MSFormer: A Multisparsity Transformer for Reconstructing Missing Traces in GPR B-Scans via Self-Supervised Learning

The integrity of ground-penetrating radar (GPR) B-scans is essential for reliable interpretation of the object's information. However, acquiring complete and high-quality B-scans is often hindered by field obstructions and measurement inefficiencies, leading to severely corrupted data. In this article, we propose a multisparsity transformer model, which jointly models global contextual dependencies and local structural information for recovering missing traces in corrupted B-scan. A self-supervised learning strategy is further introduced to enable effective training without requiring complete B-scan data as ground truth. The proposed approach is evaluated on both synthetic and measurement datasets, including linear scanning for underground sensing and circumferential scanning for defect detection in cylindrical objects. Experimental results demonstrate superior reconstruction performance over existing deep-learning methods, maintaining high accuracy even with up to 90% randomly missing traces and large consecutive missing regions. These results, together with downstream task validation, indicate that the proposed approach can substantially improve the quality of corrupted GPR data and enhance its practical utility in challenging real-world acquisition environments.