GPR B-scan analysis with machine learning algorithms

Abstract

Agriculture is a sector facing increasing threats due to climate change, particularly in regions affected by prolonged droughts and irregular precipitation patterns. Efficient water management is essential to ensure resilience, especially in irrigation systems that depend on accurate knowledge of soil moisture and root distribution. This study presents a noninvasive robotic solution developed by the Centre de Disseny d’Equips Industrials (CDEI) within the RootBot project, which integrates Ground-Penetrating Radar (GPR) to collect and analyse subsurface data in real time. This study focuses on object detection in GPR data. This thesis aims to find a well-researched computer vision algorithm for this application, generate a labelled dataset large enough to obtain a reliable solution and test the effect of postprocessing techniques and data augmentation on the results. You Only Look Once (YOLO) was selected as the most suitable algorithm for the RootBot project. YOLO was trained using the final dataset, which was formed by tree root data from two field tests, GPR scans from various other applications and simulated data (created using GPRMax). Roboflow was used as the labelling and processing tool in this study. The results of the trained YOLO model demonstrate satisfactory performance detecting objects in GPR scans. Further advancements on this project should prioritize using a dataset made up of purely tree root data to ensure the applicability of the model for the agricultural sector. Finally, other algorithms could be tested if the project demanded higher accuracy without relying on real-time performance.