Detection of underground objects from GPR data using a lightweight YOLO-based approach

Underground object detection presents significant challenges due to the noisy nature of Ground Penetrating Radar (GPR) data, the small scale of targets, and ground-related complexities. This study presents a task-specific YOLOv8n-based framework for detecting small subsurface objects in noisy GPR radargrams. The proposed model integrates GhostConv and C3Ghost modules into the architecture to reduce computational costs and enhance feature extraction capabilities. This ghost block-based structure minimizes unnecessary convolution operations, enabling more effective feature representations. The model was trained and tested on four different GPR datasets obtained from the Roboflow platform, using both individual and combined Integrated Datasets. Performance was evaluated using F1-score, mean Average Precision at IoU = 0.5 (mAP@0.5), mean Average Precision averaged over IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95), and inference time (ms/frame). Experimental results show that the proposed approach achieves improvements of up to 5.7% and 5.9% in mAP and F1 scores, respectively, on the Integrated Dataset. The findings reveal that it increases detection accuracy while reducing processing time and offers practical applicability for real-time use in non-destructive excavation and underground inspection applications.