurate scattering model and sufficient training data are not significantly necessary. The method may provide a promising way for forward scattering detection.