LassoNet-GNN: A Catalyst for Few-Shot Distorted SAR Image Classification

Few-shot learning (FSL) has revolutionized SAR image classification by enabling robust recognition with limited high-quality labeled data. However, existing methods struggle with both identifying key features in noisy, high-variation SAR data and handling geometric distortions. To address these challenges, we propose a novel contrastive sparse few-shot distorted SAR image classification framework via LassoNet-GNN, which incorporates GNN-based FSL framework with an adopted episodic training protocol and a LassoNet-based sparsification technique into an end-to-end architecture for more precise SAR image classification and informative feature selection with an optional plug-and-play contrastive learning (CL) regularization module that can be beneficial for improving robustness to geometric distortions on more complex data with richer intraclass variations in few-shot scenarios. Specifically, we first build upon a GNN-based FSL framework trained with an adopted episodic protocol that forms a learnable similarity graph over features extracted by a convolutional encoder, enabling effective message passing and strong base recognition capability under limited data. Additionally, we introduce a LassoNet-type inputto- output skip-layer across the entire GNN-based network and apply a sparse penalty to its weights, which implicitly drives the weights of unimportant nodes in the first hidden layer of the network toward zero. Finally, an auxiliary plug-andplay contrastive regularization module is employed by utilizing geometric transformation augmentations to construct positive and negative sample pairs, which can be optionally activated on few-shot complex data when beneficial for robustness. Experimental results on common few-shot SAR image datasets verify the effectiveness and efficiency of the proposed framework in classification accuracy and interpretability.