Deep Learning (DL)-based Ground Penetrating Radar (GPR) target recognition methods have been widely applied in geological exploration and infrastructure inspection. However, existing approaches face three main limitations: (1) Most GPR systems operate in single-polarization mode, leading to incomplete acquisition of target scattering information; (2) Traditional deep learning methods risk misclassification when handling B-scan images with similar hyperbolic features from different targets; (3) Direct input of 2D B-scan images into convolutional neural networks incurs high computational overhead. To address these challenges, this paper proposes a multi-polarimetric decomposition fusion method for GPR target recognition based on a lightweight MobileNetV3 network. The proposed method first acquires full-polarimetric GPR data (HH, VH, and VV polarizations) of subsurface targets. Subsequently, H-Alpha decomposition, Freeman decomposition, and Pauli decomposition are performed to extract eight polarimetric parameters characterizing the targets. These parameters are fused into an eight-dimensional feature matrix, which is then fed into a modified MobileNetV3 network integrated with a Squeeze-and-Excitation (SE) attention module for target identification. To verify the effectiveness of the method, four typical targets were classified in the experiments, and the results indicate that using the eight-dimensional feature matrix as the network input can enhance the target classification accuracy. The target classification accuracy can be further improved by incorporating the SE module into the network. Furthermore, compared to conventional ResNet18 and VGG16 networks, the improved MobileNetV3 achieves the highest recognition accuracy (98.75%) while significantly reducing parameter number and model size. The experimental results demonstrate that using an eight-dimensional feature matrix that includes target polarization information as the network input not only provides richer target information but also effectively reduces the redundant information of the input network. This improvement enhances the target classification accuracy while decreasing the matrix size of the input network. Additionally, the lightweight backbone network based on MobileNetV3, integrated with the SE attention mechanism, enhances critical feature extraction capabilities and strengthens discriminative power for target classification. The paper effectively addresses the challenges of insufficient feature discrimination and high computational load in GPR target classification.