GPR Target Intelligent Recognition with Very Few Samples Training Based on Two-stage Optimization MAML

Ground Penetrating Radar (GPR) is a critical tool for non-destructive detection and identification of subsurface targets. It enables efficient surveying and precise discrimination of buried objects through intelligent recognition techniques. However, in scenarios with limited data availability, current intelligent recognition models exhibit constrained generalization capabilities and demonstrate poor adaptability to new environments and tasks. Therefore, this paper proposes a two-stage optimization MAML approach (A²-MAML) for GPR target recognition with few-shot training. The proposed method introduces two key enhancements. An adaptive loss mechanism dynamically weights a supervised contrastive loss to improve generalization. A feature augmentation module is introduced to capture task-invariant representations. With only a limited number of training samples, the model achieves robust performance under noisy conditions and reduces the risk of overfitting. Experimental results demonstrate the effectiveness of A²-MAML. Under few-shot settings, our method achieves F1-scores of 88.5% for new environments (6-way-3-shot) and 89.7% for new tasks (4-way-3-shot). This represents relative improvements of 3.4% and 7.9% over the original MAML baseline. This gain underscores the superiority of our approach in data-scarce scenarios. A²-MAML delivers accurate recognition from very few samples, reducing costly field surveys and allowing rapid deployment across sites to offer clear practical benefits. The framework also demonstrates the potential for real-time deployment on IoT edge nodes due to its efficient inference capability.