Multi-agent framework for expert-level Ground Penetrating Radar (GPR) image interpretation

Document Type

Article

Publication Date

6-2-2026

Abstract

Ground penetrating radar (GPR) has been a widely applied measurement technique in geological engineering investigations and subsurface structure detection, directly impacting engineering decision validity and safety. However, existing intelligent GPR interpretation methods exhibit a paradigm-level limitation: they reduce interpretation to closed-set visual recognition, ignoring engineering context and lacking the hierarchical professional reasoning that expert analysts rely upon. To address these limitations, this study reconceptualizes GPR image interpretation as a multi-stage, context-aware, and structured reasoning process, and proposes a multi-agent collaborative framework that decomposes the expert analytical workflow into four functionally specialized agents, each targeting a distinct cognitive layer of the interpretation process: global semantic analysis, local anomaly detection and description, engineering context retrieval, and information fusion. The framework is designed as an open architecture in which individual modules can be updated as advances in the underlying models emerge. The framework is validated on 3000 GPR profiles from three real-world hydropower engineering projects. Experimental results show that the proposed method achieves a BERTScore of 0.8422 and 0.8397 on waveform analysis and geological forecasting, an F1 score of 0.9208 on rock mass classification, an mAP@50 of 0.9284 on anomalous waveform detection, and a BERTScore of 0.8141 on final report generation. These results demonstrate that domain adaptation and engineering context integration are essential for professional-grade GPR interpretation, and that the proposed multi-agent paradigm offers a transferable design principle for AI-assisted measurement interpretation in knowledge-intensive geotechnical applications.

Publication Source (Journal or Book title)

Measurement Journal of the International Measurement Confederation

This document is currently not available here.

Share

COinS