A Lightweight SAR Ship Classification Method based on VV/VH Dual-Polarization Fusion and Knowledge Distillation

Abstract

Synthetic Aperture Radar (SAR) can acquire clear and reliable radar images under complex weather conditions such as rain and fog, offering significant research and application value in maritime surveillance, port monitoring, and ship target recognition. However, existing SAR ship target classification models still face challenges of large parameter counts and low inference efficiency in edge computing scenarios, and the classification performance for certain target categories remains unsatisfactory in practical situations involving limited samples and imbalanced data. To address these issues, this paper adopts a lightweight SAR ship target classification approach that combines VV-VH dual-polarization feature fusion with model compression. Using OpenSARShip_uint8 as the experimental dataset, the method first constructs dual-channel inputs that leverage the complementary scattering response information from VV and VH polarization channels to enhance the model's ability to represent ship structural and textural features, thereby improving classification stability. Second, to build a high-quality teacher model, this paper employs VGG16 pre-trained on ImageNet and designs a two-stage training strategy: in the first stage, the feature extraction layers are frozen and only the classifier is trained; in the second stage, all network parameters are unfrozen, and a differentiated learning rate strategy is applied for fine-tuning the entire network-a smaller learning rate is used for the feature extraction layers to preserve pre-trained features, while a relatively larger learning rate is used for the classifier to accelerate adaptation, combined with cosine annealing scheduling to gradually decay the learning rate. This strategy helps improve the model's capacity to represent fine-grained features in SAR images and mitigates, to a certain extent, the model bias caused by data imbalance. Furthermore, knowledge distillation is employed to transfer the feature learning capability of the teacher model to a lightweight student model, significantly reducing the model size while preserving recognition accuracy as much as possible. Experimental results show that the teacher model achieves a test accuracy of 74.08%, while the distilled student model attains a test accuracy of 73.86% with only a 0.22% accuracy degradation. Despite the moderate absolute classification accuracy limited by speckle noise, inter-class scattering ambiguity and dataset imbalance, the proposed distillation framework drastically shrinks the model scale: parameters drop from 18.92M to 1.24M, storage footprint falls from 72.17MB to 4.74MB, and single-image inference latency decreases from 1.93ms to 0.71ms.