Abstract
This study presents a lightweight neural network model integrated with a microwave-based detection system for identifying buried objects. The proposed model is trained and tested exclusively on real-world measurements, enhancing its practical relevance and robustness. The system utilizes 16 × 16 scattering parameter (S-parameter) measurements, transformed into a compact 256-dimensional feature vector that captures the microwave response of subsurface materials. This representation enables a neural network architecture with reduced computational complexity while maintaining high accuracy. Experimental evaluations demonstrate that the proposed model achieves an accuracy of 99.83%, an F1 score of 0.989, and a recall of 0.979 in distinguishing hazardous from non-hazardous (safe) objects, outperforming baseline CNN, DRN, and EfficientNet architectures. These results confirm the suitability of the approach in defense and security applications.